llm_knowledge_distillation.ipynb

{
  "cells": [
    {
      "cell_type": "markdown",
      "metadata": {},
      "source": [
        "### **Simplified Implementation of knowledge distillation in LLMs**\n",
        "This is a minimal single-file script Implementation of knowledge distillation in LLMs. Knowledge distillation uses a large “teacher” model to guide the training of a smaller “student” model so that the student mimics the teacher’s core capabilities but at reduced computational cost. In this implementation, we use GPT-2 124M as student model and GPT-2 Medium 340M as teacher via reverse Kullback-Leibler (KL) divergence. \n",
        "\n",
        "**Reference Papers**: https://arxiv.org/abs/1503.02531, https://arxiv.org/abs/2306.08543 \n",
        "\n",
        "Bsically we minimize the *reverse KL* divergence, proceeding through the following steps:\n",
        "\n",
        "**(1)** We first define our overall objective:\n",
        "$$\n",
        "\\theta \\;=\\;\\arg\\min_{\\theta} L(\\theta).\n",
        "$$\n",
        "Here, $\\theta$ are the parameters of the student model, and $L(\\theta)$ is the loss we aim to minimize.\n",
        "\n",
        "**(2)** We then instantiate $L(\\theta)$ as the reverse KL divergence between the student distribution $q_{\\theta}$ and the teacher distribution $p$:\n",
        "$$\n",
        "\\theta \\;=\\;\\arg\\min_{\\theta}\\;\\mathrm{KL}\\bigl[q_{\\theta} \\,\\|\\, p\\bigr].\n",
        "$$\n",
        "This focuses training on matching the teacher’s high-probability outputs, avoiding the overemphasis on rare modes induced by forward KL.\n",
        "\n",
        "**(3)** Finally, we expand the reverse KL objective for text generation:\n",
        "$$\n",
        "\\theta \\;=\\;\\arg\\min_{\\theta}\n",
        "\\Bigl[\\,-\\,\\mathbb{E}_{\\,x \\sim p_x,\\;y\\sim q_{\\theta}}\\;\\log\\!\\biggl(\\frac{p(y\\mid x)}{\\,q_{\\theta}(y \\mid x)}\\biggr)\\Bigr].\n",
        "$$\n",
        "In practice, this leads to a *mode-seeking* behavior, aligning the student’s most likely outputs with those of the teacher while assigning lower probability to unlikely sequences.\n",
        "\n",
        "<img src=\"assets/meth.png\" />\n"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 13,
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 1000
        },
        "id": "WwYnN9Ahiyig",
        "outputId": "0f3f39df-442a-4800-cac0-71d773024889"
      },
      "outputs": [
        {
          "name": "stdout",
          "output_type": "stream",
          "text": [
            "Loading teacher and student models...\n",
            "Loading a small subset of OpenWebText dataset...\n",
            "Tokenizing dataset...\n"
          ]
        },
        {
          "name": "stderr",
          "output_type": "stream",
          "text": [
            "/usr/local/lib/python3.10/dist-packages/transformers/optimization.py:591: FutureWarning: This implementation of AdamW is deprecated and will be removed in a future version. Use the PyTorch implementation torch.optim.AdamW instead, or set `no_deprecation_warning=True` to disable this warning\n",
            "  warnings.warn(\n"
          ]
        },
        {
          "name": "stdout",
          "output_type": "stream",
          "text": [
            "\u001b[1;30;43mStreaming output truncated to the last 5000 lines.\u001b[0m\n",
            "step 17/5000, loss: 1.1983\n",
            "step 18/5000, loss: 1.0328\n",
            "step 19/5000, loss: 1.2989\n",
            "step 20/5000, loss: 1.3451\n",
            "step 21/5000, loss: 1.2157\n",
            "step 22/5000, loss: 1.1031\n",
            "step 23/5000, loss: 1.0572\n",
            "step 24/5000, loss: 0.9709\n",
            "step 25/5000, loss: 1.5088\n",
            "step 26/5000, loss: 1.1439\n",
            "step 27/5000, loss: 0.8624\n",
            "step 28/5000, loss: 1.4996\n",
            "step 29/5000, loss: 2.0317\n",
            "step 30/5000, loss: 0.9370\n",
            "step 31/5000, loss: 1.0354\n",
            "step 32/5000, loss: 1.3377\n",
            "step 33/5000, loss: 0.9648\n",
            "step 34/5000, loss: 1.2608\n",
            "step 35/5000, loss: 1.3695\n",
            "step 36/5000, loss: 1.0112\n",
            "step 37/5000, loss: 1.0084\n",
            "step 38/5000, loss: 1.0430\n",
            "step 39/5000, loss: 2.9537\n",
            "step 40/5000, loss: 1.0600\n",
            "step 41/5000, loss: 1.1919\n",
            "step 42/5000, loss: 1.5256\n",
            "step 43/5000, loss: 1.0082\n",
            "step 44/5000, loss: 1.2172\n",
            "step 45/5000, loss: 1.0341\n",
            "step 46/5000, loss: 1.0750\n",
            "step 47/5000, loss: 1.3476\n",
            "step 48/5000, loss: 3.5077\n",
            "step 49/5000, loss: 1.2183\n",
            "step 50/5000, loss: 0.9547\n",
            "step 51/5000, loss: 1.0068\n",
            "step 52/5000, loss: 1.2507\n",
            "step 53/5000, loss: 1.1025\n",
            "step 54/5000, loss: 1.0777\n",
            "step 55/5000, loss: 1.2195\n",
            "step 56/5000, loss: 1.2412\n",
            "step 57/5000, loss: 1.1300\n",
            "step 58/5000, loss: 1.1611\n",
            "step 59/5000, loss: 3.5082\n",
            "step 60/5000, loss: 1.2546\n",
            "step 61/5000, loss: 1.1623\n",
            "step 62/5000, loss: 1.6945\n",
            "step 63/5000, loss: 1.1399\n",
            "step 64/5000, loss: 1.1540\n",
            "step 65/5000, loss: 1.3702\n",
            "step 66/5000, loss: 0.9925\n",
            "step 67/5000, loss: 1.1214\n",
            "step 68/5000, loss: 1.3032\n",
            "step 69/5000, loss: 0.8705\n",
            "step 70/5000, loss: 1.0997\n",
            "step 71/5000, loss: 1.2914\n",
            "step 72/5000, loss: 1.0212\n",
            "step 73/5000, loss: 0.8346\n",
            "step 74/5000, loss: 0.9946\n",
            "step 75/5000, loss: 1.1369\n",
            "step 76/5000, loss: 1.0829\n",
            "step 77/5000, loss: 0.9337\n",
            "step 78/5000, loss: 1.0749\n",
            "step 79/5000, loss: 1.1923\n",
            "step 80/5000, loss: 0.9913\n",
            "step 81/5000, loss: 2.3528\n",
            "step 82/5000, loss: 1.1681\n",
            "step 83/5000, loss: 1.0376\n",
            "step 84/5000, loss: 1.1339\n",
            "step 85/5000, loss: 0.9449\n",
            "step 86/5000, loss: 1.2871\n",
            "step 87/5000, loss: 0.9934\n",
            "step 88/5000, loss: 0.8423\n",
            "step 89/5000, loss: 1.2037\n",
            "step 90/5000, loss: 1.0848\n",
            "step 91/5000, loss: 1.1021\n",
            "step 92/5000, loss: 1.5025\n",
            "step 93/5000, loss: 1.0970\n",
            "step 94/5000, loss: 1.0903\n",
            "step 95/5000, loss: 1.0514\n",
            "step 96/5000, loss: 1.4298\n",
            "step 97/5000, loss: 0.9762\n",
            "step 98/5000, loss: 1.1686\n",
            "step 99/5000, loss: 0.9787\n",
            "step 100/5000, loss: 0.9496\n",
            "step 101/5000, loss: 1.0375\n",
            "step 102/5000, loss: 0.9428\n",
            "step 103/5000, loss: 1.0139\n",
            "step 104/5000, loss: 0.9383\n",
            "step 105/5000, loss: 0.8680\n",
            "step 106/5000, loss: 1.3890\n",
            "step 107/5000, loss: 1.2961\n",
            "step 108/5000, loss: 1.1176\n",
            "step 109/5000, loss: 0.8895\n",
            "step 110/5000, loss: 1.1357\n",
            "step 111/5000, loss: 1.1102\n",
            "step 112/5000, loss: 1.3081\n",
            "step 113/5000, loss: 1.5310\n",
            "step 114/5000, loss: 0.8925\n",
            "step 115/5000, loss: 0.8267\n",
            "step 116/5000, loss: 1.1065\n",
            "step 117/5000, loss: 0.9138\n",
            "step 118/5000, loss: 1.4071\n",
            "step 119/5000, loss: 1.3649\n",
            "step 120/5000, loss: 0.8401\n",
            "step 121/5000, loss: 0.9635\n",
            "step 122/5000, loss: 1.0193\n",
            "step 123/5000, loss: 1.5801\n",
            "step 124/5000, loss: 1.2128\n",
            "step 125/5000, loss: 1.1033\n",
            "step 126/5000, loss: 1.1547\n",
            "step 127/5000, loss: 1.0113\n",
            "step 128/5000, loss: 0.9866\n",
            "step 129/5000, loss: 1.1621\n",
            "step 130/5000, loss: 1.0496\n",
            "step 131/5000, loss: 1.0071\n",
            "step 132/5000, loss: 1.0409\n",
            "step 133/5000, loss: 1.2029\n",
            "step 134/5000, loss: 1.1648\n",
            "step 135/5000, loss: 0.9332\n",
            "step 136/5000, loss: 1.3122\n",
            "step 137/5000, loss: 0.9534\n",
            "step 138/5000, loss: 1.1096\n",
            "step 139/5000, loss: 0.9869\n",
            "step 140/5000, loss: 1.0774\n",
            "step 141/5000, loss: 1.0600\n",
            "step 142/5000, loss: 1.1405\n",
            "step 143/5000, loss: 1.1046\n",
            "step 144/5000, loss: 1.1607\n",
            "step 145/5000, loss: 0.9080\n",
            "step 146/5000, loss: 1.2554\n",
            "step 147/5000, loss: 1.1488\n",
            "step 148/5000, loss: 0.9710\n",
            "step 149/5000, loss: 1.0813\n",
            "step 150/5000, loss: 0.8360\n",
            "step 151/5000, loss: 0.8606\n",
            "step 152/5000, loss: 1.0405\n",
            "step 153/5000, loss: 1.0697\n",
            "step 154/5000, loss: 1.0399\n",
            "step 155/5000, loss: 0.9997\n",
            "step 156/5000, loss: 1.0299\n",
            "step 157/5000, loss: 1.5898\n",
            "step 158/5000, loss: 1.2061\n",
            "step 159/5000, loss: 0.9859\n",
            "step 160/5000, loss: 1.0056\n",
            "step 161/5000, loss: 1.2486\n",
            "step 162/5000, loss: 1.5523\n",
            "step 163/5000, loss: 2.1372\n",
            "step 164/5000, loss: 0.8754\n",
            "step 165/5000, loss: 0.9947\n",
            "step 166/5000, loss: 0.9568\n",
            "step 167/5000, loss: 0.9467\n",
            "step 168/5000, loss: 0.9980\n",
            "step 169/5000, loss: 0.9627\n",
            "step 170/5000, loss: 0.7837\n",
            "step 171/5000, loss: 1.1150\n",
            "step 172/5000, loss: 1.1020\n",
            "step 173/5000, loss: 0.8871\n",
            "step 174/5000, loss: 1.0290\n",
            "step 175/5000, loss: 0.9989\n",
            "step 176/5000, loss: 1.2460\n",
            "step 177/5000, loss: 4.3528\n",
            "step 178/5000, loss: 1.0736\n",
            "step 179/5000, loss: 0.8233\n",
            "step 180/5000, loss: 0.7744\n",
            "step 181/5000, loss: 1.6955\n",
            "step 182/5000, loss: 1.4987\n",
            "step 183/5000, loss: 1.0515\n",
            "step 184/5000, loss: 1.2293\n",
            "step 185/5000, loss: 1.0914\n",
            "step 186/5000, loss: 1.2074\n",
            "step 187/5000, loss: 1.1375\n",
            "step 188/5000, loss: 0.8072\n",
            "step 189/5000, loss: 1.0810\n",
            "step 190/5000, loss: 1.0589\n",
            "step 191/5000, loss: 1.0127\n",
            "step 192/5000, loss: 0.8061\n",
            "step 193/5000, loss: 1.0849\n",
            "step 194/5000, loss: 0.9894\n",
            "step 195/5000, loss: 1.1841\n",
            "step 196/5000, loss: 1.0097\n",
            "step 197/5000, loss: 1.2228\n",
            "step 198/5000, loss: 0.9772\n",
            "step 199/5000, loss: 0.9072\n",
            "step 200/5000, loss: 0.9158\n",
            "step 201/5000, loss: 0.9218\n",
            "step 202/5000, loss: 0.9565\n",
            "step 203/5000, loss: 1.1358\n",
            "step 204/5000, loss: 1.0980\n",
            "step 205/5000, loss: 1.1123\n",
            "step 206/5000, loss: 0.8010\n",
            "step 207/5000, loss: 1.0284\n",
            "step 208/5000, loss: 1.0947\n",
            "step 209/5000, loss: 0.9308\n",
            "step 210/5000, loss: 1.0436\n",
            "step 211/5000, loss: 1.0609\n",
            "step 212/5000, loss: 1.1810\n",
            "step 213/5000, loss: 0.9320\n",
            "step 214/5000, loss: 1.1337\n",
            "step 215/5000, loss: 1.0272\n",
            "step 216/5000, loss: 0.9549\n",
            "step 217/5000, loss: 0.8834\n",
            "step 218/5000, loss: 1.1316\n",
            "step 219/5000, loss: 1.0069\n",
            "step 220/5000, loss: 1.1009\n",
            "step 221/5000, loss: 1.3218\n",
            "step 222/5000, loss: 1.3255\n",
            "step 223/5000, loss: 1.1158\n",
            "step 224/5000, loss: 1.0081\n",
            "step 225/5000, loss: 0.7773\n",
            "step 226/5000, loss: 1.0780\n",
            "step 227/5000, loss: 0.8335\n",
            "step 228/5000, loss: 1.0866\n",
            "step 229/5000, loss: 1.1753\n",
            "step 230/5000, loss: 1.0006\n",
            "step 231/5000, loss: 0.9988\n",
            "step 232/5000, loss: 1.1067\n",
            "step 233/5000, loss: 0.9861\n",
            "step 234/5000, loss: 0.9862\n",
            "step 235/5000, loss: 1.4799\n",
            "step 236/5000, loss: 1.3351\n",
            "step 237/5000, loss: 1.0593\n",
            "step 238/5000, loss: 1.1146\n",
            "step 239/5000, loss: 1.1577\n",
            "step 240/5000, loss: 0.9281\n",
            "step 241/5000, loss: 0.9574\n",
            "step 242/5000, loss: 0.9058\n",
            "step 243/5000, loss: 1.2861\n",
            "step 244/5000, loss: 2.7263\n",
            "step 245/5000, loss: 0.8928\n",
            "step 246/5000, loss: 0.9724\n",
            "step 247/5000, loss: 1.4774\n",
            "step 248/5000, loss: 1.3670\n",
            "step 249/5000, loss: 0.9730\n",
            "step 250/5000, loss: 0.9395\n",
            "step 251/5000, loss: 1.0259\n",
            "step 252/5000, loss: 0.8804\n",
            "step 253/5000, loss: 1.0955\n",
            "step 254/5000, loss: 0.9270\n",
            "step 255/5000, loss: 0.8476\n",
            "step 256/5000, loss: 1.3953\n",
            "step 257/5000, loss: 1.0047\n",
            "step 258/5000, loss: 1.2568\n",
            "step 259/5000, loss: 1.1463\n",
            "step 260/5000, loss: 1.2159\n",
            "step 261/5000, loss: 1.0535\n",
            "step 262/5000, loss: 1.0600\n",
            "step 263/5000, loss: 1.0232\n",
            "step 264/5000, loss: 1.0122\n",
            "step 265/5000, loss: 0.9735\n",
            "step 266/5000, loss: 1.1792\n",
            "step 267/5000, loss: 0.7632\n",
            "step 268/5000, loss: 1.0553\n",
            "step 269/5000, loss: 1.1794\n",
            "step 270/5000, loss: 1.1439\n",
            "step 271/5000, loss: 1.1201\n",
            "step 272/5000, loss: 1.6120\n",
            "step 273/5000, loss: 1.4917\n",
            "step 274/5000, loss: 0.8694\n",
            "step 275/5000, loss: 0.9089\n",
            "step 276/5000, loss: 1.0154\n",
            "step 277/5000, loss: 0.9304\n",
            "step 278/5000, loss: 1.1192\n",
            "step 279/5000, loss: 1.0895\n",
            "step 280/5000, loss: 1.1010\n",
            "step 281/5000, loss: 1.1908\n",
            "step 282/5000, loss: 0.9933\n",
            "step 283/5000, loss: 1.1051\n",
            "step 284/5000, loss: 1.4595\n",
            "step 285/5000, loss: 1.1285\n",
            "step 286/5000, loss: 1.1151\n",
            "step 287/5000, loss: 0.9124\n",
            "step 288/5000, loss: 1.2737\n",
            "step 289/5000, loss: 1.8975\n",
            "step 290/5000, loss: 0.9865\n",
            "step 291/5000, loss: 1.4563\n",
            "step 292/5000, loss: 1.0831\n",
            "step 293/5000, loss: 1.0109\n",
            "step 294/5000, loss: 1.1740\n",
            "step 295/5000, loss: 1.0485\n",
            "step 296/5000, loss: 1.3992\n",
            "step 297/5000, loss: 1.0155\n",
            "step 298/5000, loss: 0.9952\n",
            "step 299/5000, loss: 1.1873\n",
            "step 300/5000, loss: 0.9988\n",
            "step 301/5000, loss: 0.9095\n",
            "step 302/5000, loss: 1.2094\n",
            "step 303/5000, loss: 1.0207\n",
            "step 304/5000, loss: 0.9707\n",
            "step 305/5000, loss: 1.0649\n",
            "step 306/5000, loss: 1.3208\n",
            "step 307/5000, loss: 1.8324\n",
            "step 308/5000, loss: 1.0303\n",
            "step 309/5000, loss: 1.0059\n",
            "step 310/5000, loss: 0.9538\n",
            "step 311/5000, loss: 1.0716\n",
            "step 312/5000, loss: 0.8798\n",
            "step 313/5000, loss: 1.0288\n",
            "step 314/5000, loss: 1.0549\n",
            "step 315/5000, loss: 1.0908\n",
            "step 316/5000, loss: 1.0209\n",
            "step 317/5000, loss: 1.0434\n",
            "step 318/5000, loss: 1.0742\n",
            "step 319/5000, loss: 0.8972\n",
            "step 320/5000, loss: 1.1560\n",
            "step 321/5000, loss: 1.0875\n",
            "step 322/5000, loss: 0.9536\n",
            "step 323/5000, loss: 1.0421\n",
            "step 324/5000, loss: 0.9552\n",
            "step 325/5000, loss: 1.0269\n",
            "step 326/5000, loss: 1.0868\n",
            "step 327/5000, loss: 1.2923\n",
            "step 328/5000, loss: 1.0202\n",
            "step 329/5000, loss: 0.8809\n",
            "step 330/5000, loss: 0.9496\n",
            "step 331/5000, loss: 1.2690\n",
            "step 332/5000, loss: 1.0050\n",
            "step 333/5000, loss: 0.9922\n",
            "step 334/5000, loss: 1.0527\n",
            "step 335/5000, loss: 1.0657\n",
            "step 336/5000, loss: 0.8971\n",
            "step 337/5000, loss: 0.7581\n",
            "step 338/5000, loss: 1.1136\n",
            "step 339/5000, loss: 1.0790\n",
            "step 340/5000, loss: 1.2766\n",
            "step 341/5000, loss: 1.1708\n",
            "step 342/5000, loss: 0.9661\n",
            "step 343/5000, loss: 1.0980\n",
            "step 344/5000, loss: 0.8266\n",
            "step 345/5000, loss: 1.0618\n",
            "step 346/5000, loss: 0.8770\n",
            "step 347/5000, loss: 1.0252\n",
            "step 348/5000, loss: 1.2393\n",
            "step 349/5000, loss: 1.1392\n",
            "step 350/5000, loss: 0.8919\n",
            "step 351/5000, loss: 1.1097\n",
            "step 352/5000, loss: 1.2155\n",
            "step 353/5000, loss: 0.9232\n",
            "step 354/5000, loss: 0.8176\n",
            "step 355/5000, loss: 0.8597\n",
            "step 356/5000, loss: 0.8602\n",
            "step 357/5000, loss: 0.9476\n",
            "step 358/5000, loss: 0.9548\n",
            "step 359/5000, loss: 0.9470\n",
            "step 360/5000, loss: 0.8922\n",
            "step 361/5000, loss: 1.6696\n",
            "step 362/5000, loss: 0.9645\n",
            "step 363/5000, loss: 1.1007\n",
            "step 364/5000, loss: 1.2130\n",
            "step 365/5000, loss: 1.1838\n",
            "step 366/5000, loss: 1.2117\n",
            "step 367/5000, loss: 0.8544\n",
            "step 368/5000, loss: 1.3504\n",
            "step 369/5000, loss: 1.5245\n",
            "step 370/5000, loss: 1.0436\n",
            "step 371/5000, loss: 1.1211\n",
            "step 372/5000, loss: 0.8700\n",
            "step 373/5000, loss: 0.9893\n",
            "step 374/5000, loss: 1.0096\n",
            "step 375/5000, loss: 0.8857\n",
            "step 376/5000, loss: 0.9338\n",
            "step 377/5000, loss: 1.1811\n",
            "step 378/5000, loss: 0.9121\n",
            "step 379/5000, loss: 1.0254\n",
            "step 380/5000, loss: 0.8977\n",
            "step 381/5000, loss: 1.0467\n",
            "step 382/5000, loss: 1.0728\n",
            "step 383/5000, loss: 0.8426\n",
            "step 384/5000, loss: 1.0520\n",
            "step 385/5000, loss: 1.1164\n",
            "step 386/5000, loss: 0.9284\n",
            "step 387/5000, loss: 1.1672\n",
            "step 388/5000, loss: 1.0701\n",
            "step 389/5000, loss: 1.1577\n",
            "step 390/5000, loss: 1.0344\n",
            "step 391/5000, loss: 1.0538\n",
            "step 392/5000, loss: 1.1017\n",
            "step 393/5000, loss: 1.0724\n",
            "step 394/5000, loss: 1.1773\n",
            "step 395/5000, loss: 0.8151\n",
            "step 396/5000, loss: 1.2837\n",
            "step 397/5000, loss: 1.2079\n",
            "step 398/5000, loss: 0.8315\n",
            "step 399/5000, loss: 1.1568\n",
            "step 400/5000, loss: 0.8731\n",
            "step 401/5000, loss: 0.9476\n",
            "step 402/5000, loss: 0.9169\n",
            "step 403/5000, loss: 1.1997\n",
            "step 404/5000, loss: 0.9547\n",
            "step 405/5000, loss: 0.9665\n",
            "step 406/5000, loss: 1.0585\n",
            "step 407/5000, loss: 1.5079\n",
            "step 408/5000, loss: 0.9357\n",
            "step 409/5000, loss: 1.3220\n",
            "step 410/5000, loss: 1.1300\n",
            "step 411/5000, loss: 0.8620\n",
            "step 412/5000, loss: 1.0666\n",
            "step 413/5000, loss: 0.9125\n",
            "step 414/5000, loss: 1.1796\n",
            "step 415/5000, loss: 0.9299\n",
            "step 416/5000, loss: 2.2346\n",
            "step 417/5000, loss: 0.7677\n",
            "step 418/5000, loss: 0.9976\n",
            "step 419/5000, loss: 1.0567\n",
            "step 420/5000, loss: 0.9294\n",
            "step 421/5000, loss: 1.0042\n",
            "step 422/5000, loss: 1.1390\n",
            "step 423/5000, loss: 0.8197\n",
            "step 424/5000, loss: 1.2663\n",
            "step 425/5000, loss: 1.0379\n",
            "step 426/5000, loss: 1.5496\n",
            "step 427/5000, loss: 1.0162\n",
            "step 428/5000, loss: 1.2080\n",
            "step 429/5000, loss: 1.1356\n",
            "step 430/5000, loss: 0.8547\n",
            "step 431/5000, loss: 1.1605\n",
            "step 432/5000, loss: 1.0663\n",
            "step 433/5000, loss: 0.8812\n",
            "step 434/5000, loss: 1.0188\n",
            "step 435/5000, loss: 1.0286\n",
            "step 436/5000, loss: 0.9706\n",
            "step 437/5000, loss: 0.9117\n",
            "step 438/5000, loss: 1.1385\n",
            "step 439/5000, loss: 1.0604\n",
            "step 440/5000, loss: 0.8858\n",
            "step 441/5000, loss: 0.8464\n",
            "step 442/5000, loss: 1.1227\n",
            "step 443/5000, loss: 0.8968\n",
            "step 444/5000, loss: 0.8659\n",
            "step 445/5000, loss: 0.8167\n",
            "step 446/5000, loss: 0.9450\n",
            "step 447/5000, loss: 0.9734\n",
            "step 448/5000, loss: 0.9445\n",
            "step 449/5000, loss: 0.9850\n",
            "step 450/5000, loss: 0.9297\n",
            "step 451/5000, loss: 0.9326\n",
            "step 452/5000, loss: 0.9909\n",
            "step 453/5000, loss: 0.9350\n",
            "step 454/5000, loss: 0.9519\n",
            "step 455/5000, loss: 1.0930\n",
            "step 456/5000, loss: 1.2694\n",
            "step 457/5000, loss: 1.0144\n",
            "step 458/5000, loss: 0.7262\n",
            "step 459/5000, loss: 1.1196\n",
            "step 460/5000, loss: 0.9163\n",
            "step 461/5000, loss: 0.8630\n",
            "step 462/5000, loss: 0.8532\n",
            "step 463/5000, loss: 0.7017\n",
            "step 464/5000, loss: 0.8426\n",
            "step 465/5000, loss: 0.7803\n",
            "step 466/5000, loss: 1.0552\n",
            "step 467/5000, loss: 0.8261\n",
            "step 468/5000, loss: 1.3251\n",
            "step 469/5000, loss: 0.9437\n",
            "step 470/5000, loss: 0.8210\n",
            "step 471/5000, loss: 1.0086\n",
            "step 472/5000, loss: 0.9278\n",
            "step 473/5000, loss: 1.0236\n",
            "step 474/5000, loss: 0.8800\n",
            "step 475/5000, loss: 0.8785\n",
            "step 476/5000, loss: 0.7741\n",
            "step 477/5000, loss: 0.9956\n",
            "step 478/5000, loss: 1.1449\n",
            "step 479/5000, loss: 0.9437\n",
            "step 480/5000, loss: 0.7760\n",
            "step 481/5000, loss: 1.3267\n",
            "step 482/5000, loss: 0.7705\n",
            "step 483/5000, loss: 0.8849\n",
            "step 484/5000, loss: 0.8313\n",
            "step 485/5000, loss: 0.6991\n",
            "step 486/5000, loss: 1.0604\n",
            "step 487/5000, loss: 0.7771\n",
            "step 488/5000, loss: 0.7676\n",
            "step 489/5000, loss: 0.8010\n",
            "step 490/5000, loss: 1.6640\n",
            "step 491/5000, loss: 0.8708\n",
            "step 492/5000, loss: 0.9195\n",
            "step 493/5000, loss: 0.7660\n",
            "step 494/5000, loss: 1.0636\n",
            "step 495/5000, loss: 0.9863\n",
            "step 496/5000, loss: 0.8153\n",
            "step 497/5000, loss: 1.1934\n",
            "step 498/5000, loss: 0.8857\n",
            "step 499/5000, loss: 0.9524\n",
            "step 500/5000, loss: 0.9089\n",
            "step 501/5000, loss: 0.8521\n",
            "step 502/5000, loss: 1.0693\n",
            "step 503/5000, loss: 1.0608\n",
            "step 504/5000, loss: 0.9465\n",
            "step 505/5000, loss: 0.9205\n",
            "step 506/5000, loss: 0.9086\n",
            "step 507/5000, loss: 0.8029\n",
            "step 508/5000, loss: 0.9471\n",
            "step 509/5000, loss: 1.1237\n",
            "step 510/5000, loss: 0.9929\n",
            "step 511/5000, loss: 0.9128\n",
            "step 512/5000, loss: 1.0030\n",
            "step 513/5000, loss: 1.1275\n",
            "step 514/5000, loss: 1.1809\n",
            "step 515/5000, loss: 1.1001\n",
            "step 516/5000, loss: 0.8676\n",
            "step 517/5000, loss: 0.8582\n",
            "step 518/5000, loss: 0.9935\n",
            "step 519/5000, loss: 1.0094\n",
            "step 520/5000, loss: 0.8656\n",
            "step 521/5000, loss: 1.5260\n",
            "step 522/5000, loss: 1.1261\n",
            "step 523/5000, loss: 0.7482\n",
            "step 524/5000, loss: 0.8675\n",
            "step 525/5000, loss: 0.8020\n",
            "step 526/5000, loss: 1.3159\n",
            "step 527/5000, loss: 1.1704\n",
            "step 528/5000, loss: 0.8437\n",
            "step 529/5000, loss: 0.8290\n",
            "step 530/5000, loss: 1.0059\n",
            "step 531/5000, loss: 1.1180\n",
            "step 532/5000, loss: 0.9368\n",
            "step 533/5000, loss: 0.8155\n",
            "step 534/5000, loss: 0.8033\n",
            "step 535/5000, loss: 0.8435\n",
            "step 536/5000, loss: 0.8935\n",
            "step 537/5000, loss: 1.0569\n",
            "step 538/5000, loss: 2.2317\n",
            "step 539/5000, loss: 0.9303\n",
            "step 540/5000, loss: 0.8209\n",
            "step 541/5000, loss: 0.8672\n",
            "step 542/5000, loss: 1.0063\n",
            "step 543/5000, loss: 0.9062\n",
            "step 544/5000, loss: 0.8638\n",
            "step 545/5000, loss: 0.8897\n",
            "step 546/5000, loss: 1.2258\n",
            "step 547/5000, loss: 0.9996\n",
            "step 548/5000, loss: 1.0124\n",
            "step 549/5000, loss: 1.3951\n",
            "step 550/5000, loss: 1.0588\n",
            "step 551/5000, loss: 0.8815\n",
            "step 552/5000, loss: 0.8367\n",
            "step 553/5000, loss: 0.7534\n",
            "step 554/5000, loss: 0.8882\n",
            "step 555/5000, loss: 1.1487\n",
            "step 556/5000, loss: 0.9729\n",
            "step 557/5000, loss: 0.9757\n",
            "step 558/5000, loss: 0.8850\n",
            "step 559/5000, loss: 1.2242\n",
            "step 560/5000, loss: 0.8449\n",
            "step 561/5000, loss: 0.8459\n",
            "step 562/5000, loss: 0.7479\n",
            "step 563/5000, loss: 1.1844\n",
            "step 564/5000, loss: 0.9017\n",
            "step 565/5000, loss: 0.9054\n",
            "step 566/5000, loss: 0.9104\n",
            "step 567/5000, loss: 0.9531\n",
            "step 568/5000, loss: 0.9607\n",
            "step 569/5000, loss: 1.3127\n",
            "step 570/5000, loss: 0.9044\n",
            "step 571/5000, loss: 0.8213\n",
            "step 572/5000, loss: 0.8914\n",
            "step 573/5000, loss: 1.0854\n",
            "step 574/5000, loss: 0.9313\n",
            "step 575/5000, loss: 1.0540\n",
            "step 576/5000, loss: 1.2074\n",
            "step 577/5000, loss: 1.3300\n",
            "step 578/5000, loss: 0.9113\n",
            "step 579/5000, loss: 1.0163\n",
            "step 580/5000, loss: 0.7648\n",
            "step 581/5000, loss: 1.0735\n",
            "step 582/5000, loss: 0.9196\n",
            "step 583/5000, loss: 0.9405\n",
            "step 584/5000, loss: 0.9629\n",
            "step 585/5000, loss: 1.0089\n",
            "step 586/5000, loss: 1.0799\n",
            "step 587/5000, loss: 0.7599\n",
            "step 588/5000, loss: 0.9684\n",
            "step 589/5000, loss: 0.7714\n",
            "step 590/5000, loss: 0.8935\n",
            "step 591/5000, loss: 0.8701\n",
            "step 592/5000, loss: 0.8737\n",
            "step 593/5000, loss: 0.8369\n",
            "step 594/5000, loss: 0.8979\n",
            "step 595/5000, loss: 0.8448\n",
            "step 596/5000, loss: 1.0064\n",
            "step 597/5000, loss: 1.1075\n",
            "step 598/5000, loss: 1.0274\n",
            "step 599/5000, loss: 0.8430\n",
            "step 600/5000, loss: 0.7373\n",
            "step 601/5000, loss: 0.9336\n",
            "step 602/5000, loss: 0.7523\n",
            "step 603/5000, loss: 0.8178\n",
            "step 604/5000, loss: 1.0891\n",
            "step 605/5000, loss: 0.7604\n",
            "step 606/5000, loss: 0.8805\n",
            "step 607/5000, loss: 0.8392\n",
            "step 608/5000, loss: 0.8244\n",
            "step 609/5000, loss: 0.8602\n",
            "step 610/5000, loss: 1.4766\n",
            "step 611/5000, loss: 0.8680\n",
            "step 612/5000, loss: 0.9547\n",
            "step 613/5000, loss: 0.8637\n",
            "step 614/5000, loss: 1.0516\n",
            "step 615/5000, loss: 1.0217\n",
            "step 616/5000, loss: 0.7173\n",
            "step 617/5000, loss: 0.9221\n",
            "step 618/5000, loss: 0.7877\n",
            "step 619/5000, loss: 0.9476\n",
            "step 620/5000, loss: 0.9606\n",
            "step 621/5000, loss: 0.7695\n",
            "step 622/5000, loss: 1.0103\n",
            "step 623/5000, loss: 0.9805\n",
            "step 624/5000, loss: 0.8578\n",
            "step 625/5000, loss: 0.9194\n",
            "step 626/5000, loss: 0.9664\n",
            "step 627/5000, loss: 0.8618\n",
            "step 628/5000, loss: 0.9523\n",
            "step 629/5000, loss: 0.9502\n",
            "step 630/5000, loss: 1.0408\n",
            "step 631/5000, loss: 0.9426\n",
            "step 632/5000, loss: 1.6674\n",
            "step 633/5000, loss: 0.9419\n",
            "step 634/5000, loss: 0.8559\n",
            "step 635/5000, loss: 0.9197\n",
            "step 636/5000, loss: 0.8931\n",
            "step 637/5000, loss: 0.7954\n",
            "step 638/5000, loss: 3.1593\n",
            "step 639/5000, loss: 0.9163\n",
            "step 640/5000, loss: 1.0780\n",
            "step 641/5000, loss: 0.9471\n",
            "step 642/5000, loss: 0.8598\n",
            "step 643/5000, loss: 0.7950\n",
            "step 644/5000, loss: 1.0602\n",
            "step 645/5000, loss: 0.9844\n",
            "step 646/5000, loss: 0.8351\n",
            "step 647/5000, loss: 0.9464\n",
            "step 648/5000, loss: 0.9951\n",
            "step 649/5000, loss: 0.7863\n",
            "step 650/5000, loss: 0.7608\n",
            "step 651/5000, loss: 0.9128\n",
            "step 652/5000, loss: 0.8085\n",
            "step 653/5000, loss: 0.9025\n",
            "step 654/5000, loss: 1.0732\n",
            "step 655/5000, loss: 0.8381\n",
            "step 656/5000, loss: 0.9509\n",
            "step 657/5000, loss: 1.0462\n",
            "step 658/5000, loss: 0.9226\n",
            "step 659/5000, loss: 0.7891\n",
            "step 660/5000, loss: 0.9335\n",
            "step 661/5000, loss: 0.8481\n",
            "step 662/5000, loss: 1.0312\n",
            "step 663/5000, loss: 0.8541\n",
            "step 664/5000, loss: 1.0219\n",
            "step 665/5000, loss: 0.7648\n",
            "step 666/5000, loss: 0.9272\n",
            "step 667/5000, loss: 0.8293\n",
            "step 668/5000, loss: 1.1871\n",
            "step 669/5000, loss: 1.2642\n",
            "step 670/5000, loss: 0.9564\n",
            "step 671/5000, loss: 0.9955\n",
            "step 672/5000, loss: 0.9470\n",
            "step 673/5000, loss: 1.0103\n",
            "step 674/5000, loss: 1.3978\n",
            "step 675/5000, loss: 0.8428\n",
            "step 676/5000, loss: 0.8251\n",
            "step 677/5000, loss: 0.8145\n",
            "step 678/5000, loss: 0.7585\n",
            "step 679/5000, loss: 0.7704\n",
            "step 680/5000, loss: 0.8725\n",
            "step 681/5000, loss: 0.9020\n",
            "step 682/5000, loss: 0.7701\n",
            "step 683/5000, loss: 0.9257\n",
            "step 684/5000, loss: 0.9135\n",
            "step 685/5000, loss: 1.1792\n",
            "step 686/5000, loss: 0.8309\n",
            "step 687/5000, loss: 0.8486\n",
            "step 688/5000, loss: 0.8835\n",
            "step 689/5000, loss: 0.9897\n",
            "step 690/5000, loss: 0.7642\n",
            "step 691/5000, loss: 0.9798\n",
            "step 692/5000, loss: 0.8759\n",
            "step 693/5000, loss: 0.7198\n",
            "step 694/5000, loss: 0.9663\n",
            "step 695/5000, loss: 0.9244\n",
            "step 696/5000, loss: 0.8838\n",
            "step 697/5000, loss: 1.2710\n",
            "step 698/5000, loss: 0.8258\n",
            "step 699/5000, loss: 0.8831\n",
            "step 700/5000, loss: 1.4723\n",
            "step 701/5000, loss: 0.9399\n",
            "step 702/5000, loss: 0.7435\n",
            "step 703/5000, loss: 0.8163\n",
            "step 704/5000, loss: 1.0356\n",
            "step 705/5000, loss: 0.8712\n",
            "step 706/5000, loss: 0.8578\n",
            "step 707/5000, loss: 1.6786\n",
            "step 708/5000, loss: 0.9422\n",
            "step 709/5000, loss: 1.0259\n",
            "step 710/5000, loss: 0.8638\n",
            "step 711/5000, loss: 1.1960\n",
            "step 712/5000, loss: 0.7440\n",
            "step 713/5000, loss: 0.8631\n",
            "step 714/5000, loss: 2.8364\n",
            "step 715/5000, loss: 0.8357\n",
            "step 716/5000, loss: 1.2169\n",
            "step 717/5000, loss: 1.0243\n",
            "step 718/5000, loss: 0.9545\n",
            "step 719/5000, loss: 0.8848\n",
            "step 720/5000, loss: 0.9080\n",
            "step 721/5000, loss: 0.9367\n",
            "step 722/5000, loss: 0.9189\n",
            "step 723/5000, loss: 4.3373\n",
            "step 724/5000, loss: 1.2671\n",
            "step 725/5000, loss: 0.7685\n",
            "step 726/5000, loss: 1.3503\n",
            "step 727/5000, loss: 0.7932\n",
            "step 728/5000, loss: 0.9554\n",
            "step 729/5000, loss: 0.8149\n",
            "step 730/5000, loss: 1.0468\n",
            "step 731/5000, loss: 1.4547\n",
            "step 732/5000, loss: 0.8217\n",
            "step 733/5000, loss: 0.9819\n",
            "step 734/5000, loss: 0.7978\n",
            "step 735/5000, loss: 0.9810\n",
            "step 736/5000, loss: 0.7462\n",
            "step 737/5000, loss: 0.7520\n",
            "step 738/5000, loss: 1.0876\n",
            "step 739/5000, loss: 0.8777\n",
            "step 740/5000, loss: 1.1427\n",
            "step 741/5000, loss: 0.8304\n",
            "step 742/5000, loss: 0.7114\n",
            "step 743/5000, loss: 0.8927\n",
            "step 744/5000, loss: 0.9211\n",
            "step 745/5000, loss: 0.9018\n",
            "step 746/5000, loss: 0.8548\n",
            "step 747/5000, loss: 0.8725\n",
            "step 748/5000, loss: 0.9336\n",
            "step 749/5000, loss: 0.8023\n",
            "step 750/5000, loss: 0.8684\n",
            "step 751/5000, loss: 0.9698\n",
            "step 752/5000, loss: 0.9345\n",
            "step 753/5000, loss: 0.9527\n",
            "step 754/5000, loss: 0.8740\n",
            "step 755/5000, loss: 0.9600\n",
            "step 756/5000, loss: 0.8928\n",
            "step 757/5000, loss: 0.7391\n",
            "step 758/5000, loss: 1.0750\n",
            "step 759/5000, loss: 1.0289\n",
            "step 760/5000, loss: 0.9042\n",
            "step 761/5000, loss: 1.0290\n",
            "step 762/5000, loss: 0.7821\n",
            "step 763/5000, loss: 0.9362\n",
            "step 764/5000, loss: 0.9075\n",
            "step 765/5000, loss: 1.0054\n",
            "step 766/5000, loss: 0.9662\n",
            "step 767/5000, loss: 0.9593\n",
            "step 768/5000, loss: 0.8308\n",
            "step 769/5000, loss: 0.9417\n",
            "step 770/5000, loss: 0.6889\n",
            "step 771/5000, loss: 0.9027\n",
            "step 772/5000, loss: 0.8686\n",
            "step 773/5000, loss: 0.7152\n",
            "step 774/5000, loss: 0.7918\n",
            "step 775/5000, loss: 0.8840\n",
            "step 776/5000, loss: 0.8409\n",
            "step 777/5000, loss: 1.0791\n",
            "step 778/5000, loss: 0.9594\n",
            "step 779/5000, loss: 1.1448\n",
            "step 780/5000, loss: 1.1060\n",
            "step 781/5000, loss: 1.0623\n",
            "step 782/5000, loss: 1.1418\n",
            "step 783/5000, loss: 0.9504\n",
            "step 784/5000, loss: 1.1767\n",
            "step 785/5000, loss: 0.9045\n",
            "step 786/5000, loss: 0.7686\n",
            "step 787/5000, loss: 0.9441\n",
            "step 788/5000, loss: 1.2305\n",
            "step 789/5000, loss: 0.7500\n",
            "step 790/5000, loss: 0.7779\n",
            "step 791/5000, loss: 0.8800\n",
            "step 792/5000, loss: 0.6956\n",
            "step 793/5000, loss: 1.0782\n",
            "step 794/5000, loss: 0.8552\n",
            "step 795/5000, loss: 0.7994\n",
            "step 796/5000, loss: 0.7810\n",
            "step 797/5000, loss: 0.9458\n",
            "step 798/5000, loss: 0.8100\n",
            "step 799/5000, loss: 1.1032\n",
            "step 800/5000, loss: 1.0934\n",
            "step 801/5000, loss: 1.1318\n",
            "step 802/5000, loss: 0.9678\n",
            "step 803/5000, loss: 0.9585\n",
            "step 804/5000, loss: 0.9727\n",
            "step 805/5000, loss: 1.0451\n",
            "step 806/5000, loss: 0.8322\n",
            "step 807/5000, loss: 0.9216\n",
            "step 808/5000, loss: 0.7910\n",
            "step 809/5000, loss: 0.8767\n",
            "step 810/5000, loss: 1.0195\n",
            "step 811/5000, loss: 1.0200\n",
            "step 812/5000, loss: 0.8780\n",
            "step 813/5000, loss: 0.7368\n",
            "step 814/5000, loss: 0.9520\n",
            "step 815/5000, loss: 1.0690\n",
            "step 816/5000, loss: 0.8569\n",
            "step 817/5000, loss: 0.9696\n",
            "step 818/5000, loss: 0.7350\n",
            "step 819/5000, loss: 0.7962\n",
            "step 820/5000, loss: 1.0357\n",
            "step 821/5000, loss: 1.0130\n",
            "step 822/5000, loss: 0.9049\n",
            "step 823/5000, loss: 0.7540\n",
            "step 824/5000, loss: 0.7712\n",
            "step 825/5000, loss: 1.0141\n",
            "step 826/5000, loss: 0.8691\n",
            "step 827/5000, loss: 0.9406\n",
            "step 828/5000, loss: 0.8306\n",
            "step 829/5000, loss: 0.9650\n",
            "step 830/5000, loss: 1.3953\n",
            "step 831/5000, loss: 1.0768\n",
            "step 832/5000, loss: 1.0374\n",
            "step 833/5000, loss: 0.7850\n",
            "step 834/5000, loss: 0.8914\n",
            "step 835/5000, loss: 1.0589\n",
            "step 836/5000, loss: 0.8113\n",
            "step 837/5000, loss: 0.9877\n",
            "step 838/5000, loss: 0.9397\n",
            "step 839/5000, loss: 0.8635\n",
            "step 840/5000, loss: 0.8518\n",
            "step 841/5000, loss: 0.8533\n",
            "step 842/5000, loss: 0.7529\n",
            "step 843/5000, loss: 0.7293\n",
            "step 844/5000, loss: 1.3794\n",
            "step 845/5000, loss: 0.9834\n",
            "step 846/5000, loss: 0.9881\n",
            "step 847/5000, loss: 0.8766\n",
            "step 848/5000, loss: 0.6824\n",
            "step 849/5000, loss: 0.8247\n",
            "step 850/5000, loss: 0.8872\n",
            "step 851/5000, loss: 0.7563\n",
            "step 852/5000, loss: 1.0627\n",
            "step 853/5000, loss: 0.7656\n",
            "step 854/5000, loss: 1.0376\n",
            "step 855/5000, loss: 0.8265\n",
            "step 856/5000, loss: 1.0945\n",
            "step 857/5000, loss: 1.0923\n",
            "step 858/5000, loss: 2.4440\n",
            "step 859/5000, loss: 0.9923\n",
            "step 860/5000, loss: 0.8865\n",
            "step 861/5000, loss: 0.8206\n",
            "step 862/5000, loss: 0.8646\n",
            "step 863/5000, loss: 1.1076\n",
            "step 864/5000, loss: 0.8352\n",
            "step 865/5000, loss: 0.8332\n",
            "step 866/5000, loss: 0.8561\n",
            "step 867/5000, loss: 0.9387\n",
            "step 868/5000, loss: 0.8853\n",
            "step 869/5000, loss: 1.1935\n",
            "step 870/5000, loss: 0.9050\n",
            "step 871/5000, loss: 1.0249\n",
            "step 872/5000, loss: 0.8474\n",
            "step 873/5000, loss: 0.8406\n",
            "step 874/5000, loss: 1.0118\n",
            "step 875/5000, loss: 0.9016\n",
            "step 876/5000, loss: 0.9812\n",
            "step 877/5000, loss: 0.8612\n",
            "step 878/5000, loss: 0.7609\n",
            "step 879/5000, loss: 1.2872\n",
            "step 880/5000, loss: 0.8608\n",
            "step 881/5000, loss: 0.7909\n",
            "step 882/5000, loss: 1.0007\n",
            "step 883/5000, loss: 0.8822\n",
            "step 884/5000, loss: 0.8984\n",
            "step 885/5000, loss: 0.7777\n",
            "step 886/5000, loss: 0.7779\n",
            "step 887/5000, loss: 0.9079\n",
            "step 888/5000, loss: 0.9546\n",
            "step 889/5000, loss: 0.8659\n",
            "step 890/5000, loss: 0.7884\n",
            "step 891/5000, loss: 0.8108\n",
            "step 892/5000, loss: 0.7698\n",
            "step 893/5000, loss: 1.0135\n",
            "step 894/5000, loss: 0.7801\n",
            "step 895/5000, loss: 1.0664\n",
            "step 896/5000, loss: 0.8638\n",
            "step 897/5000, loss: 0.9943\n",
            "step 898/5000, loss: 0.7793\n",
            "step 899/5000, loss: 0.9833\n",
            "step 900/5000, loss: 0.8493\n",
            "step 901/5000, loss: 0.8676\n",
            "step 902/5000, loss: 0.7427\n",
            "step 903/5000, loss: 0.9338\n",
            "step 904/5000, loss: 0.7728\n",
            "step 905/5000, loss: 1.0911\n",
            "step 906/5000, loss: 0.9621\n",
            "step 907/5000, loss: 1.0910\n",
            "step 908/5000, loss: 0.8333\n",
            "step 909/5000, loss: 0.9455\n",
            "step 910/5000, loss: 0.9761\n",
            "step 911/5000, loss: 0.7954\n",
            "step 912/5000, loss: 0.8427\n",
            "step 913/5000, loss: 0.8552\n",
            "step 914/5000, loss: 1.0206\n",
            "step 915/5000, loss: 0.8207\n",
            "step 916/5000, loss: 1.0486\n",
            "step 917/5000, loss: 0.7902\n",
            "step 918/5000, loss: 1.0564\n",
            "step 919/5000, loss: 0.8616\n",
            "step 920/5000, loss: 0.9927\n",
            "step 921/5000, loss: 0.8789\n",
            "step 922/5000, loss: 0.7751\n",
            "step 923/5000, loss: 0.8968\n",
            "step 924/5000, loss: 0.9171\n",
            "step 925/5000, loss: 0.9648\n",
            "step 926/5000, loss: 0.7294\n",
            "step 927/5000, loss: 0.8273\n",
            "step 928/5000, loss: 2.4404\n",
            "step 929/5000, loss: 0.8569\n",
            "step 930/5000, loss: 0.9282\n",
            "step 931/5000, loss: 0.8819\n",
            "step 932/5000, loss: 0.8492\n",
            "step 933/5000, loss: 0.9236\n",
            "step 934/5000, loss: 0.9452\n",
            "step 935/5000, loss: 0.7743\n",
            "step 936/5000, loss: 0.8317\n",
            "step 937/5000, loss: 0.7321\n",
            "step 938/5000, loss: 0.7937\n",
            "step 939/5000, loss: 0.9776\n",
            "step 940/5000, loss: 0.9157\n",
            "step 941/5000, loss: 0.9012\n",
            "step 942/5000, loss: 0.7358\n",
            "step 943/5000, loss: 0.8532\n",
            "step 944/5000, loss: 0.7911\n",
            "step 945/5000, loss: 0.9822\n",
            "step 946/5000, loss: 0.8004\n",
            "step 947/5000, loss: 0.6751\n",
            "step 948/5000, loss: 0.9592\n",
            "step 949/5000, loss: 0.9522\n",
            "step 950/5000, loss: 0.9780\n",
            "step 951/5000, loss: 0.7381\n",
            "step 952/5000, loss: 0.9543\n",
            "step 953/5000, loss: 0.7794\n",
            "step 954/5000, loss: 0.7128\n",
            "step 955/5000, loss: 0.8288\n",
            "step 956/5000, loss: 0.8206\n",
            "step 957/5000, loss: 0.8721\n",
            "step 958/5000, loss: 0.8163\n",
            "step 959/5000, loss: 0.9450\n",
            "step 960/5000, loss: 0.7705\n",
            "step 961/5000, loss: 0.8672\n",
            "step 962/5000, loss: 0.8626\n",
            "step 963/5000, loss: 0.7323\n",
            "step 964/5000, loss: 0.9707\n",
            "step 965/5000, loss: 1.0998\n",
            "step 966/5000, loss: 0.8541\n",
            "step 967/5000, loss: 0.8700\n",
            "step 968/5000, loss: 0.7249\n",
            "step 969/5000, loss: 0.8455\n",
            "step 970/5000, loss: 0.7642\n",
            "step 971/5000, loss: 0.7686\n",
            "step 972/5000, loss: 1.0084\n",
            "step 973/5000, loss: 0.9631\n",
            "step 974/5000, loss: 0.7911\n",
            "step 975/5000, loss: 0.8532\n",
            "step 976/5000, loss: 0.7388\n",
            "step 977/5000, loss: 0.6668\n",
            "step 978/5000, loss: 0.8687\n",
            "step 979/5000, loss: 0.9425\n",
            "step 980/5000, loss: 0.8695\n",
            "step 981/5000, loss: 0.8108\n",
            "step 982/5000, loss: 0.7795\n",
            "step 983/5000, loss: 0.9249\n",
            "step 984/5000, loss: 1.1064\n",
            "step 985/5000, loss: 0.8620\n",
            "step 986/5000, loss: 0.7829\n",
            "step 987/5000, loss: 0.8444\n",
            "step 988/5000, loss: 0.9497\n",
            "step 989/5000, loss: 0.7685\n",
            "step 990/5000, loss: 0.7750\n",
            "step 991/5000, loss: 1.0179\n",
            "step 992/5000, loss: 1.0034\n",
            "step 993/5000, loss: 0.8642\n",
            "step 994/5000, loss: 0.9523\n",
            "step 995/5000, loss: 1.8221\n",
            "step 996/5000, loss: 0.7818\n",
            "step 997/5000, loss: 0.8162\n",
            "step 998/5000, loss: 0.7482\n",
            "step 999/5000, loss: 0.7562\n",
            "step 1000/5000, loss: 0.8973\n",
            "step 1001/5000, loss: 0.7382\n",
            "step 1002/5000, loss: 0.7760\n",
            "step 1003/5000, loss: 0.8312\n",
            "step 1004/5000, loss: 1.2784\n",
            "step 1005/5000, loss: 0.9139\n",
            "step 1006/5000, loss: 1.0322\n",
            "step 1007/5000, loss: 0.8332\n",
            "step 1008/5000, loss: 0.8251\n",
            "step 1009/5000, loss: 0.7771\n",
            "step 1010/5000, loss: 0.8932\n",
            "step 1011/5000, loss: 0.7785\n",
            "step 1012/5000, loss: 0.8486\n",
            "step 1013/5000, loss: 0.7869\n",
            "step 1014/5000, loss: 0.6799\n",
            "step 1015/5000, loss: 0.8744\n",
            "step 1016/5000, loss: 1.0182\n",
            "step 1017/5000, loss: 0.8475\n",
            "step 1018/5000, loss: 0.8706\n",
            "step 1019/5000, loss: 0.8537\n",
            "step 1020/5000, loss: 0.8892\n",
            "step 1021/5000, loss: 0.8259\n",
            "step 1022/5000, loss: 0.7514\n",
            "step 1023/5000, loss: 0.8786\n",
            "step 1024/5000, loss: 0.7951\n",
            "step 1025/5000, loss: 0.6999\n",
            "step 1026/5000, loss: 0.7654\n",
            "step 1027/5000, loss: 0.9946\n",
            "step 1028/5000, loss: 0.7731\n",
            "step 1029/5000, loss: 0.8142\n",
            "step 1030/5000, loss: 0.8197\n",
            "step 1031/5000, loss: 0.8821\n",
            "step 1032/5000, loss: 0.7563\n",
            "step 1033/5000, loss: 0.7723\n",
            "step 1034/5000, loss: 0.8046\n",
            "step 1035/5000, loss: 0.6941\n",
            "step 1036/5000, loss: 0.7063\n",
            "step 1037/5000, loss: 0.8179\n",
            "step 1038/5000, loss: 0.8135\n",
            "step 1039/5000, loss: 0.8660\n",
            "step 1040/5000, loss: 0.8835\n",
            "step 1041/5000, loss: 0.9497\n",
            "step 1042/5000, loss: 0.8341\n",
            "step 1043/5000, loss: 0.8617\n",
            "step 1044/5000, loss: 0.9419\n",
            "step 1045/5000, loss: 0.8865\n",
            "step 1046/5000, loss: 0.7792\n",
            "step 1047/5000, loss: 1.0337\n",
            "step 1048/5000, loss: 0.8424\n",
            "step 1049/5000, loss: 0.8481\n",
            "step 1050/5000, loss: 0.8789\n",
            "step 1051/5000, loss: 0.8570\n",
            "step 1052/5000, loss: 0.9672\n",
            "step 1053/5000, loss: 0.8342\n",
            "step 1054/5000, loss: 3.4715\n",
            "step 1055/5000, loss: 0.8029\n",
            "step 1056/5000, loss: 1.0082\n",
            "step 1057/5000, loss: 0.8969\n",
            "step 1058/5000, loss: 0.7185\n",
            "step 1059/5000, loss: 0.6399\n",
            "step 1060/5000, loss: 0.8792\n",
            "step 1061/5000, loss: 0.8976\n",
            "step 1062/5000, loss: 1.1443\n",
            "step 1063/5000, loss: 0.7833\n",
            "step 1064/5000, loss: 1.1216\n",
            "step 1065/5000, loss: 1.1812\n",
            "step 1066/5000, loss: 1.2090\n",
            "step 1067/5000, loss: 0.8302\n",
            "step 1068/5000, loss: 0.8091\n",
            "step 1069/5000, loss: 0.8166\n",
            "step 1070/5000, loss: 0.7121\n",
            "step 1071/5000, loss: 0.6376\n",
            "step 1072/5000, loss: 0.7790\n",
            "step 1073/5000, loss: 0.8084\n",
            "step 1074/5000, loss: 0.8984\n",
            "step 1075/5000, loss: 0.8356\n",
            "step 1076/5000, loss: 0.8032\n",
            "step 1077/5000, loss: 1.2461\n",
            "step 1078/5000, loss: 0.8521\n",
            "step 1079/5000, loss: 0.7870\n",
            "step 1080/5000, loss: 0.9064\n",
            "step 1081/5000, loss: 0.7380\n",
            "step 1082/5000, loss: 0.9135\n",
            "step 1083/5000, loss: 0.8361\n",
            "step 1084/5000, loss: 0.9126\n",
            "step 1085/5000, loss: 1.0237\n",
            "step 1086/5000, loss: 0.7709\n",
            "step 1087/5000, loss: 0.8314\n",
            "step 1088/5000, loss: 0.7818\n",
            "step 1089/5000, loss: 0.9588\n",
            "step 1090/5000, loss: 0.8012\n",
            "step 1091/5000, loss: 0.8716\n",
            "step 1092/5000, loss: 0.8091\n",
            "step 1093/5000, loss: 0.6467\n",
            "step 1094/5000, loss: 0.8082\n",
            "step 1095/5000, loss: 0.8135\n",
            "step 1096/5000, loss: 0.7616\n",
            "step 1097/5000, loss: 0.7266\n",
            "step 1098/5000, loss: 0.9018\n",
            "step 1099/5000, loss: 0.8670\n",
            "step 1100/5000, loss: 0.8374\n",
            "step 1101/5000, loss: 0.6689\n",
            "step 1102/5000, loss: 0.7693\n",
            "step 1103/5000, loss: 1.0058\n",
            "step 1104/5000, loss: 0.8221\n",
            "step 1105/5000, loss: 0.8713\n",
            "step 1106/5000, loss: 0.9355\n",
            "step 1107/5000, loss: 0.9482\n",
            "step 1108/5000, loss: 1.0001\n",
            "step 1109/5000, loss: 0.9238\n",
            "step 1110/5000, loss: 0.9453\n",
            "step 1111/5000, loss: 0.7417\n",
            "step 1112/5000, loss: 0.7417\n",
            "step 1113/5000, loss: 1.3170\n",
            "step 1114/5000, loss: 0.6505\n",
            "step 1115/5000, loss: 1.1919\n",
            "step 1116/5000, loss: 0.7011\n",
            "step 1117/5000, loss: 0.8024\n",
            "step 1118/5000, loss: 0.8991\n",
            "step 1119/5000, loss: 0.9583\n",
            "step 1120/5000, loss: 0.7743\n",
            "step 1121/5000, loss: 0.7791\n",
            "step 1122/5000, loss: 0.8391\n",
            "step 1123/5000, loss: 0.9875\n",
            "step 1124/5000, loss: 0.6982\n",
            "step 1125/5000, loss: 1.0145\n",
            "step 1126/5000, loss: 0.8111\n",
            "step 1127/5000, loss: 0.9255\n",
            "step 1128/5000, loss: 0.7458\n",
            "step 1129/5000, loss: 0.9267\n",
            "step 1130/5000, loss: 0.9001\n",
            "step 1131/5000, loss: 0.7202\n",
            "step 1132/5000, loss: 0.7554\n",
            "step 1133/5000, loss: 0.8689\n",
            "step 1134/5000, loss: 0.7132\n",
            "step 1135/5000, loss: 0.7093\n",
            "step 1136/5000, loss: 0.6623\n",
            "step 1137/5000, loss: 0.9030\n",
            "step 1138/5000, loss: 0.7703\n",
            "step 1139/5000, loss: 0.7602\n",
            "step 1140/5000, loss: 0.7152\n",
            "step 1141/5000, loss: 0.9546\n",
            "step 1142/5000, loss: 1.0425\n",
            "step 1143/5000, loss: 0.9301\n",
            "step 1144/5000, loss: 0.9193\n",
            "step 1145/5000, loss: 0.8047\n",
            "step 1146/5000, loss: 0.7324\n",
            "step 1147/5000, loss: 0.6756\n",
            "step 1148/5000, loss: 0.8279\n",
            "step 1149/5000, loss: 0.7389\n",
            "step 1150/5000, loss: 0.9363\n",
            "step 1151/5000, loss: 0.8775\n",
            "step 1152/5000, loss: 0.9354\n",
            "step 1153/5000, loss: 0.7319\n",
            "step 1154/5000, loss: 0.9534\n",
            "step 1155/5000, loss: 0.6477\n",
            "step 1156/5000, loss: 0.8180\n",
            "step 1157/5000, loss: 1.3218\n",
            "step 1158/5000, loss: 0.7700\n",
            "step 1159/5000, loss: 0.7585\n",
            "step 1160/5000, loss: 0.7947\n",
            "step 1161/5000, loss: 0.7930\n",
            "step 1162/5000, loss: 0.7607\n",
            "step 1163/5000, loss: 0.9137\n",
            "step 1164/5000, loss: 0.8581\n",
            "step 1165/5000, loss: 0.7462\n",
            "step 1166/5000, loss: 0.7553\n",
            "step 1167/5000, loss: 0.9860\n",
            "step 1168/5000, loss: 0.8903\n",
            "step 1169/5000, loss: 0.9989\n",
            "step 1170/5000, loss: 0.8131\n",
            "step 1171/5000, loss: 0.7880\n",
            "step 1172/5000, loss: 0.9493\n",
            "step 1173/5000, loss: 0.8262\n",
            "step 1174/5000, loss: 0.8632\n",
            "step 1175/5000, loss: 0.8654\n",
            "step 1176/5000, loss: 0.9021\n",
            "step 1177/5000, loss: 0.8852\n",
            "step 1178/5000, loss: 0.7391\n",
            "step 1179/5000, loss: 0.8314\n",
            "step 1180/5000, loss: 0.7520\n",
            "step 1181/5000, loss: 0.7288\n",
            "step 1182/5000, loss: 0.7861\n",
            "step 1183/5000, loss: 0.8739\n",
            "step 1184/5000, loss: 0.9376\n",
            "step 1185/5000, loss: 0.8206\n",
            "step 1186/5000, loss: 0.8620\n",
            "step 1187/5000, loss: 1.0603\n",
            "step 1188/5000, loss: 0.8422\n",
            "step 1189/5000, loss: 0.7867\n",
            "step 1190/5000, loss: 0.6980\n",
            "step 1191/5000, loss: 0.8290\n",
            "step 1192/5000, loss: 0.8749\n",
            "step 1193/5000, loss: 0.9890\n",
            "step 1194/5000, loss: 1.0673\n",
            "step 1195/5000, loss: 0.7058\n",
            "step 1196/5000, loss: 0.8693\n",
            "step 1197/5000, loss: 0.8736\n",
            "step 1198/5000, loss: 0.9557\n",
            "step 1199/5000, loss: 0.6701\n",
            "step 1200/5000, loss: 1.2321\n",
            "step 1201/5000, loss: 0.7768\n",
            "step 1202/5000, loss: 0.8234\n",
            "step 1203/5000, loss: 0.8823\n",
            "step 1204/5000, loss: 0.7393\n",
            "step 1205/5000, loss: 0.8397\n",
            "step 1206/5000, loss: 0.9140\n",
            "step 1207/5000, loss: 0.8800\n",
            "step 1208/5000, loss: 0.8900\n",
            "step 1209/5000, loss: 0.9526\n",
            "step 1210/5000, loss: 0.8803\n",
            "step 1211/5000, loss: 0.7756\n",
            "step 1212/5000, loss: 0.7552\n",
            "step 1213/5000, loss: 0.9983\n",
            "step 1214/5000, loss: 0.8223\n",
            "step 1215/5000, loss: 0.9047\n",
            "step 1216/5000, loss: 0.9607\n",
            "step 1217/5000, loss: 1.3780\n",
            "step 1218/5000, loss: 0.8268\n",
            "step 1219/5000, loss: 0.6852\n",
            "step 1220/5000, loss: 1.1015\n",
            "step 1221/5000, loss: 0.8830\n",
            "step 1222/5000, loss: 0.7461\n",
            "step 1223/5000, loss: 0.8975\n",
            "step 1224/5000, loss: 1.1195\n",
            "step 1225/5000, loss: 0.8548\n",
            "step 1226/5000, loss: 0.7849\n",
            "step 1227/5000, loss: 0.9847\n",
            "step 1228/5000, loss: 1.1187\n",
            "step 1229/5000, loss: 0.9370\n",
            "step 1230/5000, loss: 0.7923\n",
            "step 1231/5000, loss: 0.8328\n",
            "step 1232/5000, loss: 0.7207\n",
            "step 1233/5000, loss: 0.8406\n",
            "step 1234/5000, loss: 0.8262\n",
            "step 1235/5000, loss: 0.8568\n",
            "step 1236/5000, loss: 0.8394\n",
            "step 1237/5000, loss: 0.8735\n",
            "step 1238/5000, loss: 0.7730\n",
            "step 1239/5000, loss: 0.8601\n",
            "step 1240/5000, loss: 1.0908\n",
            "step 1241/5000, loss: 0.9410\n",
            "step 1242/5000, loss: 0.7202\n",
            "step 1243/5000, loss: 0.7252\n",
            "step 1244/5000, loss: 0.9000\n",
            "step 1245/5000, loss: 0.9997\n",
            "step 1246/5000, loss: 0.8066\n",
            "step 1247/5000, loss: 0.7458\n",
            "step 1248/5000, loss: 0.8162\n",
            "step 1249/5000, loss: 0.7156\n",
            "step 1250/5000, loss: 1.0519\n",
            "step 1251/5000, loss: 0.9566\n",
            "step 1252/5000, loss: 0.8574\n",
            "step 1253/5000, loss: 0.8552\n",
            "step 1254/5000, loss: 0.7459\n",
            "step 1255/5000, loss: 0.9368\n",
            "step 1256/5000, loss: 0.9250\n",
            "step 1257/5000, loss: 1.0921\n",
            "step 1258/5000, loss: 0.8210\n",
            "step 1259/5000, loss: 0.8357\n",
            "step 1260/5000, loss: 0.7566\n",
            "step 1261/5000, loss: 0.9423\n",
            "step 1262/5000, loss: 0.6843\n",
            "step 1263/5000, loss: 0.8243\n",
            "step 1264/5000, loss: 0.8242\n",
            "step 1265/5000, loss: 0.6503\n",
            "step 1266/5000, loss: 0.8647\n",
            "step 1267/5000, loss: 0.8949\n",
            "step 1268/5000, loss: 0.7512\n",
            "step 1269/5000, loss: 0.7099\n",
            "step 1270/5000, loss: 0.8389\n",
            "step 1271/5000, loss: 0.9960\n",
            "step 1272/5000, loss: 0.8838\n",
            "step 1273/5000, loss: 1.0526\n",
            "step 1274/5000, loss: 0.9707\n",
            "step 1275/5000, loss: 0.7393\n",
            "step 1276/5000, loss: 0.8598\n",
            "step 1277/5000, loss: 0.7957\n",
            "step 1278/5000, loss: 1.0376\n",
            "step 1279/5000, loss: 0.8829\n",
            "step 1280/5000, loss: 0.8614\n",
            "step 1281/5000, loss: 2.6022\n",
            "step 1282/5000, loss: 0.8969\n",
            "step 1283/5000, loss: 0.8356\n",
            "step 1284/5000, loss: 0.8538\n",
            "step 1285/5000, loss: 2.2328\n",
            "step 1286/5000, loss: 0.9470\n",
            "step 1287/5000, loss: 1.1211\n",
            "step 1288/5000, loss: 0.9579\n",
            "step 1289/5000, loss: 0.8336\n",
            "step 1290/5000, loss: 0.9156\n",
            "step 1291/5000, loss: 1.4442\n",
            "step 1292/5000, loss: 0.6543\n",
            "step 1293/5000, loss: 0.9349\n",
            "step 1294/5000, loss: 0.7547\n",
            "step 1295/5000, loss: 1.0100\n",
            "step 1296/5000, loss: 0.8758\n",
            "step 1297/5000, loss: 0.7748\n",
            "step 1298/5000, loss: 0.9332\n",
            "step 1299/5000, loss: 0.8018\n",
            "step 1300/5000, loss: 0.7786\n",
            "step 1301/5000, loss: 1.0000\n",
            "step 1302/5000, loss: 0.7716\n",
            "step 1303/5000, loss: 0.7506\n",
            "step 1304/5000, loss: 0.7813\n",
            "step 1305/5000, loss: 0.8374\n",
            "step 1306/5000, loss: 0.8158\n",
            "step 1307/5000, loss: 0.8386\n",
            "step 1308/5000, loss: 0.8408\n",
            "step 1309/5000, loss: 0.9554\n",
            "step 1310/5000, loss: 0.9892\n",
            "step 1311/5000, loss: 1.3605\n",
            "step 1312/5000, loss: 0.9448\n",
            "step 1313/5000, loss: 0.9181\n",
            "step 1314/5000, loss: 0.8593\n",
            "step 1315/5000, loss: 0.8766\n",
            "step 1316/5000, loss: 0.6556\n",
            "step 1317/5000, loss: 0.8619\n",
            "step 1318/5000, loss: 0.7329\n",
            "step 1319/5000, loss: 0.8182\n",
            "step 1320/5000, loss: 0.6560\n",
            "step 1321/5000, loss: 0.8512\n",
            "step 1322/5000, loss: 0.8398\n",
            "step 1323/5000, loss: 0.8061\n",
            "step 1324/5000, loss: 0.7405\n",
            "step 1325/5000, loss: 1.1026\n",
            "step 1326/5000, loss: 0.8833\n",
            "step 1327/5000, loss: 0.8535\n",
            "step 1328/5000, loss: 0.8903\n",
            "step 1329/5000, loss: 0.9690\n",
            "step 1330/5000, loss: 0.8003\n",
            "step 1331/5000, loss: 0.9230\n",
            "step 1332/5000, loss: 0.9876\n",
            "step 1333/5000, loss: 0.8686\n",
            "step 1334/5000, loss: 0.8171\n",
            "step 1335/5000, loss: 0.8170\n",
            "step 1336/5000, loss: 0.8050\n",
            "step 1337/5000, loss: 0.7782\n",
            "step 1338/5000, loss: 0.9114\n",
            "step 1339/5000, loss: 0.8364\n",
            "step 1340/5000, loss: 0.9796\n",
            "step 1341/5000, loss: 1.0763\n",
            "step 1342/5000, loss: 0.7059\n",
            "step 1343/5000, loss: 0.8252\n",
            "step 1344/5000, loss: 0.9260\n",
            "step 1345/5000, loss: 0.7466\n",
            "step 1346/5000, loss: 0.8178\n",
            "step 1347/5000, loss: 0.7690\n",
            "step 1348/5000, loss: 0.8256\n",
            "step 1349/5000, loss: 0.8251\n",
            "step 1350/5000, loss: 1.0730\n",
            "step 1351/5000, loss: 0.9633\n",
            "step 1352/5000, loss: 0.7268\n",
            "step 1353/5000, loss: 0.7371\n",
            "step 1354/5000, loss: 0.8686\n",
            "step 1355/5000, loss: 0.7599\n",
            "step 1356/5000, loss: 0.9184\n",
            "step 1357/5000, loss: 0.8725\n",
            "step 1358/5000, loss: 0.6984\n",
            "step 1359/5000, loss: 0.7963\n",
            "step 1360/5000, loss: 0.6471\n",
            "step 1361/5000, loss: 0.7782\n",
            "step 1362/5000, loss: 0.9676\n",
            "step 1363/5000, loss: 0.7961\n",
            "step 1364/5000, loss: 0.6437\n",
            "step 1365/5000, loss: 0.7102\n",
            "step 1366/5000, loss: 0.6629\n",
            "step 1367/5000, loss: 0.7347\n",
            "step 1368/5000, loss: 0.7272\n",
            "step 1369/5000, loss: 0.7900\n",
            "step 1370/5000, loss: 0.9063\n",
            "step 1371/5000, loss: 1.0331\n",
            "step 1372/5000, loss: 0.8731\n",
            "step 1373/5000, loss: 0.7417\n",
            "step 1374/5000, loss: 0.6614\n",
            "step 1375/5000, loss: 0.8543\n",
            "step 1376/5000, loss: 0.7133\n",
            "step 1377/5000, loss: 0.6363\n",
            "step 1378/5000, loss: 0.7660\n",
            "step 1379/5000, loss: 0.9485\n",
            "step 1380/5000, loss: 0.8751\n",
            "step 1381/5000, loss: 0.9922\n",
            "step 1382/5000, loss: 0.7570\n",
            "step 1383/5000, loss: 0.8146\n",
            "step 1384/5000, loss: 0.8775\n",
            "step 1385/5000, loss: 0.9873\n",
            "step 1386/5000, loss: 0.8257\n",
            "step 1387/5000, loss: 0.7918\n",
            "step 1388/5000, loss: 0.9526\n",
            "step 1389/5000, loss: 0.7612\n",
            "step 1390/5000, loss: 0.7935\n",
            "step 1391/5000, loss: 0.8019\n",
            "step 1392/5000, loss: 0.7063\n",
            "step 1393/5000, loss: 0.8063\n",
            "step 1394/5000, loss: 0.8741\n",
            "step 1395/5000, loss: 0.8379\n",
            "step 1396/5000, loss: 0.7680\n",
            "step 1397/5000, loss: 0.6488\n",
            "step 1398/5000, loss: 0.9347\n",
            "step 1399/5000, loss: 0.8256\n",
            "step 1400/5000, loss: 0.9094\n",
            "step 1401/5000, loss: 0.7909\n",
            "step 1402/5000, loss: 0.8330\n",
            "step 1403/5000, loss: 0.8902\n",
            "step 1404/5000, loss: 0.9048\n",
            "step 1405/5000, loss: 0.7030\n",
            "step 1406/5000, loss: 0.7469\n",
            "step 1407/5000, loss: 0.6320\n",
            "step 1408/5000, loss: 0.8217\n",
            "step 1409/5000, loss: 0.7721\n",
            "step 1410/5000, loss: 0.7211\n",
            "step 1411/5000, loss: 0.8207\n",
            "step 1412/5000, loss: 0.9140\n",
            "step 1413/5000, loss: 0.6668\n",
            "step 1414/5000, loss: 0.9744\n",
            "step 1415/5000, loss: 0.7731\n",
            "step 1416/5000, loss: 0.6810\n",
            "step 1417/5000, loss: 0.6377\n",
            "step 1418/5000, loss: 0.8331\n",
            "step 1419/5000, loss: 0.9359\n",
            "step 1420/5000, loss: 0.8192\n",
            "step 1421/5000, loss: 0.8217\n",
            "step 1422/5000, loss: 0.6678\n",
            "step 1423/5000, loss: 0.8245\n",
            "step 1424/5000, loss: 0.7414\n",
            "step 1425/5000, loss: 0.8119\n",
            "step 1426/5000, loss: 0.8510\n",
            "step 1427/5000, loss: 0.7151\n",
            "step 1428/5000, loss: 0.9004\n",
            "step 1429/5000, loss: 0.7902\n",
            "step 1430/5000, loss: 0.8655\n",
            "step 1431/5000, loss: 0.8631\n",
            "step 1432/5000, loss: 0.9650\n",
            "step 1433/5000, loss: 0.7113\n",
            "step 1434/5000, loss: 1.2216\n",
            "step 1435/5000, loss: 0.7111\n",
            "step 1436/5000, loss: 0.7033\n",
            "step 1437/5000, loss: 0.7691\n",
            "step 1438/5000, loss: 0.9775\n",
            "step 1439/5000, loss: 0.6868\n",
            "step 1440/5000, loss: 0.7541\n",
            "step 1441/5000, loss: 0.8312\n",
            "step 1442/5000, loss: 0.8474\n",
            "step 1443/5000, loss: 0.8115\n",
            "step 1444/5000, loss: 0.7572\n",
            "step 1445/5000, loss: 0.8084\n",
            "step 1446/5000, loss: 0.7503\n",
            "step 1447/5000, loss: 0.8189\n",
            "step 1448/5000, loss: 0.6585\n",
            "step 1449/5000, loss: 0.6743\n",
            "step 1450/5000, loss: 2.0190\n",
            "step 1451/5000, loss: 0.7093\n",
            "step 1452/5000, loss: 0.7622\n",
            "step 1453/5000, loss: 0.7996\n",
            "step 1454/5000, loss: 0.8533\n",
            "step 1455/5000, loss: 0.7439\n",
            "step 1456/5000, loss: 0.8635\n",
            "step 1457/5000, loss: 0.7848\n",
            "step 1458/5000, loss: 0.7095\n",
            "step 1459/5000, loss: 0.7323\n",
            "step 1460/5000, loss: 0.6965\n",
            "step 1461/5000, loss: 0.7857\n",
            "step 1462/5000, loss: 0.6571\n",
            "step 1463/5000, loss: 0.8138\n",
            "step 1464/5000, loss: 1.1886\n",
            "step 1465/5000, loss: 0.8078\n",
            "step 1466/5000, loss: 0.9508\n",
            "step 1467/5000, loss: 0.6431\n",
            "step 1468/5000, loss: 0.6807\n",
            "step 1469/5000, loss: 0.7779\n",
            "step 1470/5000, loss: 0.6992\n",
            "step 1471/5000, loss: 0.7172\n",
            "step 1472/5000, loss: 0.7551\n",
            "step 1473/5000, loss: 0.7595\n",
            "step 1474/5000, loss: 0.8209\n",
            "step 1475/5000, loss: 0.7396\n",
            "step 1476/5000, loss: 0.8619\n",
            "step 1477/5000, loss: 0.7348\n",
            "step 1478/5000, loss: 0.9019\n",
            "step 1479/5000, loss: 0.6981\n",
            "step 1480/5000, loss: 0.8005\n",
            "step 1481/5000, loss: 0.9238\n",
            "step 1482/5000, loss: 0.7398\n",
            "step 1483/5000, loss: 0.9163\n",
            "step 1484/5000, loss: 0.9155\n",
            "step 1485/5000, loss: 0.9909\n",
            "step 1486/5000, loss: 0.7683\n",
            "step 1487/5000, loss: 0.7267\n",
            "step 1488/5000, loss: 0.9892\n",
            "step 1489/5000, loss: 0.8724\n",
            "step 1490/5000, loss: 0.6819\n",
            "step 1491/5000, loss: 0.7671\n",
            "step 1492/5000, loss: 0.8490\n",
            "step 1493/5000, loss: 0.6890\n",
            "step 1494/5000, loss: 0.8718\n",
            "step 1495/5000, loss: 0.8578\n",
            "step 1496/5000, loss: 0.5994\n",
            "step 1497/5000, loss: 0.6608\n",
            "step 1498/5000, loss: 0.8059\n",
            "step 1499/5000, loss: 0.7054\n",
            "step 1500/5000, loss: 0.7131\n",
            "step 1501/5000, loss: 0.8819\n",
            "step 1502/5000, loss: 0.7158\n",
            "step 1503/5000, loss: 0.9534\n",
            "step 1504/5000, loss: 0.7988\n",
            "step 1505/5000, loss: 0.7422\n",
            "step 1506/5000, loss: 0.6376\n",
            "step 1507/5000, loss: 0.7532\n",
            "step 1508/5000, loss: 2.0079\n",
            "step 1509/5000, loss: 0.7881\n",
            "step 1510/5000, loss: 0.6943\n",
            "step 1511/5000, loss: 0.6859\n",
            "step 1512/5000, loss: 0.7758\n",
            "step 1513/5000, loss: 0.8269\n",
            "step 1514/5000, loss: 0.7774\n",
            "step 1515/5000, loss: 0.9317\n",
            "step 1516/5000, loss: 0.8495\n",
            "step 1517/5000, loss: 0.8810\n",
            "step 1518/5000, loss: 1.2290\n",
            "step 1519/5000, loss: 0.8170\n",
            "step 1520/5000, loss: 0.8316\n",
            "step 1521/5000, loss: 0.8278\n",
            "step 1522/5000, loss: 0.7005\n",
            "step 1523/5000, loss: 0.8026\n",
            "step 1524/5000, loss: 0.7433\n",
            "step 1525/5000, loss: 0.9395\n",
            "step 1526/5000, loss: 0.7436\n",
            "step 1527/5000, loss: 0.8385\n",
            "step 1528/5000, loss: 0.8834\n",
            "step 1529/5000, loss: 0.7708\n",
            "step 1530/5000, loss: 0.6996\n",
            "step 1531/5000, loss: 0.7385\n",
            "step 1532/5000, loss: 0.8854\n",
            "step 1533/5000, loss: 0.8465\n",
            "step 1534/5000, loss: 0.6793\n",
            "step 1535/5000, loss: 1.1262\n",
            "step 1536/5000, loss: 0.6891\n",
            "step 1537/5000, loss: 0.7580\n",
            "step 1538/5000, loss: 1.1042\n",
            "step 1539/5000, loss: 0.8641\n",
            "step 1540/5000, loss: 0.9180\n",
            "step 1541/5000, loss: 0.7036\n",
            "step 1542/5000, loss: 0.7511\n",
            "step 1543/5000, loss: 0.7267\n",
            "step 1544/5000, loss: 0.9018\n",
            "step 1545/5000, loss: 0.8029\n",
            "step 1546/5000, loss: 1.0253\n",
            "step 1547/5000, loss: 0.7039\n",
            "step 1548/5000, loss: 0.8414\n",
            "step 1549/5000, loss: 1.3323\n",
            "step 1550/5000, loss: 0.7530\n",
            "step 1551/5000, loss: 0.7279\n",
            "step 1552/5000, loss: 0.7109\n",
            "step 1553/5000, loss: 0.7284\n",
            "step 1554/5000, loss: 0.8477\n",
            "step 1555/5000, loss: 0.6722\n",
            "step 1556/5000, loss: 0.8036\n",
            "step 1557/5000, loss: 0.7858\n",
            "step 1558/5000, loss: 0.8391\n",
            "step 1559/5000, loss: 0.6273\n",
            "step 1560/5000, loss: 0.7709\n",
            "step 1561/5000, loss: 0.7142\n",
            "step 1562/5000, loss: 0.8071\n",
            "step 1563/5000, loss: 0.7290\n",
            "step 1564/5000, loss: 1.0021\n",
            "step 1565/5000, loss: 0.6702\n",
            "step 1566/5000, loss: 0.7309\n",
            "step 1567/5000, loss: 0.8176\n",
            "step 1568/5000, loss: 0.7216\n",
            "step 1569/5000, loss: 0.7230\n",
            "step 1570/5000, loss: 0.8597\n",
            "step 1571/5000, loss: 0.7367\n",
            "step 1572/5000, loss: 0.7863\n",
            "step 1573/5000, loss: 0.7916\n",
            "step 1574/5000, loss: 1.0144\n",
            "step 1575/5000, loss: 0.8932\n",
            "step 1576/5000, loss: 0.7236\n",
            "step 1577/5000, loss: 0.7839\n",
            "step 1578/5000, loss: 0.7946\n",
            "step 1579/5000, loss: 0.7988\n",
            "step 1580/5000, loss: 0.8346\n",
            "step 1581/5000, loss: 0.8700\n",
            "step 1582/5000, loss: 0.7142\n",
            "step 1583/5000, loss: 0.7329\n",
            "step 1584/5000, loss: 0.9350\n",
            "step 1585/5000, loss: 1.0212\n",
            "step 1586/5000, loss: 0.9861\n",
            "step 1587/5000, loss: 0.8474\n",
            "step 1588/5000, loss: 0.8907\n",
            "step 1589/5000, loss: 0.7902\n",
            "step 1590/5000, loss: 1.0383\n",
            "step 1591/5000, loss: 0.8311\n",
            "step 1592/5000, loss: 1.1123\n",
            "step 1593/5000, loss: 0.9961\n",
            "step 1594/5000, loss: 0.7224\n",
            "step 1595/5000, loss: 0.9564\n",
            "step 1596/5000, loss: 0.8775\n",
            "step 1597/5000, loss: 0.7752\n",
            "step 1598/5000, loss: 0.6677\n",
            "step 1599/5000, loss: 0.7956\n",
            "step 1600/5000, loss: 0.7212\n",
            "step 1601/5000, loss: 0.7943\n",
            "step 1602/5000, loss: 0.7470\n",
            "step 1603/5000, loss: 0.6754\n",
            "step 1604/5000, loss: 0.8866\n",
            "step 1605/5000, loss: 0.9066\n",
            "step 1606/5000, loss: 0.6115\n",
            "step 1607/5000, loss: 0.8869\n",
            "step 1608/5000, loss: 0.7918\n",
            "step 1609/5000, loss: 0.9381\n",
            "step 1610/5000, loss: 0.7528\n",
            "step 1611/5000, loss: 0.9004\n",
            "step 1612/5000, loss: 0.7808\n",
            "step 1613/5000, loss: 0.6937\n",
            "step 1614/5000, loss: 0.8991\n",
            "step 1615/5000, loss: 0.9220\n",
            "step 1616/5000, loss: 0.6055\n",
            "step 1617/5000, loss: 0.8765\n",
            "step 1618/5000, loss: 0.6424\n",
            "step 1619/5000, loss: 0.7707\n",
            "step 1620/5000, loss: 0.8258\n",
            "step 1621/5000, loss: 0.8870\n",
            "step 1622/5000, loss: 0.7544\n",
            "step 1623/5000, loss: 0.8106\n",
            "step 1624/5000, loss: 0.8577\n",
            "step 1625/5000, loss: 0.7691\n",
            "step 1626/5000, loss: 0.6508\n",
            "step 1627/5000, loss: 0.6898\n",
            "step 1628/5000, loss: 0.9130\n",
            "step 1629/5000, loss: 0.7856\n",
            "step 1630/5000, loss: 0.7082\n",
            "step 1631/5000, loss: 0.7809\n",
            "step 1632/5000, loss: 0.8062\n",
            "step 1633/5000, loss: 0.8986\n",
            "step 1634/5000, loss: 0.8177\n",
            "step 1635/5000, loss: 0.8412\n",
            "step 1636/5000, loss: 0.9105\n",
            "step 1637/5000, loss: 0.8047\n",
            "step 1638/5000, loss: 0.8398\n",
            "step 1639/5000, loss: 1.0252\n",
            "step 1640/5000, loss: 0.8307\n",
            "step 1641/5000, loss: 0.6629\n",
            "step 1642/5000, loss: 0.7270\n",
            "step 1643/5000, loss: 0.9517\n",
            "step 1644/5000, loss: 0.6955\n",
            "step 1645/5000, loss: 0.7367\n",
            "step 1646/5000, loss: 1.0333\n",
            "step 1647/5000, loss: 0.7060\n",
            "step 1648/5000, loss: 0.7023\n",
            "step 1649/5000, loss: 0.7733\n",
            "step 1650/5000, loss: 0.7549\n",
            "step 1651/5000, loss: 0.7646\n",
            "step 1652/5000, loss: 0.9749\n",
            "step 1653/5000, loss: 0.7610\n",
            "step 1654/5000, loss: 0.7777\n",
            "step 1655/5000, loss: 1.0330\n",
            "step 1656/5000, loss: 0.8457\n",
            "step 1657/5000, loss: 0.7295\n",
            "step 1658/5000, loss: 0.7543\n",
            "step 1659/5000, loss: 0.9313\n",
            "step 1660/5000, loss: 0.8003\n",
            "step 1661/5000, loss: 0.8077\n",
            "step 1662/5000, loss: 0.8512\n",
            "step 1663/5000, loss: 0.6589\n",
            "step 1664/5000, loss: 0.8422\n",
            "step 1665/5000, loss: 0.7416\n",
            "step 1666/5000, loss: 0.7311\n",
            "step 1667/5000, loss: 0.9511\n",
            "step 1668/5000, loss: 0.7917\n",
            "step 1669/5000, loss: 0.8207\n",
            "step 1670/5000, loss: 0.8123\n",
            "step 1671/5000, loss: 0.9640\n",
            "step 1672/5000, loss: 1.1429\n",
            "step 1673/5000, loss: 0.8032\n",
            "step 1674/5000, loss: 0.7886\n",
            "step 1675/5000, loss: 1.2146\n",
            "step 1676/5000, loss: 0.7088\n",
            "step 1677/5000, loss: 0.8924\n",
            "step 1678/5000, loss: 0.6775\n",
            "step 1679/5000, loss: 0.9215\n",
            "step 1680/5000, loss: 0.8147\n",
            "step 1681/5000, loss: 0.9924\n",
            "step 1682/5000, loss: 1.0156\n",
            "step 1683/5000, loss: 0.7636\n",
            "step 1684/5000, loss: 0.6370\n",
            "step 1685/5000, loss: 0.7892\n",
            "step 1686/5000, loss: 0.7206\n",
            "step 1687/5000, loss: 0.7330\n",
            "step 1688/5000, loss: 0.7620\n",
            "step 1689/5000, loss: 0.9024\n",
            "step 1690/5000, loss: 0.8087\n",
            "step 1691/5000, loss: 0.7928\n",
            "step 1692/5000, loss: 0.7203\n",
            "step 1693/5000, loss: 0.9096\n",
            "step 1694/5000, loss: 0.8806\n",
            "step 1695/5000, loss: 0.7536\n",
            "step 1696/5000, loss: 0.7825\n",
            "step 1697/5000, loss: 0.8677\n",
            "step 1698/5000, loss: 1.1506\n",
            "step 1699/5000, loss: 0.8576\n",
            "step 1700/5000, loss: 0.7357\n",
            "step 1701/5000, loss: 0.9947\n",
            "step 1702/5000, loss: 0.8844\n",
            "step 1703/5000, loss: 0.7724\n",
            "step 1704/5000, loss: 0.6910\n",
            "step 1705/5000, loss: 0.7366\n",
            "step 1706/5000, loss: 0.7296\n",
            "step 1707/5000, loss: 1.6381\n",
            "step 1708/5000, loss: 0.8870\n",
            "step 1709/5000, loss: 0.7386\n",
            "step 1710/5000, loss: 0.6692\n",
            "step 1711/5000, loss: 0.7151\n",
            "step 1712/5000, loss: 0.7440\n",
            "step 1713/5000, loss: 0.8178\n",
            "step 1714/5000, loss: 1.2275\n",
            "step 1715/5000, loss: 0.8171\n",
            "step 1716/5000, loss: 0.8154\n",
            "step 1717/5000, loss: 2.6047\n",
            "step 1718/5000, loss: 0.8460\n",
            "step 1719/5000, loss: 0.9249\n",
            "step 1720/5000, loss: 0.7143\n",
            "step 1721/5000, loss: 0.8409\n",
            "step 1722/5000, loss: 0.8979\n",
            "step 1723/5000, loss: 0.8177\n",
            "step 1724/5000, loss: 0.8704\n",
            "step 1725/5000, loss: 0.8076\n",
            "step 1726/5000, loss: 0.8655\n",
            "step 1727/5000, loss: 0.8003\n",
            "step 1728/5000, loss: 0.8832\n",
            "step 1729/5000, loss: 0.8163\n",
            "step 1730/5000, loss: 1.2947\n",
            "step 1731/5000, loss: 0.7132\n",
            "step 1732/5000, loss: 1.2642\n",
            "step 1733/5000, loss: 0.8215\n",
            "step 1734/5000, loss: 0.8447\n",
            "step 1735/5000, loss: 0.7982\n",
            "step 1736/5000, loss: 0.7908\n",
            "step 1737/5000, loss: 0.8854\n",
            "step 1738/5000, loss: 0.8033\n",
            "step 1739/5000, loss: 1.0895\n",
            "step 1740/5000, loss: 0.7820\n",
            "step 1741/5000, loss: 0.8405\n",
            "step 1742/5000, loss: 0.7806\n",
            "step 1743/5000, loss: 0.7556\n",
            "step 1744/5000, loss: 0.9111\n",
            "step 1745/5000, loss: 0.8483\n",
            "step 1746/5000, loss: 0.6721\n",
            "step 1747/5000, loss: 0.7955\n",
            "step 1748/5000, loss: 0.8612\n",
            "step 1749/5000, loss: 0.9659\n",
            "step 1750/5000, loss: 0.6891\n",
            "step 1751/5000, loss: 0.7495\n",
            "step 1752/5000, loss: 0.8438\n",
            "step 1753/5000, loss: 2.3721\n",
            "step 1754/5000, loss: 0.6517\n",
            "step 1755/5000, loss: 0.9869\n",
            "step 1756/5000, loss: 0.7330\n",
            "step 1757/5000, loss: 0.7475\n",
            "step 1758/5000, loss: 0.6689\n",
            "step 1759/5000, loss: 1.0077\n",
            "step 1760/5000, loss: 0.8206\n",
            "step 1761/5000, loss: 0.8525\n",
            "step 1762/5000, loss: 1.0035\n",
            "step 1763/5000, loss: 0.8958\n",
            "step 1764/5000, loss: 0.8673\n",
            "step 1765/5000, loss: 0.8038\n",
            "step 1766/5000, loss: 0.8430\n",
            "step 1767/5000, loss: 0.8129\n",
            "step 1768/5000, loss: 0.7403\n",
            "step 1769/5000, loss: 0.8828\n",
            "step 1770/5000, loss: 0.9943\n",
            "step 1771/5000, loss: 0.9766\n",
            "step 1772/5000, loss: 0.7775\n",
            "step 1773/5000, loss: 0.8789\n",
            "step 1774/5000, loss: 0.7159\n",
            "step 1775/5000, loss: 0.7758\n",
            "step 1776/5000, loss: 0.7544\n",
            "step 1777/5000, loss: 0.6316\n",
            "step 1778/5000, loss: 0.6809\n",
            "step 1779/5000, loss: 0.8077\n",
            "step 1780/5000, loss: 0.8692\n",
            "step 1781/5000, loss: 0.9263\n",
            "step 1782/5000, loss: 0.6492\n",
            "step 1783/5000, loss: 0.8107\n",
            "step 1784/5000, loss: 0.7151\n",
            "step 1785/5000, loss: 0.7241\n",
            "step 1786/5000, loss: 0.7699\n",
            "step 1787/5000, loss: 0.7484\n",
            "step 1788/5000, loss: 0.8723\n",
            "step 1789/5000, loss: 0.8852\n",
            "step 1790/5000, loss: 0.8220\n",
            "step 1791/5000, loss: 0.9043\n",
            "step 1792/5000, loss: 0.8723\n",
            "step 1793/5000, loss: 0.8796\n",
            "step 1794/5000, loss: 0.6812\n",
            "step 1795/5000, loss: 0.9895\n",
            "step 1796/5000, loss: 0.7491\n",
            "step 1797/5000, loss: 0.9357\n",
            "step 1798/5000, loss: 0.7028\n",
            "step 1799/5000, loss: 0.9090\n",
            "step 1800/5000, loss: 0.8049\n",
            "step 1801/5000, loss: 0.8048\n",
            "step 1802/5000, loss: 0.7899\n",
            "step 1803/5000, loss: 0.7706\n",
            "step 1804/5000, loss: 0.7932\n",
            "step 1805/5000, loss: 0.7751\n",
            "step 1806/5000, loss: 0.7159\n",
            "step 1807/5000, loss: 0.7412\n",
            "step 1808/5000, loss: 0.7799\n",
            "step 1809/5000, loss: 0.7571\n",
            "step 1810/5000, loss: 0.7684\n",
            "step 1811/5000, loss: 0.7175\n",
            "step 1812/5000, loss: 0.7260\n",
            "step 1813/5000, loss: 0.8302\n",
            "step 1814/5000, loss: 0.7231\n",
            "step 1815/5000, loss: 0.7952\n",
            "step 1816/5000, loss: 0.7726\n",
            "step 1817/5000, loss: 0.6779\n",
            "step 1818/5000, loss: 0.9073\n",
            "step 1819/5000, loss: 0.7978\n",
            "step 1820/5000, loss: 0.6389\n",
            "step 1821/5000, loss: 0.8523\n",
            "step 1822/5000, loss: 0.7316\n",
            "step 1823/5000, loss: 0.6705\n",
            "step 1824/5000, loss: 0.8128\n",
            "step 1825/5000, loss: 0.8420\n",
            "step 1826/5000, loss: 0.7197\n",
            "step 1827/5000, loss: 0.8943\n",
            "step 1828/5000, loss: 0.6232\n",
            "step 1829/5000, loss: 1.0887\n",
            "step 1830/5000, loss: 0.9601\n",
            "step 1831/5000, loss: 0.7758\n",
            "step 1832/5000, loss: 0.6534\n",
            "step 1833/5000, loss: 0.8234\n",
            "step 1834/5000, loss: 0.7378\n",
            "step 1835/5000, loss: 0.7137\n",
            "step 1836/5000, loss: 0.7800\n",
            "step 1837/5000, loss: 0.8713\n",
            "step 1838/5000, loss: 0.7189\n",
            "step 1839/5000, loss: 0.6753\n",
            "step 1840/5000, loss: 0.7065\n",
            "step 1841/5000, loss: 0.8995\n",
            "step 1842/5000, loss: 0.8224\n",
            "step 1843/5000, loss: 0.8031\n",
            "step 1844/5000, loss: 0.8522\n",
            "step 1845/5000, loss: 0.7224\n",
            "step 1846/5000, loss: 0.7830\n",
            "step 1847/5000, loss: 0.6976\n",
            "step 1848/5000, loss: 0.6623\n",
            "step 1849/5000, loss: 0.8548\n",
            "step 1850/5000, loss: 0.6662\n",
            "step 1851/5000, loss: 0.7290\n",
            "step 1852/5000, loss: 0.6701\n",
            "step 1853/5000, loss: 0.7222\n",
            "step 1854/5000, loss: 0.8441\n",
            "step 1855/5000, loss: 0.6343\n",
            "step 1856/5000, loss: 0.9004\n",
            "step 1857/5000, loss: 0.5856\n",
            "step 1858/5000, loss: 0.7014\n",
            "step 1859/5000, loss: 0.7236\n",
            "step 1860/5000, loss: 0.8416\n",
            "step 1861/5000, loss: 0.7518\n",
            "step 1862/5000, loss: 0.8207\n",
            "step 1863/5000, loss: 1.0096\n",
            "step 1864/5000, loss: 0.8487\n",
            "step 1865/5000, loss: 0.6545\n",
            "step 1866/5000, loss: 0.7526\n",
            "step 1867/5000, loss: 0.6805\n",
            "step 1868/5000, loss: 0.7628\n",
            "step 1869/5000, loss: 0.7796\n",
            "step 1870/5000, loss: 0.6660\n",
            "step 1871/5000, loss: 0.8979\n",
            "step 1872/5000, loss: 0.7625\n",
            "step 1873/5000, loss: 0.7610\n",
            "step 1874/5000, loss: 0.6818\n",
            "step 1875/5000, loss: 0.6793\n",
            "step 1876/5000, loss: 0.7356\n",
            "step 1877/5000, loss: 0.6721\n",
            "step 1878/5000, loss: 0.8462\n",
            "step 1879/5000, loss: 0.8188\n",
            "step 1880/5000, loss: 1.9370\n",
            "step 1881/5000, loss: 1.0176\n",
            "step 1882/5000, loss: 0.8313\n",
            "step 1883/5000, loss: 0.8253\n",
            "step 1884/5000, loss: 0.6578\n",
            "step 1885/5000, loss: 0.8068\n",
            "step 1886/5000, loss: 0.7884\n",
            "step 1887/5000, loss: 0.7527\n",
            "step 1888/5000, loss: 0.6724\n",
            "step 1889/5000, loss: 0.8141\n",
            "step 1890/5000, loss: 0.6013\n",
            "step 1891/5000, loss: 0.8942\n",
            "step 1892/5000, loss: 0.8084\n",
            "step 1893/5000, loss: 0.6419\n",
            "step 1894/5000, loss: 0.8943\n",
            "step 1895/5000, loss: 0.9641\n",
            "step 1896/5000, loss: 2.1738\n",
            "step 1897/5000, loss: 0.7761\n",
            "step 1898/5000, loss: 0.7934\n",
            "step 1899/5000, loss: 0.9184\n",
            "step 1900/5000, loss: 0.8188\n",
            "step 1901/5000, loss: 0.5995\n",
            "step 1902/5000, loss: 0.8029\n",
            "step 1903/5000, loss: 0.6508\n",
            "step 1904/5000, loss: 0.6500\n",
            "step 1905/5000, loss: 0.7682\n",
            "step 1906/5000, loss: 0.8194\n",
            "step 1907/5000, loss: 0.7749\n",
            "step 1908/5000, loss: 0.9008\n",
            "step 1909/5000, loss: 0.8695\n",
            "step 1910/5000, loss: 0.7298\n",
            "step 1911/5000, loss: 0.7562\n",
            "step 1912/5000, loss: 0.6912\n",
            "step 1913/5000, loss: 0.8466\n",
            "step 1914/5000, loss: 0.8429\n",
            "step 1915/5000, loss: 0.8627\n",
            "step 1916/5000, loss: 0.7282\n",
            "step 1917/5000, loss: 0.7386\n",
            "step 1918/5000, loss: 0.7274\n",
            "step 1919/5000, loss: 0.7678\n",
            "step 1920/5000, loss: 0.6852\n",
            "step 1921/5000, loss: 0.7115\n",
            "step 1922/5000, loss: 0.9202\n",
            "step 1923/5000, loss: 0.7696\n",
            "step 1924/5000, loss: 0.8503\n",
            "step 1925/5000, loss: 0.5786\n",
            "step 1926/5000, loss: 1.1595\n",
            "step 1927/5000, loss: 0.6398\n",
            "step 1928/5000, loss: 0.7840\n",
            "step 1929/5000, loss: 0.9660\n",
            "step 1930/5000, loss: 0.7057\n",
            "step 1931/5000, loss: 0.6950\n",
            "step 1932/5000, loss: 0.6135\n",
            "step 1933/5000, loss: 0.8708\n",
            "step 1934/5000, loss: 0.8094\n",
            "step 1935/5000, loss: 0.6699\n",
            "step 1936/5000, loss: 0.7791\n",
            "step 1937/5000, loss: 0.5711\n",
            "step 1938/5000, loss: 0.7116\n",
            "step 1939/5000, loss: 0.6388\n",
            "step 1940/5000, loss: 0.7649\n",
            "step 1941/5000, loss: 0.6755\n",
            "step 1942/5000, loss: 0.6129\n",
            "step 1943/5000, loss: 0.7485\n",
            "step 1944/5000, loss: 0.9829\n",
            "step 1945/5000, loss: 0.6589\n",
            "step 1946/5000, loss: 0.7553\n",
            "step 1947/5000, loss: 0.8285\n",
            "step 1948/5000, loss: 0.8177\n",
            "step 1949/5000, loss: 0.9644\n",
            "step 1950/5000, loss: 0.8464\n",
            "step 1951/5000, loss: 0.8924\n",
            "step 1952/5000, loss: 0.8440\n",
            "step 1953/5000, loss: 0.7383\n",
            "step 1954/5000, loss: 0.6865\n",
            "step 1955/5000, loss: 0.7398\n",
            "step 1956/5000, loss: 1.0459\n",
            "step 1957/5000, loss: 0.8108\n",
            "step 1958/5000, loss: 0.8236\n",
            "step 1959/5000, loss: 0.8685\n",
            "step 1960/5000, loss: 0.7883\n",
            "step 1961/5000, loss: 0.7365\n",
            "step 1962/5000, loss: 0.8848\n",
            "step 1963/5000, loss: 0.7288\n",
            "step 1964/5000, loss: 0.7390\n",
            "step 1965/5000, loss: 0.9405\n",
            "step 1966/5000, loss: 0.7649\n",
            "step 1967/5000, loss: 0.8069\n",
            "step 1968/5000, loss: 0.8890\n",
            "step 1969/5000, loss: 0.8225\n",
            "step 1970/5000, loss: 1.0815\n",
            "step 1971/5000, loss: 0.9130\n",
            "step 1972/5000, loss: 1.0591\n",
            "step 1973/5000, loss: 1.0427\n",
            "step 1974/5000, loss: 0.7525\n",
            "step 1975/5000, loss: 0.6857\n",
            "step 1976/5000, loss: 0.6963\n",
            "step 1977/5000, loss: 0.7748\n",
            "step 1978/5000, loss: 0.7214\n",
            "step 1979/5000, loss: 0.6877\n",
            "step 1980/5000, loss: 0.6722\n",
            "step 1981/5000, loss: 0.7807\n",
            "step 1982/5000, loss: 0.6490\n",
            "step 1983/5000, loss: 0.7452\n",
            "step 1984/5000, loss: 0.7418\n",
            "step 1985/5000, loss: 0.7979\n",
            "step 1986/5000, loss: 0.6950\n",
            "step 1987/5000, loss: 0.8079\n",
            "step 1988/5000, loss: 0.8061\n",
            "step 1989/5000, loss: 0.7146\n",
            "step 1990/5000, loss: 0.6512\n",
            "step 1991/5000, loss: 0.9153\n",
            "step 1992/5000, loss: 0.6172\n",
            "step 1993/5000, loss: 1.0090\n",
            "step 1994/5000, loss: 0.7495\n",
            "step 1995/5000, loss: 0.8938\n",
            "step 1996/5000, loss: 0.8910\n",
            "step 1997/5000, loss: 0.8617\n",
            "step 1998/5000, loss: 0.7836\n",
            "step 1999/5000, loss: 0.7524\n",
            "step 2000/5000, loss: 0.7332\n",
            "step 2001/5000, loss: 0.6479\n",
            "step 2002/5000, loss: 0.7999\n",
            "step 2003/5000, loss: 0.7254\n",
            "step 2004/5000, loss: 1.0183\n",
            "step 2005/5000, loss: 0.7369\n",
            "step 2006/5000, loss: 0.8068\n",
            "step 2007/5000, loss: 0.7333\n",
            "step 2008/5000, loss: 0.7978\n",
            "step 2009/5000, loss: 0.8863\n",
            "step 2010/5000, loss: 0.8232\n",
            "step 2011/5000, loss: 0.6337\n",
            "step 2012/5000, loss: 0.8289\n",
            "step 2013/5000, loss: 0.8113\n",
            "step 2014/5000, loss: 0.7512\n",
            "step 2015/5000, loss: 0.7366\n",
            "step 2016/5000, loss: 0.7518\n",
            "step 2017/5000, loss: 0.5971\n",
            "step 2018/5000, loss: 0.6927\n",
            "step 2019/5000, loss: 0.8182\n",
            "step 2020/5000, loss: 0.8107\n",
            "step 2021/5000, loss: 0.8827\n",
            "step 2022/5000, loss: 0.7114\n",
            "step 2023/5000, loss: 0.8558\n",
            "step 2024/5000, loss: 0.7745\n",
            "step 2025/5000, loss: 0.7233\n",
            "step 2026/5000, loss: 0.6778\n",
            "step 2027/5000, loss: 0.8711\n",
            "step 2028/5000, loss: 0.6816\n",
            "step 2029/5000, loss: 0.7837\n",
            "step 2030/5000, loss: 0.7172\n",
            "step 2031/5000, loss: 0.7882\n",
            "step 2032/5000, loss: 0.8231\n",
            "step 2033/5000, loss: 0.9059\n",
            "step 2034/5000, loss: 1.0044\n",
            "step 2035/5000, loss: 0.9991\n",
            "step 2036/5000, loss: 0.8911\n",
            "step 2037/5000, loss: 0.7788\n",
            "step 2038/5000, loss: 0.8458\n",
            "step 2039/5000, loss: 0.7034\n",
            "step 2040/5000, loss: 0.6260\n",
            "step 2041/5000, loss: 2.1844\n",
            "step 2042/5000, loss: 0.7610\n",
            "step 2043/5000, loss: 0.7329\n",
            "step 2044/5000, loss: 0.7547\n",
            "step 2045/5000, loss: 0.8275\n",
            "step 2046/5000, loss: 0.8518\n",
            "step 2047/5000, loss: 0.8267\n",
            "step 2048/5000, loss: 0.8573\n",
            "step 2049/5000, loss: 0.8089\n",
            "step 2050/5000, loss: 0.7975\n",
            "step 2051/5000, loss: 0.6632\n",
            "step 2052/5000, loss: 0.7388\n",
            "step 2053/5000, loss: 0.8583\n",
            "step 2054/5000, loss: 0.8254\n",
            "step 2055/5000, loss: 0.8165\n",
            "step 2056/5000, loss: 0.7385\n",
            "step 2057/5000, loss: 0.9501\n",
            "step 2058/5000, loss: 0.7623\n",
            "step 2059/5000, loss: 0.6439\n",
            "step 2060/5000, loss: 0.7029\n",
            "step 2061/5000, loss: 0.7683\n",
            "step 2062/5000, loss: 0.6545\n",
            "step 2063/5000, loss: 0.8716\n",
            "step 2064/5000, loss: 0.8015\n",
            "step 2065/5000, loss: 0.9713\n",
            "step 2066/5000, loss: 0.6777\n",
            "step 2067/5000, loss: 0.7006\n",
            "step 2068/5000, loss: 0.9904\n",
            "step 2069/5000, loss: 0.6376\n",
            "step 2070/5000, loss: 0.6829\n",
            "step 2071/5000, loss: 1.6474\n",
            "step 2072/5000, loss: 0.7561\n",
            "step 2073/5000, loss: 0.6970\n",
            "step 2074/5000, loss: 0.7186\n",
            "step 2075/5000, loss: 0.8108\n",
            "step 2076/5000, loss: 0.7101\n",
            "step 2077/5000, loss: 0.8270\n",
            "step 2078/5000, loss: 0.6671\n",
            "step 2079/5000, loss: 0.8180\n",
            "step 2080/5000, loss: 0.7338\n",
            "step 2081/5000, loss: 0.8280\n",
            "step 2082/5000, loss: 0.8994\n",
            "step 2083/5000, loss: 0.7333\n",
            "step 2084/5000, loss: 0.7858\n",
            "step 2085/5000, loss: 0.8125\n",
            "step 2086/5000, loss: 0.7830\n",
            "step 2087/5000, loss: 0.7737\n",
            "step 2088/5000, loss: 0.5912\n",
            "step 2089/5000, loss: 0.8028\n",
            "step 2090/5000, loss: 0.8695\n",
            "step 2091/5000, loss: 0.6616\n",
            "step 2092/5000, loss: 0.7417\n",
            "step 2093/5000, loss: 0.8602\n",
            "step 2094/5000, loss: 0.6642\n",
            "step 2095/5000, loss: 0.8084\n",
            "step 2096/5000, loss: 0.8191\n",
            "step 2097/5000, loss: 0.7417\n",
            "step 2098/5000, loss: 0.6824\n",
            "step 2099/5000, loss: 0.7442\n",
            "step 2100/5000, loss: 0.6085\n",
            "step 2101/5000, loss: 0.7287\n",
            "step 2102/5000, loss: 0.7129\n",
            "step 2103/5000, loss: 0.7548\n",
            "step 2104/5000, loss: 0.6288\n",
            "step 2105/5000, loss: 1.0488\n",
            "step 2106/5000, loss: 0.6816\n",
            "step 2107/5000, loss: 0.7684\n",
            "step 2108/5000, loss: 0.7012\n",
            "step 2109/5000, loss: 0.8989\n",
            "step 2110/5000, loss: 1.9993\n",
            "step 2111/5000, loss: 0.8439\n",
            "step 2112/5000, loss: 0.6524\n",
            "step 2113/5000, loss: 0.6836\n",
            "step 2114/5000, loss: 1.1930\n",
            "step 2115/5000, loss: 0.8160\n",
            "step 2116/5000, loss: 0.6864\n",
            "step 2117/5000, loss: 0.7080\n",
            "step 2118/5000, loss: 0.5949\n",
            "step 2119/5000, loss: 0.7315\n",
            "step 2120/5000, loss: 0.7795\n",
            "step 2121/5000, loss: 0.8014\n",
            "step 2122/5000, loss: 0.7049\n",
            "step 2123/5000, loss: 0.8718\n",
            "step 2124/5000, loss: 0.6747\n",
            "step 2125/5000, loss: 0.9756\n",
            "step 2126/5000, loss: 0.8293\n",
            "step 2127/5000, loss: 0.7264\n",
            "step 2128/5000, loss: 0.7566\n",
            "step 2129/5000, loss: 0.7169\n",
            "step 2130/5000, loss: 0.6940\n",
            "step 2131/5000, loss: 0.8312\n",
            "step 2132/5000, loss: 0.7745\n",
            "step 2133/5000, loss: 0.7787\n",
            "step 2134/5000, loss: 0.8981\n",
            "step 2135/5000, loss: 0.6356\n",
            "step 2136/5000, loss: 0.8122\n",
            "step 2137/5000, loss: 0.7441\n",
            "step 2138/5000, loss: 0.8152\n",
            "step 2139/5000, loss: 0.6814\n",
            "step 2140/5000, loss: 0.6447\n",
            "step 2141/5000, loss: 0.8307\n",
            "step 2142/5000, loss: 0.7263\n",
            "step 2143/5000, loss: 0.7312\n",
            "step 2144/5000, loss: 0.7325\n",
            "step 2145/5000, loss: 0.7732\n",
            "step 2146/5000, loss: 0.7342\n",
            "step 2147/5000, loss: 0.7669\n",
            "step 2148/5000, loss: 0.9706\n",
            "step 2149/5000, loss: 0.7035\n",
            "step 2150/5000, loss: 0.8433\n",
            "step 2151/5000, loss: 0.6881\n",
            "step 2152/5000, loss: 0.6787\n",
            "step 2153/5000, loss: 0.7522\n",
            "step 2154/5000, loss: 0.6335\n",
            "step 2155/5000, loss: 0.7739\n",
            "step 2156/5000, loss: 0.6603\n",
            "step 2157/5000, loss: 0.7818\n",
            "step 2158/5000, loss: 0.9807\n",
            "step 2159/5000, loss: 0.5780\n",
            "step 2160/5000, loss: 0.6352\n",
            "step 2161/5000, loss: 0.7836\n",
            "step 2162/5000, loss: 0.7566\n",
            "step 2163/5000, loss: 0.7329\n",
            "step 2164/5000, loss: 0.8124\n",
            "step 2165/5000, loss: 0.7413\n",
            "step 2166/5000, loss: 0.8589\n",
            "step 2167/5000, loss: 0.8121\n",
            "step 2168/5000, loss: 0.9430\n",
            "step 2169/5000, loss: 0.8375\n",
            "step 2170/5000, loss: 0.7557\n",
            "step 2171/5000, loss: 0.7668\n",
            "step 2172/5000, loss: 0.9395\n",
            "step 2173/5000, loss: 0.9105\n",
            "step 2174/5000, loss: 0.8084\n",
            "step 2175/5000, loss: 0.6926\n",
            "step 2176/5000, loss: 0.6545\n",
            "step 2177/5000, loss: 0.6900\n",
            "step 2178/5000, loss: 0.7736\n",
            "step 2179/5000, loss: 0.6674\n",
            "step 2180/5000, loss: 0.7491\n",
            "step 2181/5000, loss: 0.6331\n",
            "step 2182/5000, loss: 0.6649\n",
            "step 2183/5000, loss: 0.7198\n",
            "step 2184/5000, loss: 0.6698\n",
            "step 2185/5000, loss: 0.6707\n",
            "step 2186/5000, loss: 0.8645\n",
            "step 2187/5000, loss: 0.7422\n",
            "step 2188/5000, loss: 0.8008\n",
            "step 2189/5000, loss: 0.7948\n",
            "step 2190/5000, loss: 0.6862\n",
            "step 2191/5000, loss: 0.7493\n",
            "step 2192/5000, loss: 0.6968\n",
            "step 2193/5000, loss: 0.7576\n",
            "step 2194/5000, loss: 0.6809\n",
            "step 2195/5000, loss: 0.7448\n",
            "step 2196/5000, loss: 0.9646\n",
            "step 2197/5000, loss: 0.8638\n",
            "step 2198/5000, loss: 0.7090\n",
            "step 2199/5000, loss: 0.6897\n",
            "step 2200/5000, loss: 0.6599\n",
            "step 2201/5000, loss: 0.7350\n",
            "step 2202/5000, loss: 0.7114\n",
            "step 2203/5000, loss: 0.8123\n",
            "step 2204/5000, loss: 0.7601\n",
            "step 2205/5000, loss: 0.8609\n",
            "step 2206/5000, loss: 0.7746\n",
            "step 2207/5000, loss: 0.7890\n",
            "step 2208/5000, loss: 0.7308\n",
            "step 2209/5000, loss: 0.6798\n",
            "step 2210/5000, loss: 0.7676\n",
            "step 2211/5000, loss: 0.9073\n",
            "step 2212/5000, loss: 0.7512\n",
            "step 2213/5000, loss: 0.8821\n",
            "step 2214/5000, loss: 0.7285\n",
            "step 2215/5000, loss: 0.9945\n",
            "step 2216/5000, loss: 0.6564\n",
            "step 2217/5000, loss: 0.6991\n",
            "step 2218/5000, loss: 0.6881\n",
            "step 2219/5000, loss: 0.8831\n",
            "step 2220/5000, loss: 0.7041\n",
            "step 2221/5000, loss: 0.9274\n",
            "step 2222/5000, loss: 0.6684\n",
            "step 2223/5000, loss: 0.8882\n",
            "step 2224/5000, loss: 0.8119\n",
            "step 2225/5000, loss: 0.7946\n",
            "step 2226/5000, loss: 0.7468\n",
            "step 2227/5000, loss: 0.7852\n",
            "step 2228/5000, loss: 0.7167\n",
            "step 2229/5000, loss: 0.7708\n",
            "step 2230/5000, loss: 0.8203\n",
            "step 2231/5000, loss: 0.6842\n",
            "step 2232/5000, loss: 0.8789\n",
            "step 2233/5000, loss: 0.8179\n",
            "step 2234/5000, loss: 0.6818\n",
            "step 2235/5000, loss: 0.7669\n",
            "step 2236/5000, loss: 0.7630\n",
            "step 2237/5000, loss: 0.8484\n",
            "step 2238/5000, loss: 0.7266\n",
            "step 2239/5000, loss: 0.8192\n",
            "step 2240/5000, loss: 0.7681\n",
            "step 2241/5000, loss: 0.8664\n",
            "step 2242/5000, loss: 0.7771\n",
            "step 2243/5000, loss: 0.8312\n",
            "step 2244/5000, loss: 0.8257\n",
            "step 2245/5000, loss: 0.7069\n",
            "step 2246/5000, loss: 1.0323\n",
            "step 2247/5000, loss: 0.7661\n",
            "step 2248/5000, loss: 0.7787\n",
            "step 2249/5000, loss: 0.8762\n",
            "step 2250/5000, loss: 0.7062\n",
            "step 2251/5000, loss: 0.7695\n",
            "step 2252/5000, loss: 0.8167\n",
            "step 2253/5000, loss: 0.6180\n",
            "step 2254/5000, loss: 0.6203\n",
            "step 2255/5000, loss: 0.5770\n",
            "step 2256/5000, loss: 0.7552\n",
            "step 2257/5000, loss: 0.6690\n",
            "step 2258/5000, loss: 0.6555\n",
            "step 2259/5000, loss: 0.6329\n",
            "step 2260/5000, loss: 0.6701\n",
            "step 2261/5000, loss: 0.7415\n",
            "step 2262/5000, loss: 0.7569\n",
            "step 2263/5000, loss: 0.7070\n",
            "step 2264/5000, loss: 0.7611\n",
            "step 2265/5000, loss: 0.6487\n",
            "step 2266/5000, loss: 1.4981\n",
            "step 2267/5000, loss: 0.7350\n",
            "step 2268/5000, loss: 0.7409\n",
            "step 2269/5000, loss: 0.9460\n",
            "step 2270/5000, loss: 0.7593\n",
            "step 2271/5000, loss: 0.7178\n",
            "step 2272/5000, loss: 0.7061\n",
            "step 2273/5000, loss: 0.6855\n",
            "step 2274/5000, loss: 0.7235\n",
            "step 2275/5000, loss: 0.7041\n",
            "step 2276/5000, loss: 0.6660\n",
            "step 2277/5000, loss: 0.7041\n",
            "step 2278/5000, loss: 0.6687\n",
            "step 2279/5000, loss: 0.6973\n",
            "step 2280/5000, loss: 0.7078\n",
            "step 2281/5000, loss: 0.8751\n",
            "step 2282/5000, loss: 0.7899\n",
            "step 2283/5000, loss: 0.7020\n",
            "step 2284/5000, loss: 0.6622\n",
            "step 2285/5000, loss: 0.7905\n",
            "step 2286/5000, loss: 0.7490\n",
            "step 2287/5000, loss: 0.7102\n",
            "step 2288/5000, loss: 0.9407\n",
            "step 2289/5000, loss: 0.7377\n",
            "step 2290/5000, loss: 0.8496\n",
            "step 2291/5000, loss: 0.8388\n",
            "step 2292/5000, loss: 0.8060\n",
            "step 2293/5000, loss: 0.6483\n",
            "step 2294/5000, loss: 0.7277\n",
            "step 2295/5000, loss: 0.6753\n",
            "step 2296/5000, loss: 0.8607\n",
            "step 2297/5000, loss: 0.8675\n",
            "step 2298/5000, loss: 0.6289\n",
            "step 2299/5000, loss: 0.7367\n",
            "step 2300/5000, loss: 0.8403\n",
            "step 2301/5000, loss: 0.6995\n",
            "step 2302/5000, loss: 0.6820\n",
            "step 2303/5000, loss: 0.6209\n",
            "step 2304/5000, loss: 0.7408\n",
            "step 2305/5000, loss: 0.6128\n",
            "step 2306/5000, loss: 0.7529\n",
            "step 2307/5000, loss: 0.7915\n",
            "step 2308/5000, loss: 0.7898\n",
            "step 2309/5000, loss: 0.6602\n",
            "step 2310/5000, loss: 0.8039\n",
            "step 2311/5000, loss: 0.6851\n",
            "step 2312/5000, loss: 0.7616\n",
            "step 2313/5000, loss: 0.7910\n",
            "step 2314/5000, loss: 0.6903\n",
            "step 2315/5000, loss: 0.7952\n",
            "step 2316/5000, loss: 0.6324\n",
            "step 2317/5000, loss: 0.6813\n",
            "step 2318/5000, loss: 0.6690\n",
            "step 2319/5000, loss: 0.7461\n",
            "step 2320/5000, loss: 1.0265\n",
            "step 2321/5000, loss: 0.6622\n",
            "step 2322/5000, loss: 0.7412\n",
            "step 2323/5000, loss: 0.8646\n",
            "step 2324/5000, loss: 0.9003\n",
            "step 2325/5000, loss: 0.7151\n",
            "step 2326/5000, loss: 0.6246\n",
            "step 2327/5000, loss: 0.7788\n",
            "step 2328/5000, loss: 0.7095\n",
            "step 2329/5000, loss: 1.8888\n",
            "step 2330/5000, loss: 0.8198\n",
            "step 2331/5000, loss: 0.6667\n",
            "step 2332/5000, loss: 0.6728\n",
            "step 2333/5000, loss: 0.6660\n",
            "step 2334/5000, loss: 0.6262\n",
            "step 2335/5000, loss: 0.7269\n",
            "step 2336/5000, loss: 0.6812\n",
            "step 2337/5000, loss: 0.6940\n",
            "step 2338/5000, loss: 0.7016\n",
            "step 2339/5000, loss: 0.6981\n",
            "step 2340/5000, loss: 0.8421\n",
            "step 2341/5000, loss: 0.7401\n",
            "step 2342/5000, loss: 0.7558\n",
            "step 2343/5000, loss: 0.8063\n",
            "step 2344/5000, loss: 0.8136\n",
            "step 2345/5000, loss: 0.7545\n",
            "step 2346/5000, loss: 0.8062\n",
            "step 2347/5000, loss: 0.8068\n",
            "step 2348/5000, loss: 0.7568\n",
            "step 2349/5000, loss: 0.8252\n",
            "step 2350/5000, loss: 0.9331\n",
            "step 2351/5000, loss: 0.7466\n",
            "step 2352/5000, loss: 0.6239\n",
            "step 2353/5000, loss: 0.7325\n",
            "step 2354/5000, loss: 0.7810\n",
            "step 2355/5000, loss: 0.7306\n",
            "step 2356/5000, loss: 0.8282\n",
            "step 2357/5000, loss: 0.7417\n",
            "step 2358/5000, loss: 0.8752\n",
            "step 2359/5000, loss: 1.0291\n",
            "step 2360/5000, loss: 0.6679\n",
            "step 2361/5000, loss: 0.7280\n",
            "step 2362/5000, loss: 0.7407\n",
            "step 2363/5000, loss: 0.7334\n",
            "step 2364/5000, loss: 0.7652\n",
            "step 2365/5000, loss: 0.9170\n",
            "step 2366/5000, loss: 0.8318\n",
            "step 2367/5000, loss: 0.7806\n",
            "step 2368/5000, loss: 1.0010\n",
            "step 2369/5000, loss: 0.8044\n",
            "step 2370/5000, loss: 0.8730\n",
            "step 2371/5000, loss: 0.6828\n",
            "step 2372/5000, loss: 0.6332\n",
            "step 2373/5000, loss: 0.7138\n",
            "step 2374/5000, loss: 0.6325\n",
            "step 2375/5000, loss: 0.7112\n",
            "step 2376/5000, loss: 0.7837\n",
            "step 2377/5000, loss: 0.7587\n",
            "step 2378/5000, loss: 0.8310\n",
            "step 2379/5000, loss: 0.8537\n",
            "step 2380/5000, loss: 0.7314\n",
            "step 2381/5000, loss: 0.8440\n",
            "step 2382/5000, loss: 0.6546\n",
            "step 2383/5000, loss: 0.8553\n",
            "step 2384/5000, loss: 0.7865\n",
            "step 2385/5000, loss: 0.6739\n",
            "step 2386/5000, loss: 0.6673\n",
            "step 2387/5000, loss: 0.8026\n",
            "step 2388/5000, loss: 0.7895\n",
            "step 2389/5000, loss: 0.6640\n",
            "step 2390/5000, loss: 0.7751\n",
            "step 2391/5000, loss: 0.7867\n",
            "step 2392/5000, loss: 0.7632\n",
            "step 2393/5000, loss: 0.7080\n",
            "step 2394/5000, loss: 0.7216\n",
            "step 2395/5000, loss: 0.6942\n",
            "step 2396/5000, loss: 0.9654\n",
            "step 2397/5000, loss: 1.3621\n",
            "step 2398/5000, loss: 0.8360\n",
            "step 2399/5000, loss: 1.0968\n",
            "step 2400/5000, loss: 0.6351\n",
            "step 2401/5000, loss: 0.6760\n",
            "step 2402/5000, loss: 0.6449\n",
            "step 2403/5000, loss: 0.8918\n",
            "step 2404/5000, loss: 0.8197\n",
            "step 2405/5000, loss: 0.6206\n",
            "step 2406/5000, loss: 0.7697\n",
            "step 2407/5000, loss: 0.6611\n",
            "step 2408/5000, loss: 0.7541\n",
            "step 2409/5000, loss: 0.8662\n",
            "step 2410/5000, loss: 0.7240\n",
            "step 2411/5000, loss: 0.6841\n",
            "step 2412/5000, loss: 0.7362\n",
            "step 2413/5000, loss: 0.5760\n",
            "step 2414/5000, loss: 0.7195\n",
            "step 2415/5000, loss: 0.8664\n",
            "step 2416/5000, loss: 0.7566\n",
            "step 2417/5000, loss: 0.6650\n",
            "step 2418/5000, loss: 1.0139\n",
            "step 2419/5000, loss: 0.7234\n",
            "step 2420/5000, loss: 0.8276\n",
            "step 2421/5000, loss: 0.7470\n",
            "step 2422/5000, loss: 0.8565\n",
            "step 2423/5000, loss: 0.6514\n",
            "step 2424/5000, loss: 0.6397\n",
            "step 2425/5000, loss: 0.8485\n",
            "step 2426/5000, loss: 0.9348\n",
            "step 2427/5000, loss: 0.8001\n",
            "step 2428/5000, loss: 0.8737\n",
            "step 2429/5000, loss: 0.6483\n",
            "step 2430/5000, loss: 0.8080\n",
            "step 2431/5000, loss: 0.7279\n",
            "step 2432/5000, loss: 0.6917\n",
            "step 2433/5000, loss: 0.6301\n",
            "step 2434/5000, loss: 1.2375\n",
            "step 2435/5000, loss: 0.8189\n",
            "step 2436/5000, loss: 0.6890\n",
            "step 2437/5000, loss: 0.8189\n",
            "step 2438/5000, loss: 0.7681\n",
            "step 2439/5000, loss: 0.7264\n",
            "step 2440/5000, loss: 0.7665\n",
            "step 2441/5000, loss: 0.6621\n",
            "step 2442/5000, loss: 0.6431\n",
            "step 2443/5000, loss: 0.6968\n",
            "step 2444/5000, loss: 0.6877\n",
            "step 2445/5000, loss: 0.6899\n",
            "step 2446/5000, loss: 0.9322\n",
            "step 2447/5000, loss: 0.7042\n",
            "step 2448/5000, loss: 1.7471\n",
            "step 2449/5000, loss: 0.6706\n",
            "step 2450/5000, loss: 0.7841\n",
            "step 2451/5000, loss: 0.8272\n",
            "step 2452/5000, loss: 0.6215\n",
            "step 2453/5000, loss: 0.6680\n",
            "step 2454/5000, loss: 0.6701\n",
            "step 2455/5000, loss: 0.7958\n",
            "step 2456/5000, loss: 0.9042\n",
            "step 2457/5000, loss: 0.7346\n",
            "step 2458/5000, loss: 0.7896\n",
            "step 2459/5000, loss: 0.7573\n",
            "step 2460/5000, loss: 0.7770\n",
            "step 2461/5000, loss: 0.6533\n",
            "step 2462/5000, loss: 0.8398\n",
            "step 2463/5000, loss: 0.6952\n",
            "step 2464/5000, loss: 0.7599\n",
            "step 2465/5000, loss: 0.7002\n",
            "step 2466/5000, loss: 0.8928\n",
            "step 2467/5000, loss: 0.6526\n",
            "step 2468/5000, loss: 0.8798\n",
            "step 2469/5000, loss: 0.8442\n",
            "step 2470/5000, loss: 0.9343\n",
            "step 2471/5000, loss: 0.7582\n",
            "step 2472/5000, loss: 0.7640\n",
            "step 2473/5000, loss: 0.8091\n",
            "step 2474/5000, loss: 0.7682\n",
            "step 2475/5000, loss: 0.7153\n",
            "step 2476/5000, loss: 0.6407\n",
            "step 2477/5000, loss: 0.8418\n",
            "step 2478/5000, loss: 0.7984\n",
            "step 2479/5000, loss: 0.6046\n",
            "step 2480/5000, loss: 0.7520\n",
            "step 2481/5000, loss: 0.7663\n",
            "step 2482/5000, loss: 0.8169\n",
            "step 2483/5000, loss: 0.6482\n",
            "step 2484/5000, loss: 0.7485\n",
            "step 2485/5000, loss: 0.5621\n",
            "step 2486/5000, loss: 0.9054\n",
            "step 2487/5000, loss: 0.6901\n",
            "step 2488/5000, loss: 0.8413\n",
            "step 2489/5000, loss: 0.6265\n",
            "step 2490/5000, loss: 0.9097\n",
            "step 2491/5000, loss: 0.6759\n",
            "step 2492/5000, loss: 0.8020\n",
            "step 2493/5000, loss: 0.7215\n",
            "step 2494/5000, loss: 0.6801\n",
            "step 2495/5000, loss: 0.6025\n",
            "step 2496/5000, loss: 0.7705\n",
            "step 2497/5000, loss: 0.6920\n",
            "step 2498/5000, loss: 0.7384\n",
            "step 2499/5000, loss: 0.6295\n",
            "step 2500/5000, loss: 0.6383\n",
            "step 2501/5000, loss: 0.7988\n",
            "step 2502/5000, loss: 0.8446\n",
            "step 2503/5000, loss: 0.6160\n",
            "step 2504/5000, loss: 0.6649\n",
            "step 2505/5000, loss: 0.8493\n",
            "step 2506/5000, loss: 0.8099\n",
            "step 2507/5000, loss: 0.7833\n",
            "step 2508/5000, loss: 0.9506\n",
            "step 2509/5000, loss: 0.7268\n",
            "step 2510/5000, loss: 0.7041\n",
            "step 2511/5000, loss: 0.7809\n",
            "step 2512/5000, loss: 0.7089\n",
            "step 2513/5000, loss: 0.7600\n",
            "step 2514/5000, loss: 0.8046\n",
            "step 2515/5000, loss: 0.6425\n",
            "step 2516/5000, loss: 0.6234\n",
            "step 2517/5000, loss: 0.7285\n",
            "step 2518/5000, loss: 0.7687\n",
            "step 2519/5000, loss: 0.6662\n",
            "step 2520/5000, loss: 0.6196\n",
            "step 2521/5000, loss: 0.7987\n",
            "step 2522/5000, loss: 0.7696\n",
            "step 2523/5000, loss: 0.5893\n",
            "step 2524/5000, loss: 0.6949\n",
            "step 2525/5000, loss: 0.6145\n",
            "step 2526/5000, loss: 0.7662\n",
            "step 2527/5000, loss: 0.6300\n",
            "step 2528/5000, loss: 0.8266\n",
            "step 2529/5000, loss: 0.7463\n",
            "step 2530/5000, loss: 0.6916\n",
            "step 2531/5000, loss: 0.7670\n",
            "step 2532/5000, loss: 0.7644\n",
            "step 2533/5000, loss: 1.0226\n",
            "step 2534/5000, loss: 0.7007\n",
            "step 2535/5000, loss: 0.5513\n",
            "step 2536/5000, loss: 0.6178\n",
            "step 2537/5000, loss: 2.1350\n",
            "step 2538/5000, loss: 0.7019\n",
            "step 2539/5000, loss: 0.7168\n",
            "step 2540/5000, loss: 0.6430\n",
            "step 2541/5000, loss: 0.7315\n",
            "step 2542/5000, loss: 0.9547\n",
            "step 2543/5000, loss: 0.5889\n",
            "step 2544/5000, loss: 0.7375\n",
            "step 2545/5000, loss: 0.8784\n",
            "step 2546/5000, loss: 0.7326\n",
            "step 2547/5000, loss: 0.7146\n",
            "step 2548/5000, loss: 0.8340\n",
            "step 2549/5000, loss: 0.7812\n",
            "step 2550/5000, loss: 0.7371\n",
            "step 2551/5000, loss: 0.5617\n",
            "step 2552/5000, loss: 0.8672\n",
            "step 2553/5000, loss: 0.6740\n",
            "step 2554/5000, loss: 0.6656\n",
            "step 2555/5000, loss: 0.7406\n",
            "step 2556/5000, loss: 0.6866\n",
            "step 2557/5000, loss: 0.7079\n",
            "step 2558/5000, loss: 0.9005\n",
            "step 2559/5000, loss: 0.7175\n",
            "step 2560/5000, loss: 0.7509\n",
            "step 2561/5000, loss: 0.6807\n",
            "step 2562/5000, loss: 0.7259\n",
            "step 2563/5000, loss: 0.7530\n",
            "step 2564/5000, loss: 0.7530\n",
            "step 2565/5000, loss: 0.7514\n",
            "step 2566/5000, loss: 0.6217\n",
            "step 2567/5000, loss: 0.7646\n",
            "step 2568/5000, loss: 0.7074\n",
            "step 2569/5000, loss: 0.7792\n",
            "step 2570/5000, loss: 0.8098\n",
            "step 2571/5000, loss: 0.6901\n",
            "step 2572/5000, loss: 0.7103\n",
            "step 2573/5000, loss: 0.6770\n",
            "step 2574/5000, loss: 0.6654\n",
            "step 2575/5000, loss: 0.7838\n",
            "step 2576/5000, loss: 0.6874\n",
            "step 2577/5000, loss: 0.9218\n",
            "step 2578/5000, loss: 0.7548\n",
            "step 2579/5000, loss: 0.6302\n",
            "step 2580/5000, loss: 0.7907\n",
            "step 2581/5000, loss: 0.7738\n",
            "step 2582/5000, loss: 0.7818\n",
            "step 2583/5000, loss: 0.6712\n",
            "step 2584/5000, loss: 0.6142\n",
            "step 2585/5000, loss: 0.7838\n",
            "step 2586/5000, loss: 0.8451\n",
            "step 2587/5000, loss: 0.8493\n",
            "step 2588/5000, loss: 0.8101\n",
            "step 2589/5000, loss: 0.7106\n",
            "step 2590/5000, loss: 0.5910\n",
            "step 2591/5000, loss: 0.9291\n",
            "step 2592/5000, loss: 0.9130\n",
            "step 2593/5000, loss: 0.8069\n",
            "step 2594/5000, loss: 0.7412\n",
            "step 2595/5000, loss: 0.6587\n",
            "step 2596/5000, loss: 0.7122\n",
            "step 2597/5000, loss: 0.5972\n",
            "step 2598/5000, loss: 0.7402\n",
            "step 2599/5000, loss: 0.7055\n",
            "step 2600/5000, loss: 0.7222\n",
            "step 2601/5000, loss: 0.7448\n",
            "step 2602/5000, loss: 0.7258\n",
            "step 2603/5000, loss: 0.7250\n",
            "step 2604/5000, loss: 0.7139\n",
            "step 2605/5000, loss: 0.6467\n",
            "step 2606/5000, loss: 0.7380\n",
            "step 2607/5000, loss: 0.8531\n",
            "step 2608/5000, loss: 0.6926\n",
            "step 2609/5000, loss: 0.7112\n",
            "step 2610/5000, loss: 0.7423\n",
            "step 2611/5000, loss: 0.7820\n",
            "step 2612/5000, loss: 0.8228\n",
            "step 2613/5000, loss: 1.8349\n",
            "step 2614/5000, loss: 0.7937\n",
            "step 2615/5000, loss: 0.7855\n",
            "step 2616/5000, loss: 0.6878\n",
            "step 2617/5000, loss: 0.6908\n",
            "step 2618/5000, loss: 0.6941\n",
            "step 2619/5000, loss: 0.8218\n",
            "step 2620/5000, loss: 0.7675\n",
            "step 2621/5000, loss: 0.7264\n",
            "step 2622/5000, loss: 0.8896\n",
            "step 2623/5000, loss: 0.7983\n",
            "step 2624/5000, loss: 0.7265\n",
            "step 2625/5000, loss: 0.6777\n",
            "step 2626/5000, loss: 0.6999\n",
            "step 2627/5000, loss: 0.7267\n",
            "step 2628/5000, loss: 0.7454\n",
            "step 2629/5000, loss: 0.6998\n",
            "step 2630/5000, loss: 0.7426\n",
            "step 2631/5000, loss: 0.8090\n",
            "step 2632/5000, loss: 0.8395\n",
            "step 2633/5000, loss: 0.8836\n",
            "step 2634/5000, loss: 0.7437\n",
            "step 2635/5000, loss: 0.6067\n",
            "step 2636/5000, loss: 0.6869\n",
            "step 2637/5000, loss: 0.6182\n",
            "step 2638/5000, loss: 0.7517\n",
            "step 2639/5000, loss: 0.9078\n",
            "step 2640/5000, loss: 0.7184\n",
            "step 2641/5000, loss: 0.6536\n",
            "step 2642/5000, loss: 0.8282\n",
            "step 2643/5000, loss: 0.6393\n",
            "step 2644/5000, loss: 0.7844\n",
            "step 2645/5000, loss: 0.7415\n",
            "step 2646/5000, loss: 0.7579\n",
            "step 2647/5000, loss: 0.7313\n",
            "step 2648/5000, loss: 0.7698\n",
            "step 2649/5000, loss: 0.6753\n",
            "step 2650/5000, loss: 0.6824\n",
            "step 2651/5000, loss: 0.9982\n",
            "step 2652/5000, loss: 0.6829\n",
            "step 2653/5000, loss: 0.6646\n",
            "step 2654/5000, loss: 0.8771\n",
            "step 2655/5000, loss: 0.8540\n",
            "step 2656/5000, loss: 0.8220\n",
            "step 2657/5000, loss: 0.8361\n",
            "step 2658/5000, loss: 0.7283\n",
            "step 2659/5000, loss: 0.6260\n",
            "step 2660/5000, loss: 1.0906\n",
            "step 2661/5000, loss: 0.7037\n",
            "step 2662/5000, loss: 0.7519\n",
            "step 2663/5000, loss: 0.8102\n",
            "step 2664/5000, loss: 0.9418\n",
            "step 2665/5000, loss: 0.8326\n",
            "step 2666/5000, loss: 0.6462\n",
            "step 2667/5000, loss: 0.7334\n",
            "step 2668/5000, loss: 0.7849\n",
            "step 2669/5000, loss: 0.8619\n",
            "step 2670/5000, loss: 0.6815\n",
            "step 2671/5000, loss: 0.8331\n",
            "step 2672/5000, loss: 0.8359\n",
            "step 2673/5000, loss: 0.7366\n",
            "step 2674/5000, loss: 0.7139\n",
            "step 2675/5000, loss: 0.7204\n",
            "step 2676/5000, loss: 0.7200\n",
            "step 2677/5000, loss: 0.6373\n",
            "step 2678/5000, loss: 0.7360\n",
            "step 2679/5000, loss: 0.6620\n",
            "step 2680/5000, loss: 0.7278\n",
            "step 2681/5000, loss: 0.6821\n",
            "step 2682/5000, loss: 0.7996\n",
            "step 2683/5000, loss: 0.7161\n",
            "step 2684/5000, loss: 0.7857\n",
            "step 2685/5000, loss: 0.7142\n",
            "step 2686/5000, loss: 0.6767\n",
            "step 2687/5000, loss: 0.7702\n",
            "step 2688/5000, loss: 0.7276\n",
            "step 2689/5000, loss: 0.7500\n",
            "step 2690/5000, loss: 0.6709\n",
            "step 2691/5000, loss: 0.7776\n",
            "step 2692/5000, loss: 1.1027\n",
            "step 2693/5000, loss: 0.7004\n",
            "step 2694/5000, loss: 0.8646\n",
            "step 2695/5000, loss: 0.6568\n",
            "step 2696/5000, loss: 0.7333\n",
            "step 2697/5000, loss: 0.7333\n",
            "step 2698/5000, loss: 0.6823\n",
            "step 2699/5000, loss: 0.6199\n",
            "step 2700/5000, loss: 0.7698\n",
            "step 2701/5000, loss: 0.5765\n",
            "step 2702/5000, loss: 0.6648\n",
            "step 2703/5000, loss: 0.7318\n",
            "step 2704/5000, loss: 0.7768\n",
            "step 2705/5000, loss: 0.6872\n",
            "step 2706/5000, loss: 0.6046\n",
            "step 2707/5000, loss: 0.8150\n",
            "step 2708/5000, loss: 0.6877\n",
            "step 2709/5000, loss: 0.7272\n",
            "step 2710/5000, loss: 0.8962\n",
            "step 2711/5000, loss: 0.8315\n",
            "step 2712/5000, loss: 0.7148\n",
            "step 2713/5000, loss: 0.7638\n",
            "step 2714/5000, loss: 0.8641\n",
            "step 2715/5000, loss: 0.6992\n",
            "step 2716/5000, loss: 0.7433\n",
            "step 2717/5000, loss: 0.8517\n",
            "step 2718/5000, loss: 0.7402\n",
            "step 2719/5000, loss: 0.6317\n",
            "step 2720/5000, loss: 0.6117\n",
            "step 2721/5000, loss: 0.8649\n",
            "step 2722/5000, loss: 0.6877\n",
            "step 2723/5000, loss: 0.7160\n",
            "step 2724/5000, loss: 0.9918\n",
            "step 2725/5000, loss: 0.6860\n",
            "step 2726/5000, loss: 0.7497\n",
            "step 2727/5000, loss: 0.6916\n",
            "step 2728/5000, loss: 0.6489\n",
            "step 2729/5000, loss: 0.9037\n",
            "step 2730/5000, loss: 0.7699\n",
            "step 2731/5000, loss: 0.8138\n",
            "step 2732/5000, loss: 0.6629\n",
            "step 2733/5000, loss: 0.6357\n",
            "step 2734/5000, loss: 0.5939\n",
            "step 2735/5000, loss: 0.6308\n",
            "step 2736/5000, loss: 0.6354\n",
            "step 2737/5000, loss: 0.6770\n",
            "step 2738/5000, loss: 0.7431\n",
            "step 2739/5000, loss: 0.7172\n",
            "step 2740/5000, loss: 0.7230\n",
            "step 2741/5000, loss: 0.5961\n",
            "step 2742/5000, loss: 0.6712\n",
            "step 2743/5000, loss: 1.3696\n",
            "step 2744/5000, loss: 0.6614\n",
            "step 2745/5000, loss: 0.7446\n",
            "step 2746/5000, loss: 0.6781\n",
            "step 2747/5000, loss: 0.6708\n",
            "step 2748/5000, loss: 0.9442\n",
            "step 2749/5000, loss: 0.8239\n",
            "step 2750/5000, loss: 0.6583\n",
            "step 2751/5000, loss: 0.7578\n",
            "step 2752/5000, loss: 0.6855\n",
            "step 2753/5000, loss: 0.6151\n",
            "step 2754/5000, loss: 0.6890\n",
            "step 2755/5000, loss: 0.7284\n",
            "step 2756/5000, loss: 0.8035\n",
            "step 2757/5000, loss: 0.7916\n",
            "step 2758/5000, loss: 0.6189\n",
            "step 2759/5000, loss: 0.5687\n",
            "step 2760/5000, loss: 0.7835\n",
            "step 2761/5000, loss: 0.6511\n",
            "step 2762/5000, loss: 0.8005\n",
            "step 2763/5000, loss: 0.6781\n",
            "step 2764/5000, loss: 0.7055\n",
            "step 2765/5000, loss: 0.5799\n",
            "step 2766/5000, loss: 0.6536\n",
            "step 2767/5000, loss: 0.7350\n",
            "step 2768/5000, loss: 0.6788\n",
            "step 2769/5000, loss: 0.6425\n",
            "step 2770/5000, loss: 0.7493\n",
            "step 2771/5000, loss: 0.6570\n",
            "step 2772/5000, loss: 0.5918\n",
            "step 2773/5000, loss: 0.8043\n",
            "step 2774/5000, loss: 0.7252\n",
            "step 2775/5000, loss: 0.6614\n",
            "step 2776/5000, loss: 0.7759\n",
            "step 2777/5000, loss: 1.0467\n",
            "step 2778/5000, loss: 0.6844\n",
            "step 2779/5000, loss: 0.7382\n",
            "step 2780/5000, loss: 0.5774\n",
            "step 2781/5000, loss: 0.6825\n",
            "step 2782/5000, loss: 0.8234\n",
            "step 2783/5000, loss: 0.9100\n",
            "step 2784/5000, loss: 0.8184\n",
            "step 2785/5000, loss: 0.8024\n",
            "step 2786/5000, loss: 0.7542\n",
            "step 2787/5000, loss: 0.7218\n",
            "step 2788/5000, loss: 0.6223\n",
            "step 2789/5000, loss: 0.7027\n",
            "step 2790/5000, loss: 0.7505\n",
            "step 2791/5000, loss: 0.7116\n",
            "step 2792/5000, loss: 0.6257\n",
            "step 2793/5000, loss: 0.6784\n",
            "step 2794/5000, loss: 0.6876\n",
            "step 2795/5000, loss: 0.6289\n",
            "step 2796/5000, loss: 1.1045\n",
            "step 2797/5000, loss: 0.7352\n",
            "step 2798/5000, loss: 0.6299\n",
            "step 2799/5000, loss: 0.7958\n",
            "step 2800/5000, loss: 0.8063\n",
            "step 2801/5000, loss: 0.7099\n",
            "step 2802/5000, loss: 0.5484\n",
            "step 2803/5000, loss: 0.7981\n",
            "step 2804/5000, loss: 0.8165\n",
            "step 2805/5000, loss: 0.8306\n",
            "step 2806/5000, loss: 0.6786\n",
            "step 2807/5000, loss: 0.7250\n",
            "step 2808/5000, loss: 0.6566\n",
            "step 2809/5000, loss: 0.7845\n",
            "step 2810/5000, loss: 0.7085\n",
            "step 2811/5000, loss: 0.7512\n",
            "step 2812/5000, loss: 0.6719\n",
            "step 2813/5000, loss: 0.6583\n",
            "step 2814/5000, loss: 0.6819\n",
            "step 2815/5000, loss: 0.6610\n",
            "step 2816/5000, loss: 0.7455\n",
            "step 2817/5000, loss: 0.6681\n",
            "step 2818/5000, loss: 0.6357\n",
            "step 2819/5000, loss: 0.6361\n",
            "step 2820/5000, loss: 0.8312\n",
            "step 2821/5000, loss: 0.7273\n",
            "step 2822/5000, loss: 0.6724\n",
            "step 2823/5000, loss: 0.6525\n",
            "step 2824/5000, loss: 0.8020\n",
            "step 2825/5000, loss: 0.9086\n",
            "step 2826/5000, loss: 0.7013\n",
            "step 2827/5000, loss: 0.8159\n",
            "step 2828/5000, loss: 0.7241\n",
            "step 2829/5000, loss: 0.7059\n",
            "step 2830/5000, loss: 0.6617\n",
            "step 2831/5000, loss: 0.6970\n",
            "step 2832/5000, loss: 0.7276\n",
            "step 2833/5000, loss: 0.6914\n",
            "step 2834/5000, loss: 0.5960\n",
            "step 2835/5000, loss: 0.7561\n",
            "step 2836/5000, loss: 0.5747\n",
            "step 2837/5000, loss: 0.7314\n",
            "step 2838/5000, loss: 0.7092\n",
            "step 2839/5000, loss: 0.9800\n",
            "step 2840/5000, loss: 0.6823\n",
            "step 2841/5000, loss: 0.6874\n",
            "step 2842/5000, loss: 0.6499\n",
            "step 2843/5000, loss: 0.7558\n",
            "step 2844/5000, loss: 0.5882\n",
            "step 2845/5000, loss: 0.6699\n",
            "step 2846/5000, loss: 0.9446\n",
            "step 2847/5000, loss: 0.7374\n",
            "step 2848/5000, loss: 0.8943\n",
            "step 2849/5000, loss: 0.5573\n",
            "step 2850/5000, loss: 0.5856\n",
            "step 2851/5000, loss: 0.6602\n",
            "step 2852/5000, loss: 0.6463\n",
            "step 2853/5000, loss: 0.7339\n",
            "step 2854/5000, loss: 0.7355\n",
            "step 2855/5000, loss: 0.7838\n",
            "step 2856/5000, loss: 1.6933\n",
            "step 2857/5000, loss: 0.7677\n",
            "step 2858/5000, loss: 0.6905\n",
            "step 2859/5000, loss: 0.7390\n",
            "step 2860/5000, loss: 0.6584\n",
            "step 2861/5000, loss: 0.7317\n",
            "step 2862/5000, loss: 0.7134\n",
            "step 2863/5000, loss: 0.7451\n",
            "step 2864/5000, loss: 0.6409\n",
            "step 2865/5000, loss: 0.8864\n",
            "step 2866/5000, loss: 0.7454\n",
            "step 2867/5000, loss: 0.6700\n",
            "step 2868/5000, loss: 0.7211\n",
            "step 2869/5000, loss: 2.0693\n",
            "step 2870/5000, loss: 0.7281\n",
            "step 2871/5000, loss: 0.6375\n",
            "step 2872/5000, loss: 0.6256\n",
            "step 2873/5000, loss: 0.6455\n",
            "step 2874/5000, loss: 0.8065\n",
            "step 2875/5000, loss: 0.6393\n",
            "step 2876/5000, loss: 0.6770\n",
            "step 2877/5000, loss: 0.7299\n",
            "step 2878/5000, loss: 0.7732\n",
            "step 2879/5000, loss: 0.9409\n",
            "step 2880/5000, loss: 0.7739\n",
            "step 2881/5000, loss: 0.7580\n",
            "step 2882/5000, loss: 0.7687\n",
            "step 2883/5000, loss: 0.9284\n",
            "step 2884/5000, loss: 0.7013\n",
            "step 2885/5000, loss: 0.8203\n",
            "step 2886/5000, loss: 0.7759\n",
            "step 2887/5000, loss: 0.6521\n",
            "step 2888/5000, loss: 0.7673\n",
            "step 2889/5000, loss: 0.6710\n",
            "step 2890/5000, loss: 0.7532\n",
            "step 2891/5000, loss: 0.8036\n",
            "step 2892/5000, loss: 0.7474\n",
            "step 2893/5000, loss: 0.7459\n",
            "step 2894/5000, loss: 0.6891\n",
            "step 2895/5000, loss: 0.7742\n",
            "step 2896/5000, loss: 0.5713\n",
            "step 2897/5000, loss: 0.7099\n",
            "step 2898/5000, loss: 0.7165\n",
            "step 2899/5000, loss: 0.7252\n",
            "step 2900/5000, loss: 0.7618\n",
            "step 2901/5000, loss: 0.6491\n",
            "step 2902/5000, loss: 0.6709\n",
            "step 2903/5000, loss: 0.6309\n",
            "step 2904/5000, loss: 0.8272\n",
            "step 2905/5000, loss: 0.7691\n",
            "step 2906/5000, loss: 0.7244\n",
            "step 2907/5000, loss: 0.7010\n",
            "step 2908/5000, loss: 0.6426\n",
            "step 2909/5000, loss: 0.6988\n",
            "step 2910/5000, loss: 0.6108\n",
            "step 2911/5000, loss: 0.7163\n",
            "step 2912/5000, loss: 0.8597\n",
            "step 2913/5000, loss: 0.6891\n",
            "step 2914/5000, loss: 0.6517\n",
            "step 2915/5000, loss: 0.8581\n",
            "step 2916/5000, loss: 0.7716\n",
            "step 2917/5000, loss: 0.7362\n",
            "step 2918/5000, loss: 0.7031\n",
            "step 2919/5000, loss: 0.8296\n",
            "step 2920/5000, loss: 0.7700\n",
            "step 2921/5000, loss: 0.8876\n",
            "step 2922/5000, loss: 0.8012\n",
            "step 2923/5000, loss: 0.7181\n",
            "step 2924/5000, loss: 0.7032\n",
            "step 2925/5000, loss: 0.6427\n",
            "step 2926/5000, loss: 0.6561\n",
            "step 2927/5000, loss: 0.6541\n",
            "step 2928/5000, loss: 0.8333\n",
            "step 2929/5000, loss: 0.7178\n",
            "step 2930/5000, loss: 0.6846\n",
            "step 2931/5000, loss: 0.7902\n",
            "step 2932/5000, loss: 0.9218\n",
            "step 2933/5000, loss: 0.7351\n",
            "step 2934/5000, loss: 0.8497\n",
            "step 2935/5000, loss: 0.7716\n",
            "step 2936/5000, loss: 0.8077\n",
            "step 2937/5000, loss: 0.7131\n",
            "step 2938/5000, loss: 0.7096\n",
            "step 2939/5000, loss: 0.9427\n",
            "step 2940/5000, loss: 1.1112\n",
            "step 2941/5000, loss: 0.6922\n",
            "step 2942/5000, loss: 0.5660\n",
            "step 2943/5000, loss: 0.6479\n",
            "step 2944/5000, loss: 0.7058\n",
            "step 2945/5000, loss: 0.6801\n",
            "step 2946/5000, loss: 0.7157\n",
            "step 2947/5000, loss: 0.6796\n",
            "step 2948/5000, loss: 0.7902\n",
            "step 2949/5000, loss: 0.7196\n",
            "step 2950/5000, loss: 0.6179\n",
            "step 2951/5000, loss: 0.6401\n",
            "step 2952/5000, loss: 1.8279\n",
            "step 2953/5000, loss: 0.6570\n",
            "step 2954/5000, loss: 0.6142\n",
            "step 2955/5000, loss: 0.7380\n",
            "step 2956/5000, loss: 0.7151\n",
            "step 2957/5000, loss: 0.6847\n",
            "step 2958/5000, loss: 0.6727\n",
            "step 2959/5000, loss: 0.7247\n",
            "step 2960/5000, loss: 0.7100\n",
            "step 2961/5000, loss: 0.7126\n",
            "step 2962/5000, loss: 0.6255\n",
            "step 2963/5000, loss: 0.6312\n",
            "step 2964/5000, loss: 0.6164\n",
            "step 2965/5000, loss: 0.6970\n",
            "step 2966/5000, loss: 1.0191\n",
            "step 2967/5000, loss: 0.7696\n",
            "step 2968/5000, loss: 0.9026\n",
            "step 2969/5000, loss: 0.8772\n",
            "step 2970/5000, loss: 0.7011\n",
            "step 2971/5000, loss: 0.5587\n",
            "step 2972/5000, loss: 0.6990\n",
            "step 2973/5000, loss: 0.8955\n",
            "step 2974/5000, loss: 0.7098\n",
            "step 2975/5000, loss: 0.7257\n",
            "step 2976/5000, loss: 0.6785\n",
            "step 2977/5000, loss: 0.7326\n",
            "step 2978/5000, loss: 0.6707\n",
            "step 2979/5000, loss: 0.7641\n",
            "step 2980/5000, loss: 0.6921\n",
            "step 2981/5000, loss: 0.8159\n",
            "step 2982/5000, loss: 0.6566\n",
            "step 2983/5000, loss: 0.7870\n",
            "step 2984/5000, loss: 0.7378\n",
            "step 2985/5000, loss: 0.8195\n",
            "step 2986/5000, loss: 0.8413\n",
            "step 2987/5000, loss: 0.6976\n",
            "step 2988/5000, loss: 0.8448\n",
            "step 2989/5000, loss: 0.6543\n",
            "step 2990/5000, loss: 0.7150\n",
            "step 2991/5000, loss: 0.5908\n",
            "step 2992/5000, loss: 0.7911\n",
            "step 2993/5000, loss: 0.9335\n",
            "step 2994/5000, loss: 0.9137\n",
            "step 2995/5000, loss: 0.7794\n",
            "step 2996/5000, loss: 0.6877\n",
            "step 2997/5000, loss: 0.7156\n",
            "step 2998/5000, loss: 0.7094\n",
            "step 2999/5000, loss: 0.6686\n",
            "step 3000/5000, loss: 0.7146\n",
            "step 3001/5000, loss: 0.6025\n",
            "step 3002/5000, loss: 1.0311\n",
            "step 3003/5000, loss: 0.7545\n",
            "step 3004/5000, loss: 0.7142\n",
            "step 3005/5000, loss: 0.5943\n",
            "step 3006/5000, loss: 0.7167\n",
            "step 3007/5000, loss: 0.7335\n",
            "step 3008/5000, loss: 0.7290\n",
            "step 3009/5000, loss: 0.6898\n",
            "step 3010/5000, loss: 0.6671\n",
            "step 3011/5000, loss: 0.6967\n",
            "step 3012/5000, loss: 0.6231\n",
            "step 3013/5000, loss: 0.7121\n",
            "step 3014/5000, loss: 0.6851\n",
            "step 3015/5000, loss: 0.6727\n",
            "step 3016/5000, loss: 0.6536\n",
            "step 3017/5000, loss: 0.6947\n",
            "step 3018/5000, loss: 0.7225\n",
            "step 3019/5000, loss: 0.8469\n",
            "step 3020/5000, loss: 0.8180\n",
            "step 3021/5000, loss: 0.7916\n",
            "step 3022/5000, loss: 0.7407\n",
            "step 3023/5000, loss: 0.5985\n",
            "step 3024/5000, loss: 0.7373\n",
            "step 3025/5000, loss: 0.8919\n",
            "step 3026/5000, loss: 0.7248\n",
            "step 3027/5000, loss: 0.7376\n",
            "step 3028/5000, loss: 0.7496\n",
            "step 3029/5000, loss: 0.7221\n",
            "step 3030/5000, loss: 0.6614\n",
            "step 3031/5000, loss: 0.7145\n",
            "step 3032/5000, loss: 0.9072\n",
            "step 3033/5000, loss: 0.6518\n",
            "step 3034/5000, loss: 0.6900\n",
            "step 3035/5000, loss: 0.6305\n",
            "step 3036/5000, loss: 0.8061\n",
            "step 3037/5000, loss: 0.6795\n",
            "step 3038/5000, loss: 0.6723\n",
            "step 3039/5000, loss: 0.6490\n",
            "step 3040/5000, loss: 0.8005\n",
            "step 3041/5000, loss: 0.6577\n",
            "step 3042/5000, loss: 1.0183\n",
            "step 3043/5000, loss: 0.7746\n",
            "step 3044/5000, loss: 0.7131\n",
            "step 3045/5000, loss: 0.7763\n",
            "step 3046/5000, loss: 0.7965\n",
            "step 3047/5000, loss: 0.8076\n",
            "step 3048/5000, loss: 0.8118\n",
            "step 3049/5000, loss: 0.5328\n",
            "step 3050/5000, loss: 0.7175\n",
            "step 3051/5000, loss: 0.6844\n",
            "step 3052/5000, loss: 0.6982\n",
            "step 3053/5000, loss: 0.7556\n",
            "step 3054/5000, loss: 0.7268\n",
            "step 3055/5000, loss: 0.6445\n",
            "step 3056/5000, loss: 0.7105\n",
            "step 3057/5000, loss: 0.7225\n",
            "step 3058/5000, loss: 0.6541\n",
            "step 3059/5000, loss: 0.6997\n",
            "step 3060/5000, loss: 0.8549\n",
            "step 3061/5000, loss: 0.6519\n",
            "step 3062/5000, loss: 0.9491\n",
            "step 3063/5000, loss: 0.6834\n",
            "step 3064/5000, loss: 0.7324\n",
            "step 3065/5000, loss: 0.6939\n",
            "step 3066/5000, loss: 0.7690\n",
            "step 3067/5000, loss: 0.6599\n",
            "step 3068/5000, loss: 0.7251\n",
            "step 3069/5000, loss: 0.8351\n",
            "step 3070/5000, loss: 0.7379\n",
            "step 3071/5000, loss: 1.0835\n",
            "step 3072/5000, loss: 0.5997\n",
            "step 3073/5000, loss: 0.8936\n",
            "step 3074/5000, loss: 0.6166\n",
            "step 3075/5000, loss: 0.7558\n",
            "step 3076/5000, loss: 0.6224\n",
            "step 3077/5000, loss: 0.7471\n",
            "step 3078/5000, loss: 0.5890\n",
            "step 3079/5000, loss: 0.6849\n",
            "step 3080/5000, loss: 0.8150\n",
            "step 3081/5000, loss: 0.7854\n",
            "step 3082/5000, loss: 0.5747\n",
            "step 3083/5000, loss: 0.7537\n",
            "step 3084/5000, loss: 0.6958\n",
            "step 3085/5000, loss: 0.9361\n",
            "step 3086/5000, loss: 0.8898\n",
            "step 3087/5000, loss: 0.6951\n",
            "step 3088/5000, loss: 0.8175\n",
            "step 3089/5000, loss: 0.7319\n",
            "step 3090/5000, loss: 0.8339\n",
            "step 3091/5000, loss: 0.6879\n",
            "step 3092/5000, loss: 0.6394\n",
            "step 3093/5000, loss: 0.7567\n",
            "step 3094/5000, loss: 0.7574\n",
            "step 3095/5000, loss: 0.7332\n",
            "step 3096/5000, loss: 0.8799\n",
            "step 3097/5000, loss: 0.7408\n",
            "step 3098/5000, loss: 0.9244\n",
            "step 3099/5000, loss: 0.7409\n",
            "step 3100/5000, loss: 0.6092\n",
            "step 3101/5000, loss: 0.7385\n",
            "step 3102/5000, loss: 0.7409\n",
            "step 3103/5000, loss: 0.7498\n",
            "step 3104/5000, loss: 0.6987\n",
            "step 3105/5000, loss: 0.7940\n",
            "step 3106/5000, loss: 0.6253\n",
            "step 3107/5000, loss: 0.6742\n",
            "step 3108/5000, loss: 0.9155\n",
            "step 3109/5000, loss: 0.6381\n",
            "step 3110/5000, loss: 0.6831\n",
            "step 3111/5000, loss: 0.7699\n",
            "step 3112/5000, loss: 0.7169\n",
            "step 3113/5000, loss: 0.6520\n",
            "step 3114/5000, loss: 0.7791\n",
            "step 3115/5000, loss: 0.6571\n",
            "step 3116/5000, loss: 0.6076\n",
            "step 3117/5000, loss: 0.6205\n",
            "step 3118/5000, loss: 0.7273\n",
            "step 3119/5000, loss: 0.7083\n",
            "step 3120/5000, loss: 0.6645\n",
            "step 3121/5000, loss: 0.5595\n",
            "step 3122/5000, loss: 1.7068\n",
            "step 3123/5000, loss: 0.6405\n",
            "step 3124/5000, loss: 0.6517\n",
            "step 3125/5000, loss: 0.8499\n",
            "step 3126/5000, loss: 0.6680\n",
            "step 3127/5000, loss: 0.6815\n",
            "step 3128/5000, loss: 0.7754\n",
            "step 3129/5000, loss: 0.7342\n",
            "step 3130/5000, loss: 0.8580\n",
            "step 3131/5000, loss: 0.6887\n",
            "step 3132/5000, loss: 0.6774\n",
            "step 3133/5000, loss: 0.6625\n",
            "step 3134/5000, loss: 0.6616\n",
            "step 3135/5000, loss: 0.7224\n",
            "step 3136/5000, loss: 0.5636\n",
            "step 3137/5000, loss: 0.9335\n",
            "step 3138/5000, loss: 0.6822\n",
            "step 3139/5000, loss: 0.6854\n",
            "step 3140/5000, loss: 0.7226\n",
            "step 3141/5000, loss: 0.8109\n",
            "step 3142/5000, loss: 0.6297\n",
            "step 3143/5000, loss: 0.7842\n",
            "step 3144/5000, loss: 0.6993\n",
            "step 3145/5000, loss: 0.8148\n",
            "step 3146/5000, loss: 0.6437\n",
            "step 3147/5000, loss: 0.6593\n",
            "step 3148/5000, loss: 0.6404\n",
            "step 3149/5000, loss: 0.6536\n",
            "step 3150/5000, loss: 0.6149\n",
            "step 3151/5000, loss: 0.6842\n",
            "step 3152/5000, loss: 1.0438\n",
            "step 3153/5000, loss: 0.7185\n",
            "step 3154/5000, loss: 0.6595\n",
            "step 3155/5000, loss: 0.7149\n",
            "step 3156/5000, loss: 0.6838\n",
            "step 3157/5000, loss: 0.7346\n",
            "step 3158/5000, loss: 0.5913\n",
            "step 3159/5000, loss: 0.8174\n",
            "step 3160/5000, loss: 0.7321\n",
            "step 3161/5000, loss: 0.6777\n",
            "step 3162/5000, loss: 0.7446\n",
            "step 3163/5000, loss: 0.8022\n",
            "step 3164/5000, loss: 0.6191\n",
            "step 3165/5000, loss: 0.7065\n",
            "step 3166/5000, loss: 0.7252\n",
            "step 3167/5000, loss: 0.5903\n",
            "step 3168/5000, loss: 0.6633\n",
            "step 3169/5000, loss: 0.6649\n",
            "step 3170/5000, loss: 0.6710\n",
            "step 3171/5000, loss: 0.6657\n",
            "step 3172/5000, loss: 0.7789\n",
            "step 3173/5000, loss: 0.6243\n",
            "step 3174/5000, loss: 0.5504\n",
            "step 3175/5000, loss: 0.7075\n",
            "step 3176/5000, loss: 0.7567\n",
            "step 3177/5000, loss: 0.6865\n",
            "step 3178/5000, loss: 0.5832\n",
            "step 3179/5000, loss: 0.6280\n",
            "step 3180/5000, loss: 0.8468\n",
            "step 3181/5000, loss: 0.7403\n",
            "step 3182/5000, loss: 0.7892\n",
            "step 3183/5000, loss: 0.9363\n",
            "step 3184/5000, loss: 0.9072\n",
            "step 3185/5000, loss: 0.6795\n",
            "step 3186/5000, loss: 0.6927\n",
            "step 3187/5000, loss: 0.6657\n",
            "step 3188/5000, loss: 0.6291\n",
            "step 3189/5000, loss: 0.7648\n",
            "step 3190/5000, loss: 0.7230\n",
            "step 3191/5000, loss: 0.7422\n",
            "step 3192/5000, loss: 0.6913\n",
            "step 3193/5000, loss: 0.6620\n",
            "step 3194/5000, loss: 0.7933\n",
            "step 3195/5000, loss: 0.6813\n",
            "step 3196/5000, loss: 1.0165\n",
            "step 3197/5000, loss: 0.6946\n",
            "step 3198/5000, loss: 0.6045\n",
            "step 3199/5000, loss: 0.5847\n",
            "step 3200/5000, loss: 0.6740\n",
            "step 3201/5000, loss: 0.6743\n",
            "step 3202/5000, loss: 0.6995\n",
            "step 3203/5000, loss: 0.7097\n",
            "step 3204/5000, loss: 0.8243\n",
            "step 3205/5000, loss: 0.6984\n",
            "step 3206/5000, loss: 0.6987\n",
            "step 3207/5000, loss: 0.6928\n",
            "step 3208/5000, loss: 0.7208\n",
            "step 3209/5000, loss: 0.6133\n",
            "step 3210/5000, loss: 0.7237\n",
            "step 3211/5000, loss: 0.6999\n",
            "step 3212/5000, loss: 0.7436\n",
            "step 3213/5000, loss: 0.7427\n",
            "step 3214/5000, loss: 0.8367\n",
            "step 3215/5000, loss: 0.8339\n",
            "step 3216/5000, loss: 0.6121\n",
            "step 3217/5000, loss: 0.6981\n",
            "step 3218/5000, loss: 0.7326\n",
            "step 3219/5000, loss: 0.6330\n",
            "step 3220/5000, loss: 0.7206\n",
            "step 3221/5000, loss: 0.8921\n",
            "step 3222/5000, loss: 0.6613\n",
            "step 3223/5000, loss: 0.7583\n",
            "step 3224/5000, loss: 2.0397\n",
            "step 3225/5000, loss: 0.8785\n",
            "step 3226/5000, loss: 0.6881\n",
            "step 3227/5000, loss: 0.8375\n",
            "step 3228/5000, loss: 0.6811\n",
            "step 3229/5000, loss: 0.7206\n",
            "step 3230/5000, loss: 0.6393\n",
            "step 3231/5000, loss: 0.7414\n",
            "step 3232/5000, loss: 0.7067\n",
            "step 3233/5000, loss: 0.6650\n",
            "step 3234/5000, loss: 0.6170\n",
            "step 3235/5000, loss: 0.6927\n",
            "step 3236/5000, loss: 0.5860\n",
            "step 3237/5000, loss: 0.9248\n",
            "step 3238/5000, loss: 0.8279\n",
            "step 3239/5000, loss: 0.5830\n",
            "step 3240/5000, loss: 0.7218\n",
            "step 3241/5000, loss: 0.7276\n",
            "step 3242/5000, loss: 0.9975\n",
            "step 3243/5000, loss: 0.7501\n",
            "step 3244/5000, loss: 0.6282\n",
            "step 3245/5000, loss: 0.6173\n",
            "step 3246/5000, loss: 0.5964\n",
            "step 3247/5000, loss: 0.6653\n",
            "step 3248/5000, loss: 0.9288\n",
            "step 3249/5000, loss: 0.6808\n",
            "step 3250/5000, loss: 0.6395\n",
            "step 3251/5000, loss: 0.6321\n",
            "step 3252/5000, loss: 0.8454\n",
            "step 3253/5000, loss: 0.6574\n",
            "step 3254/5000, loss: 0.8470\n",
            "step 3255/5000, loss: 0.6069\n",
            "step 3256/5000, loss: 0.7314\n",
            "step 3257/5000, loss: 0.7610\n",
            "step 3258/5000, loss: 0.8155\n",
            "step 3259/5000, loss: 0.6432\n",
            "step 3260/5000, loss: 0.5928\n",
            "step 3261/5000, loss: 0.6632\n",
            "step 3262/5000, loss: 0.6529\n",
            "step 3263/5000, loss: 0.5616\n",
            "step 3264/5000, loss: 0.9290\n",
            "step 3265/5000, loss: 0.7034\n",
            "step 3266/5000, loss: 0.6465\n",
            "step 3267/5000, loss: 0.5537\n",
            "step 3268/5000, loss: 0.6246\n",
            "step 3269/5000, loss: 0.6729\n",
            "step 3270/5000, loss: 0.5452\n",
            "step 3271/5000, loss: 0.6514\n",
            "step 3272/5000, loss: 0.6597\n",
            "step 3273/5000, loss: 0.7721\n",
            "step 3274/5000, loss: 0.6083\n",
            "step 3275/5000, loss: 0.6383\n",
            "step 3276/5000, loss: 0.6809\n",
            "step 3277/5000, loss: 0.6518\n",
            "step 3278/5000, loss: 0.8050\n",
            "step 3279/5000, loss: 0.7033\n",
            "step 3280/5000, loss: 0.6329\n",
            "step 3281/5000, loss: 0.7430\n",
            "step 3282/5000, loss: 0.6677\n",
            "step 3283/5000, loss: 0.6770\n",
            "step 3284/5000, loss: 0.6887\n",
            "step 3285/5000, loss: 0.6222\n",
            "step 3286/5000, loss: 0.6056\n",
            "step 3287/5000, loss: 0.6763\n",
            "step 3288/5000, loss: 0.7069\n",
            "step 3289/5000, loss: 0.7918\n",
            "step 3290/5000, loss: 0.7980\n",
            "step 3291/5000, loss: 0.7392\n",
            "step 3292/5000, loss: 0.8091\n",
            "step 3293/5000, loss: 0.6928\n",
            "step 3294/5000, loss: 0.6912\n",
            "step 3295/5000, loss: 0.8268\n",
            "step 3296/5000, loss: 0.7291\n",
            "step 3297/5000, loss: 0.6115\n",
            "step 3298/5000, loss: 0.8791\n",
            "step 3299/5000, loss: 0.7722\n",
            "step 3300/5000, loss: 0.6597\n",
            "step 3301/5000, loss: 1.5944\n",
            "step 3302/5000, loss: 0.9233\n",
            "step 3303/5000, loss: 0.8182\n",
            "step 3304/5000, loss: 0.6375\n",
            "step 3305/5000, loss: 0.6381\n",
            "step 3306/5000, loss: 0.7221\n",
            "step 3307/5000, loss: 0.8150\n",
            "step 3308/5000, loss: 0.8741\n",
            "step 3309/5000, loss: 0.7095\n",
            "step 3310/5000, loss: 0.6105\n",
            "step 3311/5000, loss: 0.5998\n",
            "step 3312/5000, loss: 0.7219\n",
            "step 3313/5000, loss: 0.7446\n",
            "step 3314/5000, loss: 0.7210\n",
            "step 3315/5000, loss: 0.7415\n",
            "step 3316/5000, loss: 0.9119\n",
            "step 3317/5000, loss: 0.7102\n",
            "step 3318/5000, loss: 0.7339\n",
            "step 3319/5000, loss: 0.6785\n",
            "step 3320/5000, loss: 0.6510\n",
            "step 3321/5000, loss: 0.8059\n",
            "step 3322/5000, loss: 0.7525\n",
            "step 3323/5000, loss: 0.6865\n",
            "step 3324/5000, loss: 0.9477\n",
            "step 3325/5000, loss: 0.7489\n",
            "step 3326/5000, loss: 0.6519\n",
            "step 3327/5000, loss: 0.7181\n",
            "step 3328/5000, loss: 0.7566\n",
            "step 3329/5000, loss: 0.6675\n",
            "step 3330/5000, loss: 0.7565\n",
            "step 3331/5000, loss: 0.6879\n",
            "step 3332/5000, loss: 0.7139\n",
            "step 3333/5000, loss: 0.7650\n",
            "step 3334/5000, loss: 0.7114\n",
            "step 3335/5000, loss: 0.7411\n",
            "step 3336/5000, loss: 0.6677\n",
            "step 3337/5000, loss: 0.7520\n",
            "step 3338/5000, loss: 0.7479\n",
            "step 3339/5000, loss: 0.7608\n",
            "step 3340/5000, loss: 0.7608\n",
            "step 3341/5000, loss: 0.5852\n",
            "step 3342/5000, loss: 0.6112\n",
            "step 3343/5000, loss: 0.5816\n",
            "step 3344/5000, loss: 0.8132\n",
            "step 3345/5000, loss: 0.7352\n",
            "step 3346/5000, loss: 0.7140\n",
            "step 3347/5000, loss: 0.7533\n",
            "step 3348/5000, loss: 0.6640\n",
            "step 3349/5000, loss: 0.6555\n",
            "step 3350/5000, loss: 0.5608\n",
            "step 3351/5000, loss: 0.5672\n",
            "step 3352/5000, loss: 0.6957\n",
            "step 3353/5000, loss: 0.6833\n",
            "step 3354/5000, loss: 0.7065\n",
            "step 3355/5000, loss: 0.7057\n",
            "step 3356/5000, loss: 0.6375\n",
            "step 3357/5000, loss: 0.7451\n",
            "step 3358/5000, loss: 0.7180\n",
            "step 3359/5000, loss: 0.7208\n",
            "step 3360/5000, loss: 0.8420\n",
            "step 3361/5000, loss: 0.6749\n",
            "step 3362/5000, loss: 0.5592\n",
            "step 3363/5000, loss: 0.6863\n",
            "step 3364/5000, loss: 0.7363\n",
            "step 3365/5000, loss: 0.7302\n",
            "step 3366/5000, loss: 0.5766\n",
            "step 3367/5000, loss: 0.6717\n",
            "step 3368/5000, loss: 0.8924\n",
            "step 3369/5000, loss: 0.7110\n",
            "step 3370/5000, loss: 0.6722\n",
            "step 3371/5000, loss: 0.7383\n",
            "step 3372/5000, loss: 0.8253\n",
            "step 3373/5000, loss: 0.7181\n",
            "step 3374/5000, loss: 0.7376\n",
            "step 3375/5000, loss: 0.7283\n",
            "step 3376/5000, loss: 0.6929\n",
            "step 3377/5000, loss: 0.7489\n",
            "step 3378/5000, loss: 0.6562\n",
            "step 3379/5000, loss: 0.7662\n",
            "step 3380/5000, loss: 0.5807\n",
            "step 3381/5000, loss: 0.6729\n",
            "step 3382/5000, loss: 0.6671\n",
            "step 3383/5000, loss: 0.7343\n",
            "step 3384/5000, loss: 0.7353\n",
            "step 3385/5000, loss: 0.6529\n",
            "step 3386/5000, loss: 0.6553\n",
            "step 3387/5000, loss: 0.6816\n",
            "step 3388/5000, loss: 0.7707\n",
            "step 3389/5000, loss: 0.5636\n",
            "step 3390/5000, loss: 0.7221\n",
            "step 3391/5000, loss: 0.7555\n",
            "step 3392/5000, loss: 0.8340\n",
            "step 3393/5000, loss: 0.8945\n",
            "step 3394/5000, loss: 0.6648\n",
            "step 3395/5000, loss: 0.7817\n",
            "step 3396/5000, loss: 0.7320\n",
            "step 3397/5000, loss: 0.7078\n",
            "step 3398/5000, loss: 0.5890\n",
            "step 3399/5000, loss: 0.7407\n",
            "step 3400/5000, loss: 0.7692\n",
            "step 3401/5000, loss: 0.8375\n",
            "step 3402/5000, loss: 0.6906\n",
            "step 3403/5000, loss: 0.7525\n",
            "step 3404/5000, loss: 0.7652\n",
            "step 3405/5000, loss: 0.5860\n",
            "step 3406/5000, loss: 0.6582\n",
            "step 3407/5000, loss: 0.7147\n",
            "step 3408/5000, loss: 0.6162\n",
            "step 3409/5000, loss: 0.5548\n",
            "step 3410/5000, loss: 0.8545\n",
            "step 3411/5000, loss: 1.3030\n",
            "step 3412/5000, loss: 0.8251\n",
            "step 3413/5000, loss: 0.5970\n",
            "step 3414/5000, loss: 0.6615\n",
            "step 3415/5000, loss: 0.7414\n",
            "step 3416/5000, loss: 0.6192\n",
            "step 3417/5000, loss: 2.1294\n",
            "step 3418/5000, loss: 0.7561\n",
            "step 3419/5000, loss: 0.6184\n",
            "step 3420/5000, loss: 0.7823\n",
            "step 3421/5000, loss: 0.6478\n",
            "step 3422/5000, loss: 0.6080\n",
            "step 3423/5000, loss: 0.6040\n",
            "step 3424/5000, loss: 0.5879\n",
            "step 3425/5000, loss: 0.7245\n",
            "step 3426/5000, loss: 0.7169\n",
            "step 3427/5000, loss: 0.6720\n",
            "step 3428/5000, loss: 0.6687\n",
            "step 3429/5000, loss: 0.7591\n",
            "step 3430/5000, loss: 0.8338\n",
            "step 3431/5000, loss: 0.6157\n",
            "step 3432/5000, loss: 0.7013\n",
            "step 3433/5000, loss: 0.6420\n",
            "step 3434/5000, loss: 0.6707\n",
            "step 3435/5000, loss: 0.7160\n",
            "step 3436/5000, loss: 0.7401\n",
            "step 3437/5000, loss: 0.8373\n",
            "step 3438/5000, loss: 0.6272\n",
            "step 3439/5000, loss: 0.6836\n",
            "step 3440/5000, loss: 0.7889\n",
            "step 3441/5000, loss: 0.6906\n",
            "step 3442/5000, loss: 0.6795\n",
            "step 3443/5000, loss: 0.5855\n",
            "step 3444/5000, loss: 0.7111\n",
            "step 3445/5000, loss: 0.7090\n",
            "step 3446/5000, loss: 0.5959\n",
            "step 3447/5000, loss: 0.6518\n",
            "step 3448/5000, loss: 0.8012\n",
            "step 3449/5000, loss: 0.6588\n",
            "step 3450/5000, loss: 0.7742\n",
            "step 3451/5000, loss: 0.7070\n",
            "step 3452/5000, loss: 0.6905\n",
            "step 3453/5000, loss: 0.9496\n",
            "step 3454/5000, loss: 0.6234\n",
            "step 3455/5000, loss: 0.7051\n",
            "step 3456/5000, loss: 0.8039\n",
            "step 3457/5000, loss: 0.6535\n",
            "step 3458/5000, loss: 0.9741\n",
            "step 3459/5000, loss: 0.6953\n",
            "step 3460/5000, loss: 0.6725\n",
            "step 3461/5000, loss: 0.7506\n",
            "step 3462/5000, loss: 0.7945\n",
            "step 3463/5000, loss: 0.7864\n",
            "step 3464/5000, loss: 0.7322\n",
            "step 3465/5000, loss: 0.9068\n",
            "step 3466/5000, loss: 0.6731\n",
            "step 3467/5000, loss: 0.6715\n",
            "step 3468/5000, loss: 0.8919\n",
            "step 3469/5000, loss: 0.7323\n",
            "step 3470/5000, loss: 0.6584\n",
            "step 3471/5000, loss: 0.6866\n",
            "step 3472/5000, loss: 0.6490\n",
            "step 3473/5000, loss: 0.6923\n",
            "step 3474/5000, loss: 0.6793\n",
            "step 3475/5000, loss: 0.6779\n",
            "step 3476/5000, loss: 0.6120\n",
            "step 3477/5000, loss: 0.7301\n",
            "step 3478/5000, loss: 0.6744\n",
            "step 3479/5000, loss: 0.7529\n",
            "step 3480/5000, loss: 0.6723\n",
            "step 3481/5000, loss: 0.5923\n",
            "step 3482/5000, loss: 0.6794\n",
            "step 3483/5000, loss: 0.5734\n",
            "step 3484/5000, loss: 0.7210\n",
            "step 3485/5000, loss: 0.8401\n",
            "step 3486/5000, loss: 0.7686\n",
            "step 3487/5000, loss: 0.6910\n",
            "step 3488/5000, loss: 0.6392\n",
            "step 3489/5000, loss: 0.7075\n",
            "step 3490/5000, loss: 0.7004\n",
            "step 3491/5000, loss: 0.7185\n",
            "step 3492/5000, loss: 0.7355\n",
            "step 3493/5000, loss: 0.6750\n",
            "step 3494/5000, loss: 0.7758\n",
            "step 3495/5000, loss: 0.7663\n",
            "step 3496/5000, loss: 0.5770\n",
            "step 3497/5000, loss: 0.7308\n",
            "step 3498/5000, loss: 0.5954\n",
            "step 3499/5000, loss: 0.6037\n",
            "step 3500/5000, loss: 0.9283\n",
            "step 3501/5000, loss: 0.8198\n",
            "step 3502/5000, loss: 0.8653\n",
            "step 3503/5000, loss: 0.8870\n",
            "step 3504/5000, loss: 0.7286\n",
            "step 3505/5000, loss: 0.8107\n",
            "step 3506/5000, loss: 0.5999\n",
            "step 3507/5000, loss: 0.7175\n",
            "step 3508/5000, loss: 0.8964\n",
            "step 3509/5000, loss: 0.6965\n",
            "step 3510/5000, loss: 0.7014\n",
            "step 3511/5000, loss: 0.6388\n",
            "step 3512/5000, loss: 0.6131\n",
            "step 3513/5000, loss: 0.6576\n",
            "step 3514/5000, loss: 0.7561\n",
            "step 3515/5000, loss: 0.8429\n",
            "step 3516/5000, loss: 0.8366\n",
            "step 3517/5000, loss: 0.5173\n",
            "step 3518/5000, loss: 0.7069\n",
            "step 3519/5000, loss: 0.7064\n",
            "step 3520/5000, loss: 0.6991\n",
            "step 3521/5000, loss: 0.6758\n",
            "step 3522/5000, loss: 0.6366\n",
            "step 3523/5000, loss: 0.6091\n",
            "step 3524/5000, loss: 0.7269\n",
            "step 3525/5000, loss: 0.7440\n",
            "step 3526/5000, loss: 0.7093\n",
            "step 3527/5000, loss: 0.5686\n",
            "step 3528/5000, loss: 0.6790\n",
            "step 3529/5000, loss: 0.5844\n",
            "step 3530/5000, loss: 0.5875\n",
            "step 3531/5000, loss: 0.6781\n",
            "step 3532/5000, loss: 1.4888\n",
            "step 3533/5000, loss: 0.6747\n",
            "step 3534/5000, loss: 0.7323\n",
            "step 3535/5000, loss: 0.7384\n",
            "step 3536/5000, loss: 0.6058\n",
            "step 3537/5000, loss: 1.1994\n",
            "step 3538/5000, loss: 0.6984\n",
            "step 3539/5000, loss: 0.7810\n",
            "step 3540/5000, loss: 0.6867\n",
            "step 3541/5000, loss: 0.6843\n",
            "step 3542/5000, loss: 0.7547\n",
            "step 3543/5000, loss: 0.6490\n",
            "step 3544/5000, loss: 0.7674\n",
            "step 3545/5000, loss: 0.7536\n",
            "step 3546/5000, loss: 0.7460\n",
            "step 3547/5000, loss: 0.5963\n",
            "step 3548/5000, loss: 0.7897\n",
            "step 3549/5000, loss: 0.7153\n",
            "step 3550/5000, loss: 0.5644\n",
            "step 3551/5000, loss: 0.7011\n",
            "step 3552/5000, loss: 0.6955\n",
            "step 3553/5000, loss: 0.6692\n",
            "step 3554/5000, loss: 0.8316\n",
            "step 3555/5000, loss: 0.8454\n",
            "step 3556/5000, loss: 0.7299\n",
            "step 3557/5000, loss: 0.7232\n",
            "step 3558/5000, loss: 0.5669\n",
            "step 3559/5000, loss: 0.6334\n",
            "step 3560/5000, loss: 0.6514\n",
            "step 3561/5000, loss: 0.7704\n",
            "step 3562/5000, loss: 0.6814\n",
            "step 3563/5000, loss: 0.6155\n",
            "step 3564/5000, loss: 0.7526\n",
            "step 3565/5000, loss: 0.6631\n",
            "step 3566/5000, loss: 0.8390\n",
            "step 3567/5000, loss: 0.6542\n",
            "step 3568/5000, loss: 0.7136\n",
            "step 3569/5000, loss: 0.6733\n",
            "step 3570/5000, loss: 0.7222\n",
            "step 3571/5000, loss: 0.6577\n",
            "step 3572/5000, loss: 0.7904\n",
            "step 3573/5000, loss: 0.5998\n",
            "step 3574/5000, loss: 0.7115\n",
            "step 3575/5000, loss: 0.6973\n",
            "step 3576/5000, loss: 0.6697\n",
            "step 3577/5000, loss: 0.6553\n",
            "step 3578/5000, loss: 0.6890\n",
            "step 3579/5000, loss: 0.7491\n",
            "step 3580/5000, loss: 0.6452\n",
            "step 3581/5000, loss: 0.7956\n",
            "step 3582/5000, loss: 0.7205\n",
            "step 3583/5000, loss: 0.7477\n",
            "step 3584/5000, loss: 0.6389\n",
            "step 3585/5000, loss: 0.7388\n",
            "step 3586/5000, loss: 0.6384\n",
            "step 3587/5000, loss: 0.7028\n",
            "step 3588/5000, loss: 0.6403\n",
            "step 3589/5000, loss: 0.6283\n",
            "step 3590/5000, loss: 0.6702\n",
            "step 3591/5000, loss: 0.7142\n",
            "step 3592/5000, loss: 0.7860\n",
            "step 3593/5000, loss: 0.7975\n",
            "step 3594/5000, loss: 0.8901\n",
            "step 3595/5000, loss: 0.7728\n",
            "step 3596/5000, loss: 0.7549\n",
            "step 3597/5000, loss: 0.6886\n",
            "step 3598/5000, loss: 0.7775\n",
            "step 3599/5000, loss: 0.7536\n",
            "step 3600/5000, loss: 0.5925\n",
            "step 3601/5000, loss: 0.6933\n",
            "step 3602/5000, loss: 0.6953\n",
            "step 3603/5000, loss: 0.7440\n",
            "step 3604/5000, loss: 0.6273\n",
            "step 3605/5000, loss: 0.6844\n",
            "step 3606/5000, loss: 0.7675\n",
            "step 3607/5000, loss: 0.6066\n",
            "step 3608/5000, loss: 0.6457\n",
            "step 3609/5000, loss: 0.8660\n",
            "step 3610/5000, loss: 0.5911\n",
            "step 3611/5000, loss: 0.8734\n",
            "step 3612/5000, loss: 0.7434\n",
            "step 3613/5000, loss: 0.6434\n",
            "step 3614/5000, loss: 0.6745\n",
            "step 3615/5000, loss: 0.5874\n",
            "step 3616/5000, loss: 0.8182\n",
            "step 3617/5000, loss: 0.7394\n",
            "step 3618/5000, loss: 0.8625\n",
            "step 3619/5000, loss: 0.7216\n",
            "step 3620/5000, loss: 0.6565\n",
            "step 3621/5000, loss: 0.6702\n",
            "step 3622/5000, loss: 0.6300\n",
            "step 3623/5000, loss: 0.6702\n",
            "step 3624/5000, loss: 0.6616\n",
            "step 3625/5000, loss: 0.6833\n",
            "step 3626/5000, loss: 0.5944\n",
            "step 3627/5000, loss: 0.7017\n",
            "step 3628/5000, loss: 0.5847\n",
            "step 3629/5000, loss: 0.6835\n",
            "step 3630/5000, loss: 0.6438\n",
            "step 3631/5000, loss: 0.6837\n",
            "step 3632/5000, loss: 0.6203\n",
            "step 3633/5000, loss: 0.7153\n",
            "step 3634/5000, loss: 0.7029\n",
            "step 3635/5000, loss: 0.6205\n",
            "step 3636/5000, loss: 0.6764\n",
            "step 3637/5000, loss: 0.8082\n",
            "step 3638/5000, loss: 0.6752\n",
            "step 3639/5000, loss: 0.7466\n",
            "step 3640/5000, loss: 0.6920\n",
            "step 3641/5000, loss: 0.6877\n",
            "step 3642/5000, loss: 0.6553\n",
            "step 3643/5000, loss: 0.6487\n",
            "step 3644/5000, loss: 0.5718\n",
            "step 3645/5000, loss: 0.5905\n",
            "step 3646/5000, loss: 0.8074\n",
            "step 3647/5000, loss: 0.6404\n",
            "step 3648/5000, loss: 0.6436\n",
            "step 3649/5000, loss: 0.8383\n",
            "step 3650/5000, loss: 0.6037\n",
            "step 3651/5000, loss: 0.7005\n",
            "step 3652/5000, loss: 0.7228\n",
            "step 3653/5000, loss: 0.8140\n",
            "step 3654/5000, loss: 0.5746\n",
            "step 3655/5000, loss: 0.6035\n",
            "step 3656/5000, loss: 2.0150\n",
            "step 3657/5000, loss: 0.5957\n",
            "step 3658/5000, loss: 0.6793\n",
            "step 3659/5000, loss: 0.6203\n",
            "step 3660/5000, loss: 0.6162\n",
            "step 3661/5000, loss: 0.7293\n",
            "step 3662/5000, loss: 0.5950\n",
            "step 3663/5000, loss: 0.6479\n",
            "step 3664/5000, loss: 0.6945\n",
            "step 3665/5000, loss: 0.8278\n",
            "step 3666/5000, loss: 0.6607\n",
            "step 3667/5000, loss: 0.6878\n",
            "step 3668/5000, loss: 0.6862\n",
            "step 3669/5000, loss: 0.9223\n",
            "step 3670/5000, loss: 0.7344\n",
            "step 3671/5000, loss: 0.6926\n",
            "step 3672/5000, loss: 0.7690\n",
            "step 3673/5000, loss: 0.6954\n",
            "step 3674/5000, loss: 0.7637\n",
            "step 3675/5000, loss: 0.6432\n",
            "step 3676/5000, loss: 0.5849\n",
            "step 3677/5000, loss: 0.7235\n",
            "step 3678/5000, loss: 0.6284\n",
            "step 3679/5000, loss: 0.7287\n",
            "step 3680/5000, loss: 0.7966\n",
            "step 3681/5000, loss: 0.6507\n",
            "step 3682/5000, loss: 0.7735\n",
            "step 3683/5000, loss: 0.7639\n",
            "step 3684/5000, loss: 0.7057\n",
            "step 3685/5000, loss: 0.7348\n",
            "step 3686/5000, loss: 0.7988\n",
            "step 3687/5000, loss: 0.6429\n",
            "step 3688/5000, loss: 0.8429\n",
            "step 3689/5000, loss: 0.6361\n",
            "step 3690/5000, loss: 0.7166\n",
            "step 3691/5000, loss: 0.7471\n",
            "step 3692/5000, loss: 0.6885\n",
            "step 3693/5000, loss: 0.6649\n",
            "step 3694/5000, loss: 0.8282\n",
            "step 3695/5000, loss: 0.6512\n",
            "step 3696/5000, loss: 0.7370\n",
            "step 3697/5000, loss: 0.7443\n",
            "step 3698/5000, loss: 0.6211\n",
            "step 3699/5000, loss: 0.6351\n",
            "step 3700/5000, loss: 0.6516\n",
            "step 3701/5000, loss: 0.6416\n",
            "step 3702/5000, loss: 0.6954\n",
            "step 3703/5000, loss: 0.6547\n",
            "step 3704/5000, loss: 0.6513\n",
            "step 3705/5000, loss: 1.6767\n",
            "step 3706/5000, loss: 0.7476\n",
            "step 3707/5000, loss: 0.6405\n",
            "step 3708/5000, loss: 0.7053\n",
            "step 3709/5000, loss: 0.6251\n",
            "step 3710/5000, loss: 0.6382\n",
            "step 3711/5000, loss: 0.7154\n",
            "step 3712/5000, loss: 0.7006\n",
            "step 3713/5000, loss: 0.6474\n",
            "step 3714/5000, loss: 0.5656\n",
            "step 3715/5000, loss: 0.6490\n",
            "step 3716/5000, loss: 0.6843\n",
            "step 3717/5000, loss: 0.6311\n",
            "step 3718/5000, loss: 0.7200\n",
            "step 3719/5000, loss: 0.7906\n",
            "step 3720/5000, loss: 0.6887\n",
            "step 3721/5000, loss: 0.6488\n",
            "step 3722/5000, loss: 0.6841\n",
            "step 3723/5000, loss: 0.6948\n",
            "step 3724/5000, loss: 0.7359\n",
            "step 3725/5000, loss: 0.6727\n",
            "step 3726/5000, loss: 0.7119\n",
            "step 3727/5000, loss: 1.0309\n",
            "step 3728/5000, loss: 0.6817\n",
            "step 3729/5000, loss: 0.6107\n",
            "step 3730/5000, loss: 0.6311\n",
            "step 3731/5000, loss: 0.7873\n",
            "step 3732/5000, loss: 0.6130\n",
            "step 3733/5000, loss: 0.8410\n",
            "step 3734/5000, loss: 0.5771\n",
            "step 3735/5000, loss: 0.7611\n",
            "step 3736/5000, loss: 0.7415\n",
            "step 3737/5000, loss: 0.6336\n",
            "step 3738/5000, loss: 0.6246\n",
            "step 3739/5000, loss: 0.6995\n",
            "step 3740/5000, loss: 0.8015\n",
            "step 3741/5000, loss: 0.7250\n",
            "step 3742/5000, loss: 0.6206\n",
            "step 3743/5000, loss: 0.7255\n",
            "step 3744/5000, loss: 0.6464\n",
            "step 3745/5000, loss: 0.8037\n",
            "step 3746/5000, loss: 1.0194\n",
            "step 3747/5000, loss: 0.7959\n",
            "step 3748/5000, loss: 0.7794\n",
            "step 3749/5000, loss: 0.6885\n",
            "step 3750/5000, loss: 0.7492\n",
            "step 3751/5000, loss: 0.6775\n",
            "step 3752/5000, loss: 0.6095\n",
            "step 3753/5000, loss: 0.7145\n",
            "step 3754/5000, loss: 0.6846\n",
            "step 3755/5000, loss: 0.7269\n",
            "step 3756/5000, loss: 0.9269\n",
            "step 3757/5000, loss: 0.6958\n",
            "step 3758/5000, loss: 0.6155\n",
            "step 3759/5000, loss: 0.6857\n",
            "step 3760/5000, loss: 0.6889\n",
            "step 3761/5000, loss: 0.7372\n",
            "step 3762/5000, loss: 0.9260\n",
            "step 3763/5000, loss: 0.7376\n",
            "step 3764/5000, loss: 0.6281\n",
            "step 3765/5000, loss: 0.7107\n",
            "step 3766/5000, loss: 0.8260\n",
            "step 3767/5000, loss: 0.6649\n",
            "step 3768/5000, loss: 0.6423\n",
            "step 3769/5000, loss: 0.6414\n",
            "step 3770/5000, loss: 0.6864\n",
            "step 3771/5000, loss: 0.8005\n",
            "step 3772/5000, loss: 0.5834\n",
            "step 3773/5000, loss: 0.7525\n",
            "step 3774/5000, loss: 0.8577\n",
            "step 3775/5000, loss: 0.8470\n",
            "step 3776/5000, loss: 0.5785\n",
            "step 3777/5000, loss: 0.8331\n",
            "step 3778/5000, loss: 0.7536\n",
            "step 3779/5000, loss: 0.6715\n",
            "step 3780/5000, loss: 1.5865\n",
            "step 3781/5000, loss: 0.6442\n",
            "step 3782/5000, loss: 0.7067\n",
            "step 3783/5000, loss: 0.7966\n",
            "step 3784/5000, loss: 0.8241\n",
            "step 3785/5000, loss: 0.5611\n",
            "step 3786/5000, loss: 0.6593\n",
            "step 3787/5000, loss: 0.6816\n",
            "step 3788/5000, loss: 0.5126\n",
            "step 3789/5000, loss: 0.6573\n",
            "step 3790/5000, loss: 0.5842\n",
            "step 3791/5000, loss: 0.7646\n",
            "step 3792/5000, loss: 0.6772\n",
            "step 3793/5000, loss: 0.6264\n",
            "step 3794/5000, loss: 0.6521\n",
            "step 3795/5000, loss: 0.6155\n",
            "step 3796/5000, loss: 0.7231\n",
            "step 3797/5000, loss: 0.7096\n",
            "step 3798/5000, loss: 0.6787\n",
            "step 3799/5000, loss: 0.8384\n",
            "step 3800/5000, loss: 0.6857\n",
            "step 3801/5000, loss: 0.7227\n",
            "step 3802/5000, loss: 0.7016\n",
            "step 3803/5000, loss: 0.7592\n",
            "step 3804/5000, loss: 0.6669\n",
            "step 3805/5000, loss: 0.7170\n",
            "step 3806/5000, loss: 0.6928\n",
            "step 3807/5000, loss: 0.6983\n",
            "step 3808/5000, loss: 0.6566\n",
            "step 3809/5000, loss: 0.6622\n",
            "step 3810/5000, loss: 0.7006\n",
            "step 3811/5000, loss: 0.6737\n",
            "step 3812/5000, loss: 0.8299\n",
            "step 3813/5000, loss: 0.8073\n",
            "step 3814/5000, loss: 0.6053\n",
            "step 3815/5000, loss: 0.5483\n",
            "step 3816/5000, loss: 0.6368\n",
            "step 3817/5000, loss: 0.7742\n",
            "step 3818/5000, loss: 0.7418\n",
            "step 3819/5000, loss: 0.6250\n",
            "step 3820/5000, loss: 0.6406\n",
            "step 3821/5000, loss: 0.6694\n",
            "step 3822/5000, loss: 0.6455\n",
            "step 3823/5000, loss: 0.6206\n",
            "step 3824/5000, loss: 0.8524\n",
            "step 3825/5000, loss: 0.7953\n",
            "step 3826/5000, loss: 0.6927\n",
            "step 3827/5000, loss: 0.6367\n",
            "step 3828/5000, loss: 0.6589\n",
            "step 3829/5000, loss: 0.6283\n",
            "step 3830/5000, loss: 0.7190\n",
            "step 3831/5000, loss: 0.6466\n",
            "step 3832/5000, loss: 0.7097\n",
            "step 3833/5000, loss: 0.6509\n",
            "step 3834/5000, loss: 0.8009\n",
            "step 3835/5000, loss: 0.6880\n",
            "step 3836/5000, loss: 0.6774\n",
            "step 3837/5000, loss: 0.7846\n",
            "step 3838/5000, loss: 0.6313\n",
            "step 3839/5000, loss: 0.6139\n",
            "step 3840/5000, loss: 0.6277\n",
            "step 3841/5000, loss: 0.8529\n",
            "step 3842/5000, loss: 0.6351\n",
            "step 3843/5000, loss: 0.8386\n",
            "step 3844/5000, loss: 0.6754\n",
            "step 3845/5000, loss: 0.6331\n",
            "step 3846/5000, loss: 0.7730\n",
            "step 3847/5000, loss: 0.6743\n",
            "step 3848/5000, loss: 0.6246\n",
            "step 3849/5000, loss: 0.6138\n",
            "step 3850/5000, loss: 0.5829\n",
            "step 3851/5000, loss: 0.6153\n",
            "step 3852/5000, loss: 0.6791\n",
            "step 3853/5000, loss: 0.5855\n",
            "step 3854/5000, loss: 0.6973\n",
            "step 3855/5000, loss: 0.8460\n",
            "step 3856/5000, loss: 0.8021\n",
            "step 3857/5000, loss: 0.6110\n",
            "step 3858/5000, loss: 0.6435\n",
            "step 3859/5000, loss: 0.6633\n",
            "step 3860/5000, loss: 0.7154\n",
            "step 3861/5000, loss: 0.6179\n",
            "step 3862/5000, loss: 0.7903\n",
            "step 3863/5000, loss: 0.6119\n",
            "step 3864/5000, loss: 0.6321\n",
            "step 3865/5000, loss: 0.6454\n",
            "step 3866/5000, loss: 0.6867\n",
            "step 3867/5000, loss: 0.6929\n",
            "step 3868/5000, loss: 0.7240\n",
            "step 3869/5000, loss: 0.6339\n",
            "step 3870/5000, loss: 0.5570\n",
            "step 3871/5000, loss: 0.6212\n",
            "step 3872/5000, loss: 0.5865\n",
            "step 3873/5000, loss: 0.7839\n",
            "step 3874/5000, loss: 0.7710\n",
            "step 3875/5000, loss: 0.7786\n",
            "step 3876/5000, loss: 0.5777\n",
            "step 3877/5000, loss: 0.7738\n",
            "step 3878/5000, loss: 0.7073\n",
            "step 3879/5000, loss: 0.5960\n",
            "step 3880/5000, loss: 0.6161\n",
            "step 3881/5000, loss: 0.6956\n",
            "step 3882/5000, loss: 0.6496\n",
            "step 3883/5000, loss: 0.5335\n",
            "step 3884/5000, loss: 0.8428\n",
            "step 3885/5000, loss: 0.8581\n",
            "step 3886/5000, loss: 0.6324\n",
            "step 3887/5000, loss: 0.7963\n",
            "step 3888/5000, loss: 0.6685\n",
            "step 3889/5000, loss: 0.6518\n",
            "step 3890/5000, loss: 0.6799\n",
            "step 3891/5000, loss: 0.6202\n",
            "step 3892/5000, loss: 0.5727\n",
            "step 3893/5000, loss: 0.7559\n",
            "step 3894/5000, loss: 0.8124\n",
            "step 3895/5000, loss: 0.6358\n",
            "step 3896/5000, loss: 0.8370\n",
            "step 3897/5000, loss: 0.6102\n",
            "step 3898/5000, loss: 0.7305\n",
            "step 3899/5000, loss: 0.7851\n",
            "step 3900/5000, loss: 0.8038\n",
            "step 3901/5000, loss: 0.7537\n",
            "step 3902/5000, loss: 0.6661\n",
            "step 3903/5000, loss: 0.7399\n",
            "step 3904/5000, loss: 0.6620\n",
            "step 3905/5000, loss: 0.6801\n",
            "step 3906/5000, loss: 0.7129\n",
            "step 3907/5000, loss: 0.5571\n",
            "step 3908/5000, loss: 0.6198\n",
            "step 3909/5000, loss: 0.6629\n",
            "step 3910/5000, loss: 0.5803\n",
            "step 3911/5000, loss: 0.9075\n",
            "step 3912/5000, loss: 0.6932\n",
            "step 3913/5000, loss: 0.8272\n",
            "step 3914/5000, loss: 0.7194\n",
            "step 3915/5000, loss: 0.8840\n",
            "step 3916/5000, loss: 0.6434\n",
            "step 3917/5000, loss: 0.7028\n",
            "step 3918/5000, loss: 0.8262\n",
            "step 3919/5000, loss: 1.0540\n",
            "step 3920/5000, loss: 0.6150\n",
            "step 3921/5000, loss: 0.6939\n",
            "step 3922/5000, loss: 0.6715\n",
            "step 3923/5000, loss: 0.6417\n",
            "step 3924/5000, loss: 0.7627\n",
            "step 3925/5000, loss: 0.6619\n",
            "step 3926/5000, loss: 0.6720\n",
            "step 3927/5000, loss: 0.7111\n",
            "step 3928/5000, loss: 0.7649\n",
            "step 3929/5000, loss: 0.7481\n",
            "step 3930/5000, loss: 0.6229\n",
            "step 3931/5000, loss: 0.6398\n",
            "step 3932/5000, loss: 0.6158\n",
            "step 3933/5000, loss: 0.6455\n",
            "step 3934/5000, loss: 0.8660\n",
            "step 3935/5000, loss: 0.8112\n",
            "step 3936/5000, loss: 0.7190\n",
            "step 3937/5000, loss: 0.6039\n",
            "step 3938/5000, loss: 0.8388\n",
            "step 3939/5000, loss: 0.6719\n",
            "step 3940/5000, loss: 0.6766\n",
            "step 3941/5000, loss: 0.6549\n",
            "step 3942/5000, loss: 0.5639\n",
            "step 3943/5000, loss: 0.7797\n",
            "step 3944/5000, loss: 0.7949\n",
            "step 3945/5000, loss: 0.6872\n",
            "step 3946/5000, loss: 0.6697\n",
            "step 3947/5000, loss: 0.7176\n",
            "step 3948/5000, loss: 0.6706\n",
            "step 3949/5000, loss: 0.6477\n",
            "step 3950/5000, loss: 0.7871\n",
            "step 3951/5000, loss: 0.6931\n",
            "step 3952/5000, loss: 0.6841\n",
            "step 3953/5000, loss: 0.7822\n",
            "step 3954/5000, loss: 0.7282\n",
            "step 3955/5000, loss: 0.6936\n",
            "step 3956/5000, loss: 0.6335\n",
            "step 3957/5000, loss: 0.7290\n",
            "step 3958/5000, loss: 0.6365\n",
            "step 3959/5000, loss: 0.7613\n",
            "step 3960/5000, loss: 0.7093\n",
            "step 3961/5000, loss: 0.9028\n",
            "step 3962/5000, loss: 0.9274\n",
            "step 3963/5000, loss: 0.6721\n",
            "step 3964/5000, loss: 0.6730\n",
            "step 3965/5000, loss: 0.7429\n",
            "step 3966/5000, loss: 0.6133\n",
            "step 3967/5000, loss: 0.5604\n",
            "step 3968/5000, loss: 0.6969\n",
            "step 3969/5000, loss: 0.6726\n",
            "step 3970/5000, loss: 0.6829\n",
            "step 3971/5000, loss: 0.7013\n",
            "step 3972/5000, loss: 0.6486\n",
            "step 3973/5000, loss: 0.7846\n",
            "step 3974/5000, loss: 0.6572\n",
            "step 3975/5000, loss: 0.6695\n",
            "step 3976/5000, loss: 0.7069\n",
            "step 3977/5000, loss: 0.6178\n",
            "step 3978/5000, loss: 0.7752\n",
            "step 3979/5000, loss: 0.5851\n",
            "step 3980/5000, loss: 0.7649\n",
            "step 3981/5000, loss: 0.7116\n",
            "step 3982/5000, loss: 0.7473\n",
            "step 3983/5000, loss: 0.7190\n",
            "step 3984/5000, loss: 0.5804\n",
            "step 3985/5000, loss: 0.6359\n",
            "step 3986/5000, loss: 0.6464\n",
            "step 3987/5000, loss: 0.6487\n",
            "step 3988/5000, loss: 0.7316\n",
            "step 3989/5000, loss: 0.7197\n",
            "step 3990/5000, loss: 0.7970\n",
            "step 3991/5000, loss: 0.7324\n",
            "step 3992/5000, loss: 0.5698\n",
            "step 3993/5000, loss: 0.6365\n",
            "step 3994/5000, loss: 0.7341\n",
            "step 3995/5000, loss: 1.3438\n",
            "step 3996/5000, loss: 0.6475\n",
            "step 3997/5000, loss: 0.7363\n",
            "step 3998/5000, loss: 0.6695\n",
            "step 3999/5000, loss: 0.7846\n",
            "step 4000/5000, loss: 0.7274\n",
            "step 4001/5000, loss: 0.6996\n",
            "step 4002/5000, loss: 0.6469\n",
            "step 4003/5000, loss: 0.7096\n",
            "step 4004/5000, loss: 0.7347\n",
            "step 4005/5000, loss: 0.6157\n",
            "step 4006/5000, loss: 0.6365\n",
            "step 4007/5000, loss: 0.5466\n",
            "step 4008/5000, loss: 0.7542\n",
            "step 4009/5000, loss: 0.6113\n",
            "step 4010/5000, loss: 0.7130\n",
            "step 4011/5000, loss: 0.5782\n",
            "step 4012/5000, loss: 0.6395\n",
            "step 4013/5000, loss: 0.8039\n",
            "step 4014/5000, loss: 0.6620\n",
            "step 4015/5000, loss: 0.6374\n",
            "step 4016/5000, loss: 0.6722\n",
            "step 4017/5000, loss: 0.7575\n",
            "step 4018/5000, loss: 1.3607\n",
            "step 4019/5000, loss: 0.8472\n",
            "step 4020/5000, loss: 0.7791\n",
            "step 4021/5000, loss: 0.6911\n",
            "step 4022/5000, loss: 0.7740\n",
            "step 4023/5000, loss: 0.6074\n",
            "step 4024/5000, loss: 0.7729\n",
            "step 4025/5000, loss: 0.7188\n",
            "step 4026/5000, loss: 0.7746\n",
            "step 4027/5000, loss: 0.8245\n",
            "step 4028/5000, loss: 0.5951\n",
            "step 4029/5000, loss: 0.7044\n",
            "step 4030/5000, loss: 0.6288\n",
            "step 4031/5000, loss: 0.6396\n",
            "step 4032/5000, loss: 0.6637\n",
            "step 4033/5000, loss: 0.6335\n",
            "step 4034/5000, loss: 0.5529\n",
            "step 4035/5000, loss: 0.8107\n",
            "step 4036/5000, loss: 0.6322\n",
            "step 4037/5000, loss: 0.6278\n",
            "step 4038/5000, loss: 0.6522\n",
            "step 4039/5000, loss: 0.6185\n",
            "step 4040/5000, loss: 0.6788\n",
            "step 4041/5000, loss: 0.6430\n",
            "step 4042/5000, loss: 0.8560\n",
            "step 4043/5000, loss: 0.7005\n",
            "step 4044/5000, loss: 0.6117\n",
            "step 4045/5000, loss: 0.7010\n",
            "step 4046/5000, loss: 0.7932\n",
            "step 4047/5000, loss: 0.8116\n",
            "step 4048/5000, loss: 0.7236\n",
            "step 4049/5000, loss: 0.7277\n",
            "step 4050/5000, loss: 0.6822\n",
            "step 4051/5000, loss: 0.6543\n",
            "step 4052/5000, loss: 0.5996\n",
            "step 4053/5000, loss: 0.7677\n",
            "step 4054/5000, loss: 0.7446\n",
            "step 4055/5000, loss: 0.6896\n",
            "step 4056/5000, loss: 0.6859\n",
            "step 4057/5000, loss: 0.6603\n",
            "step 4058/5000, loss: 0.6215\n",
            "step 4059/5000, loss: 0.7275\n",
            "step 4060/5000, loss: 0.7675\n",
            "step 4061/5000, loss: 0.7357\n",
            "step 4062/5000, loss: 0.6385\n",
            "step 4063/5000, loss: 0.7539\n",
            "step 4064/5000, loss: 0.5966\n",
            "step 4065/5000, loss: 0.7092\n",
            "step 4066/5000, loss: 0.6186\n",
            "step 4067/5000, loss: 0.6864\n",
            "step 4068/5000, loss: 1.3129\n",
            "step 4069/5000, loss: 0.7746\n",
            "step 4070/5000, loss: 0.5962\n",
            "step 4071/5000, loss: 0.7021\n",
            "step 4072/5000, loss: 0.6210\n",
            "step 4073/5000, loss: 0.6821\n",
            "step 4074/5000, loss: 0.6064\n",
            "step 4075/5000, loss: 0.6844\n",
            "step 4076/5000, loss: 0.7950\n",
            "step 4077/5000, loss: 0.9995\n",
            "step 4078/5000, loss: 0.6580\n",
            "step 4079/5000, loss: 0.7355\n",
            "step 4080/5000, loss: 0.6931\n",
            "step 4081/5000, loss: 0.6720\n",
            "step 4082/5000, loss: 0.6103\n",
            "step 4083/5000, loss: 0.6899\n",
            "step 4084/5000, loss: 0.6118\n",
            "step 4085/5000, loss: 0.6086\n",
            "step 4086/5000, loss: 0.5658\n",
            "step 4087/5000, loss: 0.7229\n",
            "step 4088/5000, loss: 0.6898\n",
            "step 4089/5000, loss: 0.6443\n",
            "step 4090/5000, loss: 0.6774\n",
            "step 4091/5000, loss: 0.7181\n",
            "step 4092/5000, loss: 0.8556\n",
            "step 4093/5000, loss: 0.7317\n",
            "step 4094/5000, loss: 0.6988\n",
            "step 4095/5000, loss: 0.6344\n",
            "step 4096/5000, loss: 0.7725\n",
            "step 4097/5000, loss: 0.7459\n",
            "step 4098/5000, loss: 0.7116\n",
            "step 4099/5000, loss: 0.5201\n",
            "step 4100/5000, loss: 0.5756\n",
            "step 4101/5000, loss: 0.6948\n",
            "step 4102/5000, loss: 0.7231\n",
            "step 4103/5000, loss: 0.7592\n",
            "step 4104/5000, loss: 0.6025\n",
            "step 4105/5000, loss: 0.5954\n",
            "step 4106/5000, loss: 0.7441\n",
            "step 4107/5000, loss: 0.6516\n",
            "step 4108/5000, loss: 0.7873\n",
            "step 4109/5000, loss: 0.7658\n",
            "step 4110/5000, loss: 0.6636\n",
            "step 4111/5000, loss: 0.9147\n",
            "step 4112/5000, loss: 0.6199\n",
            "step 4113/5000, loss: 0.8882\n",
            "step 4114/5000, loss: 0.7655\n",
            "step 4115/5000, loss: 0.6255\n",
            "step 4116/5000, loss: 0.6789\n",
            "step 4117/5000, loss: 0.8121\n",
            "step 4118/5000, loss: 0.6565\n",
            "step 4119/5000, loss: 0.6969\n",
            "step 4120/5000, loss: 0.7398\n",
            "step 4121/5000, loss: 0.6161\n",
            "step 4122/5000, loss: 0.5802\n",
            "step 4123/5000, loss: 0.7719\n",
            "step 4124/5000, loss: 0.6339\n",
            "step 4125/5000, loss: 0.5741\n",
            "step 4126/5000, loss: 0.8875\n",
            "step 4127/5000, loss: 0.6370\n",
            "step 4128/5000, loss: 0.6694\n",
            "step 4129/5000, loss: 0.7857\n",
            "step 4130/5000, loss: 0.7275\n",
            "step 4131/5000, loss: 0.7203\n",
            "step 4132/5000, loss: 0.5950\n",
            "step 4133/5000, loss: 0.6084\n",
            "step 4134/5000, loss: 0.7204\n",
            "step 4135/5000, loss: 0.6458\n",
            "step 4136/5000, loss: 0.7926\n",
            "step 4137/5000, loss: 0.6614\n",
            "step 4138/5000, loss: 0.7220\n",
            "step 4139/5000, loss: 0.6306\n",
            "step 4140/5000, loss: 0.6416\n",
            "step 4141/5000, loss: 0.6639\n",
            "step 4142/5000, loss: 0.8321\n",
            "step 4143/5000, loss: 0.6238\n",
            "step 4144/5000, loss: 0.6686\n",
            "step 4145/5000, loss: 0.8108\n",
            "step 4146/5000, loss: 0.6425\n",
            "step 4147/5000, loss: 0.6593\n",
            "step 4148/5000, loss: 0.6441\n",
            "step 4149/5000, loss: 0.7179\n",
            "step 4150/5000, loss: 0.7002\n",
            "step 4151/5000, loss: 0.7509\n",
            "step 4152/5000, loss: 0.7427\n",
            "step 4153/5000, loss: 0.5685\n",
            "step 4154/5000, loss: 0.6114\n",
            "step 4155/5000, loss: 1.0204\n",
            "step 4156/5000, loss: 0.6732\n",
            "step 4157/5000, loss: 0.6999\n",
            "step 4158/5000, loss: 0.8290\n",
            "step 4159/5000, loss: 0.6688\n",
            "step 4160/5000, loss: 0.6741\n",
            "step 4161/5000, loss: 0.7217\n",
            "step 4162/5000, loss: 0.7098\n",
            "step 4163/5000, loss: 0.7414\n",
            "step 4164/5000, loss: 0.7160\n",
            "step 4165/5000, loss: 0.7509\n",
            "step 4166/5000, loss: 0.7081\n",
            "step 4167/5000, loss: 0.6855\n",
            "step 4168/5000, loss: 0.7000\n",
            "step 4169/5000, loss: 0.7350\n",
            "step 4170/5000, loss: 0.8189\n",
            "step 4171/5000, loss: 0.7021\n",
            "step 4172/5000, loss: 0.6292\n",
            "step 4173/5000, loss: 0.6410\n",
            "step 4174/5000, loss: 0.6653\n",
            "step 4175/5000, loss: 0.5831\n",
            "step 4176/5000, loss: 0.6642\n",
            "step 4177/5000, loss: 0.7344\n",
            "step 4178/5000, loss: 0.6412\n",
            "step 4179/5000, loss: 0.6179\n",
            "step 4180/5000, loss: 0.6223\n",
            "step 4181/5000, loss: 0.6171\n",
            "step 4182/5000, loss: 0.6722\n",
            "step 4183/5000, loss: 0.7363\n",
            "step 4184/5000, loss: 0.6246\n",
            "step 4185/5000, loss: 0.7060\n",
            "step 4186/5000, loss: 0.6579\n",
            "step 4187/5000, loss: 0.7851\n",
            "step 4188/5000, loss: 0.7862\n",
            "step 4189/5000, loss: 0.6850\n",
            "step 4190/5000, loss: 0.5864\n",
            "step 4191/5000, loss: 0.6562\n",
            "step 4192/5000, loss: 0.6289\n",
            "step 4193/5000, loss: 0.6728\n",
            "step 4194/5000, loss: 0.6281\n",
            "step 4195/5000, loss: 0.6866\n",
            "step 4196/5000, loss: 0.6384\n",
            "step 4197/5000, loss: 0.6378\n",
            "step 4198/5000, loss: 0.7081\n",
            "step 4199/5000, loss: 0.5651\n",
            "step 4200/5000, loss: 0.6817\n",
            "step 4201/5000, loss: 0.6752\n",
            "step 4202/5000, loss: 0.5603\n",
            "step 4203/5000, loss: 0.7093\n",
            "step 4204/5000, loss: 0.6307\n",
            "step 4205/5000, loss: 0.5956\n",
            "step 4206/5000, loss: 0.6620\n",
            "step 4207/5000, loss: 0.6161\n",
            "step 4208/5000, loss: 0.7282\n",
            "step 4209/5000, loss: 0.7644\n",
            "step 4210/5000, loss: 0.7865\n",
            "step 4211/5000, loss: 0.6465\n",
            "step 4212/5000, loss: 0.6466\n",
            "step 4213/5000, loss: 0.6207\n",
            "step 4214/5000, loss: 0.5903\n",
            "step 4215/5000, loss: 0.8041\n",
            "step 4216/5000, loss: 0.7029\n",
            "step 4217/5000, loss: 0.7448\n",
            "step 4218/5000, loss: 0.6188\n",
            "step 4219/5000, loss: 0.5383\n",
            "step 4220/5000, loss: 0.6503\n",
            "step 4221/5000, loss: 0.7665\n",
            "step 4222/5000, loss: 0.7124\n",
            "step 4223/5000, loss: 0.6798\n",
            "step 4224/5000, loss: 0.6503\n",
            "step 4225/5000, loss: 0.6401\n",
            "step 4226/5000, loss: 0.7510\n",
            "step 4227/5000, loss: 0.7132\n",
            "step 4228/5000, loss: 0.7753\n",
            "step 4229/5000, loss: 0.7027\n",
            "step 4230/5000, loss: 0.7197\n",
            "step 4231/5000, loss: 0.6541\n",
            "step 4232/5000, loss: 0.6957\n",
            "step 4233/5000, loss: 0.7196\n",
            "step 4234/5000, loss: 0.6370\n",
            "step 4235/5000, loss: 0.7703\n",
            "step 4236/5000, loss: 0.6439\n",
            "step 4237/5000, loss: 0.6703\n",
            "step 4238/5000, loss: 0.7947\n",
            "step 4239/5000, loss: 0.6222\n",
            "step 4240/5000, loss: 0.7216\n",
            "step 4241/5000, loss: 0.6597\n",
            "step 4242/5000, loss: 0.6835\n",
            "step 4243/5000, loss: 0.6655\n",
            "step 4244/5000, loss: 0.6311\n",
            "step 4245/5000, loss: 0.8876\n",
            "step 4246/5000, loss: 0.6486\n",
            "step 4247/5000, loss: 0.7166\n",
            "step 4248/5000, loss: 0.6115\n",
            "step 4249/5000, loss: 0.6732\n",
            "step 4250/5000, loss: 0.5362\n",
            "step 4251/5000, loss: 0.5323\n",
            "step 4252/5000, loss: 1.0609\n",
            "step 4253/5000, loss: 0.6408\n",
            "step 4254/5000, loss: 0.6420\n",
            "step 4255/5000, loss: 0.5844\n",
            "step 4256/5000, loss: 0.6993\n",
            "step 4257/5000, loss: 0.6435\n",
            "step 4258/5000, loss: 0.7436\n",
            "step 4259/5000, loss: 0.7465\n",
            "step 4260/5000, loss: 0.6725\n",
            "step 4261/5000, loss: 0.6557\n",
            "step 4262/5000, loss: 0.6819\n",
            "step 4263/5000, loss: 0.6480\n",
            "step 4264/5000, loss: 0.6449\n",
            "step 4265/5000, loss: 0.6106\n",
            "step 4266/5000, loss: 0.6537\n",
            "step 4267/5000, loss: 0.6429\n",
            "step 4268/5000, loss: 0.7214\n",
            "step 4269/5000, loss: 0.7487\n",
            "step 4270/5000, loss: 1.6591\n",
            "step 4271/5000, loss: 0.5952\n",
            "step 4272/5000, loss: 0.8220\n",
            "step 4273/5000, loss: 0.8598\n",
            "step 4274/5000, loss: 0.6486\n",
            "step 4275/5000, loss: 0.7178\n",
            "step 4276/5000, loss: 0.7571\n",
            "step 4277/5000, loss: 0.6684\n",
            "step 4278/5000, loss: 0.7362\n",
            "step 4279/5000, loss: 0.6476\n",
            "step 4280/5000, loss: 0.6030\n",
            "step 4281/5000, loss: 0.6596\n",
            "step 4282/5000, loss: 0.6956\n",
            "step 4283/5000, loss: 0.7415\n",
            "step 4284/5000, loss: 0.7158\n",
            "step 4285/5000, loss: 0.7911\n",
            "step 4286/5000, loss: 0.6664\n",
            "step 4287/5000, loss: 0.6249\n",
            "step 4288/5000, loss: 0.6860\n",
            "step 4289/5000, loss: 0.8573\n",
            "step 4290/5000, loss: 0.7470\n",
            "step 4291/5000, loss: 0.7217\n",
            "step 4292/5000, loss: 0.6111\n",
            "step 4293/5000, loss: 0.6609\n",
            "step 4294/5000, loss: 0.7360\n",
            "step 4295/5000, loss: 0.6366\n",
            "step 4296/5000, loss: 0.5446\n",
            "step 4297/5000, loss: 0.5660\n",
            "step 4298/5000, loss: 0.5768\n",
            "step 4299/5000, loss: 0.7455\n",
            "step 4300/5000, loss: 1.5447\n",
            "step 4301/5000, loss: 0.7625\n",
            "step 4302/5000, loss: 0.7224\n",
            "step 4303/5000, loss: 0.8536\n",
            "step 4304/5000, loss: 0.5891\n",
            "step 4305/5000, loss: 0.6881\n",
            "step 4306/5000, loss: 0.8585\n",
            "step 4307/5000, loss: 0.5931\n",
            "step 4308/5000, loss: 0.6635\n",
            "step 4309/5000, loss: 0.6875\n",
            "step 4310/5000, loss: 0.5974\n",
            "step 4311/5000, loss: 0.6042\n",
            "step 4312/5000, loss: 0.6649\n",
            "step 4313/5000, loss: 0.5972\n",
            "step 4314/5000, loss: 0.6760\n",
            "step 4315/5000, loss: 0.6605\n",
            "step 4316/5000, loss: 0.6629\n",
            "step 4317/5000, loss: 0.6853\n",
            "step 4318/5000, loss: 0.7561\n",
            "step 4319/5000, loss: 0.6747\n",
            "step 4320/5000, loss: 0.7596\n",
            "step 4321/5000, loss: 0.5682\n",
            "step 4322/5000, loss: 0.7426\n",
            "step 4323/5000, loss: 0.6303\n",
            "step 4324/5000, loss: 0.6447\n",
            "step 4325/5000, loss: 0.7038\n",
            "step 4326/5000, loss: 0.6698\n",
            "step 4327/5000, loss: 0.6692\n",
            "step 4328/5000, loss: 0.8111\n",
            "step 4329/5000, loss: 0.5993\n",
            "step 4330/5000, loss: 0.6723\n",
            "step 4331/5000, loss: 0.9757\n",
            "step 4332/5000, loss: 0.7876\n",
            "step 4333/5000, loss: 0.6809\n",
            "step 4334/5000, loss: 0.6970\n",
            "step 4335/5000, loss: 0.7005\n",
            "step 4336/5000, loss: 0.8508\n",
            "step 4337/5000, loss: 0.6516\n",
            "step 4338/5000, loss: 1.9083\n",
            "step 4339/5000, loss: 0.7048\n",
            "step 4340/5000, loss: 0.8285\n",
            "step 4341/5000, loss: 0.7660\n",
            "step 4342/5000, loss: 0.7487\n",
            "step 4343/5000, loss: 0.6706\n",
            "step 4344/5000, loss: 0.6232\n",
            "step 4345/5000, loss: 0.5209\n",
            "step 4346/5000, loss: 0.7428\n",
            "step 4347/5000, loss: 0.7756\n",
            "step 4348/5000, loss: 0.6099\n",
            "step 4349/5000, loss: 0.6613\n",
            "step 4350/5000, loss: 0.6945\n",
            "step 4351/5000, loss: 0.7778\n",
            "step 4352/5000, loss: 0.7562\n",
            "step 4353/5000, loss: 0.6897\n",
            "step 4354/5000, loss: 0.5920\n",
            "step 4355/5000, loss: 0.6158\n",
            "step 4356/5000, loss: 0.5053\n",
            "step 4357/5000, loss: 0.8211\n",
            "step 4358/5000, loss: 0.6913\n",
            "step 4359/5000, loss: 0.7754\n",
            "step 4360/5000, loss: 0.6976\n",
            "step 4361/5000, loss: 0.6817\n",
            "step 4362/5000, loss: 0.6298\n",
            "step 4363/5000, loss: 0.6587\n",
            "step 4364/5000, loss: 0.6280\n",
            "step 4365/5000, loss: 0.6201\n",
            "step 4366/5000, loss: 0.5428\n",
            "step 4367/5000, loss: 0.6099\n",
            "step 4368/5000, loss: 0.7703\n",
            "step 4369/5000, loss: 0.6742\n",
            "step 4370/5000, loss: 0.7397\n",
            "step 4371/5000, loss: 0.5754\n",
            "step 4372/5000, loss: 0.5182\n",
            "step 4373/5000, loss: 0.7409\n",
            "step 4374/5000, loss: 0.6722\n",
            "step 4375/5000, loss: 0.7155\n",
            "step 4376/5000, loss: 0.6301\n",
            "step 4377/5000, loss: 0.6732\n",
            "step 4378/5000, loss: 0.7581\n",
            "step 4379/5000, loss: 0.7555\n",
            "step 4380/5000, loss: 0.8236\n",
            "step 4381/5000, loss: 0.7013\n",
            "step 4382/5000, loss: 0.6333\n",
            "step 4383/5000, loss: 0.7513\n",
            "step 4384/5000, loss: 0.6855\n",
            "step 4385/5000, loss: 0.6991\n",
            "step 4386/5000, loss: 0.7726\n",
            "step 4387/5000, loss: 0.7290\n",
            "step 4388/5000, loss: 0.7240\n",
            "step 4389/5000, loss: 0.6898\n",
            "step 4390/5000, loss: 0.6990\n",
            "step 4391/5000, loss: 0.6280\n",
            "step 4392/5000, loss: 0.8321\n",
            "step 4393/5000, loss: 0.5955\n",
            "step 4394/5000, loss: 0.8296\n",
            "step 4395/5000, loss: 0.6207\n",
            "step 4396/5000, loss: 0.5718\n",
            "step 4397/5000, loss: 0.6180\n",
            "step 4398/5000, loss: 0.6055\n",
            "step 4399/5000, loss: 0.7317\n",
            "step 4400/5000, loss: 0.6581\n",
            "step 4401/5000, loss: 0.6615\n",
            "step 4402/5000, loss: 0.7193\n",
            "step 4403/5000, loss: 0.7565\n",
            "step 4404/5000, loss: 0.5898\n",
            "step 4405/5000, loss: 0.5579\n",
            "step 4406/5000, loss: 0.7133\n",
            "step 4407/5000, loss: 0.5461\n",
            "step 4408/5000, loss: 0.7094\n",
            "step 4409/5000, loss: 0.7474\n",
            "step 4410/5000, loss: 1.0001\n",
            "step 4411/5000, loss: 1.0851\n",
            "step 4412/5000, loss: 0.7165\n",
            "step 4413/5000, loss: 0.7816\n",
            "step 4414/5000, loss: 0.6420\n",
            "step 4415/5000, loss: 0.6341\n",
            "step 4416/5000, loss: 0.6623\n",
            "step 4417/5000, loss: 0.5787\n",
            "step 4418/5000, loss: 0.6698\n",
            "step 4419/5000, loss: 0.5528\n",
            "step 4420/5000, loss: 0.6817\n",
            "step 4421/5000, loss: 0.6147\n",
            "step 4422/5000, loss: 0.8328\n",
            "step 4423/5000, loss: 0.6408\n",
            "step 4424/5000, loss: 0.5897\n",
            "step 4425/5000, loss: 0.6568\n",
            "step 4426/5000, loss: 0.6803\n",
            "step 4427/5000, loss: 0.6869\n",
            "step 4428/5000, loss: 0.7158\n",
            "step 4429/5000, loss: 0.5530\n",
            "step 4430/5000, loss: 0.5330\n",
            "step 4431/5000, loss: 0.6373\n",
            "step 4432/5000, loss: 0.7738\n",
            "step 4433/5000, loss: 0.6650\n",
            "step 4434/5000, loss: 0.5730\n",
            "step 4435/5000, loss: 0.5699\n",
            "step 4436/5000, loss: 0.6141\n",
            "step 4437/5000, loss: 0.6577\n",
            "step 4438/5000, loss: 0.6998\n",
            "step 4439/5000, loss: 0.6964\n",
            "step 4440/5000, loss: 0.7856\n",
            "step 4441/5000, loss: 0.7317\n",
            "step 4442/5000, loss: 0.5974\n",
            "step 4443/5000, loss: 0.6167\n",
            "step 4444/5000, loss: 0.6153\n",
            "step 4445/5000, loss: 0.6686\n",
            "step 4446/5000, loss: 0.5953\n",
            "step 4447/5000, loss: 0.7062\n",
            "step 4448/5000, loss: 0.6271\n",
            "step 4449/5000, loss: 0.7681\n",
            "step 4450/5000, loss: 0.5636\n",
            "step 4451/5000, loss: 0.6329\n",
            "step 4452/5000, loss: 0.7037\n",
            "step 4453/5000, loss: 0.7432\n",
            "step 4454/5000, loss: 0.7172\n",
            "step 4455/5000, loss: 0.6117\n",
            "step 4456/5000, loss: 0.7483\n",
            "step 4457/5000, loss: 0.8044\n",
            "step 4458/5000, loss: 1.4668\n",
            "step 4459/5000, loss: 0.6190\n",
            "step 4460/5000, loss: 0.6027\n",
            "step 4461/5000, loss: 0.6238\n",
            "step 4462/5000, loss: 0.7091\n",
            "step 4463/5000, loss: 0.5888\n",
            "step 4464/5000, loss: 0.7243\n",
            "step 4465/5000, loss: 0.7123\n",
            "step 4466/5000, loss: 0.7438\n",
            "step 4467/5000, loss: 0.6885\n",
            "step 4468/5000, loss: 0.6365\n",
            "step 4469/5000, loss: 0.5676\n",
            "step 4470/5000, loss: 0.8726\n",
            "step 4471/5000, loss: 0.7573\n",
            "step 4472/5000, loss: 0.6246\n",
            "step 4473/5000, loss: 0.8685\n",
            "step 4474/5000, loss: 0.8237\n",
            "step 4475/5000, loss: 0.7692\n",
            "step 4476/5000, loss: 0.6074\n",
            "step 4477/5000, loss: 0.6431\n",
            "step 4478/5000, loss: 0.6642\n",
            "step 4479/5000, loss: 0.6876\n",
            "step 4480/5000, loss: 0.7439\n",
            "step 4481/5000, loss: 0.7443\n",
            "step 4482/5000, loss: 0.6710\n",
            "step 4483/5000, loss: 0.5864\n",
            "step 4484/5000, loss: 0.7326\n",
            "step 4485/5000, loss: 0.7640\n",
            "step 4486/5000, loss: 0.7732\n",
            "step 4487/5000, loss: 0.6695\n",
            "step 4488/5000, loss: 0.6790\n",
            "step 4489/5000, loss: 0.6838\n",
            "step 4490/5000, loss: 0.6535\n",
            "step 4491/5000, loss: 0.6930\n",
            "step 4492/5000, loss: 0.7850\n",
            "step 4493/5000, loss: 0.8047\n",
            "step 4494/5000, loss: 0.7918\n",
            "step 4495/5000, loss: 0.6785\n",
            "step 4496/5000, loss: 0.5702\n",
            "step 4497/5000, loss: 0.6955\n",
            "step 4498/5000, loss: 0.6692\n",
            "step 4499/5000, loss: 0.7567\n",
            "step 4500/5000, loss: 0.6012\n",
            "step 4501/5000, loss: 0.6032\n",
            "step 4502/5000, loss: 0.5828\n",
            "step 4503/5000, loss: 0.6996\n",
            "step 4504/5000, loss: 0.6777\n",
            "step 4505/5000, loss: 0.7385\n",
            "step 4506/5000, loss: 0.6678\n",
            "step 4507/5000, loss: 0.6294\n",
            "step 4508/5000, loss: 0.6838\n",
            "step 4509/5000, loss: 0.6243\n",
            "step 4510/5000, loss: 0.8046\n",
            "step 4511/5000, loss: 0.6234\n",
            "step 4512/5000, loss: 0.5971\n",
            "step 4513/5000, loss: 0.7206\n",
            "step 4514/5000, loss: 0.7083\n",
            "step 4515/5000, loss: 0.6811\n",
            "step 4516/5000, loss: 0.6409\n",
            "step 4517/5000, loss: 0.5880\n",
            "step 4518/5000, loss: 0.9589\n",
            "step 4519/5000, loss: 0.7113\n",
            "step 4520/5000, loss: 0.6935\n",
            "step 4521/5000, loss: 0.5952\n",
            "step 4522/5000, loss: 0.6286\n",
            "step 4523/5000, loss: 0.6266\n",
            "step 4524/5000, loss: 0.6936\n",
            "step 4525/5000, loss: 0.6507\n",
            "step 4526/5000, loss: 0.6683\n",
            "step 4527/5000, loss: 0.6204\n",
            "step 4528/5000, loss: 0.6699\n",
            "step 4529/5000, loss: 0.6313\n",
            "step 4530/5000, loss: 0.7441\n",
            "step 4531/5000, loss: 0.7275\n",
            "step 4532/5000, loss: 0.5922\n",
            "step 4533/5000, loss: 0.5488\n",
            "step 4534/5000, loss: 0.6870\n",
            "step 4535/5000, loss: 0.6464\n",
            "step 4536/5000, loss: 1.5778\n",
            "step 4537/5000, loss: 0.5947\n",
            "step 4538/5000, loss: 0.7078\n",
            "step 4539/5000, loss: 0.5835\n",
            "step 4540/5000, loss: 0.7414\n",
            "step 4541/5000, loss: 0.5816\n",
            "step 4542/5000, loss: 0.6507\n",
            "step 4543/5000, loss: 0.6604\n",
            "step 4544/5000, loss: 0.7132\n",
            "step 4545/5000, loss: 0.9850\n",
            "step 4546/5000, loss: 0.6539\n",
            "step 4547/5000, loss: 0.6495\n",
            "step 4548/5000, loss: 0.7407\n",
            "step 4549/5000, loss: 1.0401\n",
            "step 4550/5000, loss: 0.6283\n",
            "step 4551/5000, loss: 0.7524\n",
            "step 4552/5000, loss: 0.5743\n",
            "step 4553/5000, loss: 0.7492\n",
            "step 4554/5000, loss: 0.6855\n",
            "step 4555/5000, loss: 0.6528\n",
            "step 4556/5000, loss: 0.6417\n",
            "step 4557/5000, loss: 0.5790\n",
            "step 4558/5000, loss: 0.6834\n",
            "step 4559/5000, loss: 0.7400\n",
            "step 4560/5000, loss: 0.6807\n",
            "step 4561/5000, loss: 0.6780\n",
            "step 4562/5000, loss: 0.6357\n",
            "step 4563/5000, loss: 0.7074\n",
            "step 4564/5000, loss: 0.6467\n",
            "step 4565/5000, loss: 0.8803\n",
            "step 4566/5000, loss: 0.7613\n",
            "step 4567/5000, loss: 0.5739\n",
            "step 4568/5000, loss: 0.6756\n",
            "step 4569/5000, loss: 0.6467\n",
            "step 4570/5000, loss: 0.6652\n",
            "step 4571/5000, loss: 0.6124\n",
            "step 4572/5000, loss: 0.7350\n",
            "step 4573/5000, loss: 0.6367\n",
            "step 4574/5000, loss: 0.6361\n",
            "step 4575/5000, loss: 0.5823\n",
            "step 4576/5000, loss: 0.5952\n",
            "step 4577/5000, loss: 0.5727\n",
            "step 4578/5000, loss: 0.5922\n",
            "step 4579/5000, loss: 0.6792\n",
            "step 4580/5000, loss: 0.6598\n",
            "step 4581/5000, loss: 0.8178\n",
            "step 4582/5000, loss: 0.7237\n",
            "step 4583/5000, loss: 0.5712\n",
            "step 4584/5000, loss: 0.7044\n",
            "step 4585/5000, loss: 0.6977\n",
            "step 4586/5000, loss: 0.7517\n",
            "step 4587/5000, loss: 0.6772\n",
            "step 4588/5000, loss: 0.5610\n",
            "step 4589/5000, loss: 0.7230\n",
            "step 4590/5000, loss: 0.8920\n",
            "step 4591/5000, loss: 0.6520\n",
            "step 4592/5000, loss: 0.8578\n",
            "step 4593/5000, loss: 0.6539\n",
            "step 4594/5000, loss: 0.7415\n",
            "step 4595/5000, loss: 0.5524\n",
            "step 4596/5000, loss: 0.5937\n",
            "step 4597/5000, loss: 0.6988\n",
            "step 4598/5000, loss: 0.6595\n",
            "step 4599/5000, loss: 0.8584\n",
            "step 4600/5000, loss: 0.7023\n",
            "step 4601/5000, loss: 0.7490\n",
            "step 4602/5000, loss: 0.8483\n",
            "step 4603/5000, loss: 0.6174\n",
            "step 4604/5000, loss: 0.7577\n",
            "step 4605/5000, loss: 0.5687\n",
            "step 4606/5000, loss: 0.8077\n",
            "step 4607/5000, loss: 0.6978\n",
            "step 4608/5000, loss: 0.7123\n",
            "step 4609/5000, loss: 0.6502\n",
            "step 4610/5000, loss: 0.7223\n",
            "step 4611/5000, loss: 0.6194\n",
            "step 4612/5000, loss: 0.8088\n",
            "step 4613/5000, loss: 0.7550\n",
            "step 4614/5000, loss: 0.6018\n",
            "step 4615/5000, loss: 0.6023\n",
            "step 4616/5000, loss: 0.5706\n",
            "step 4617/5000, loss: 0.6794\n",
            "step 4618/5000, loss: 1.0616\n",
            "step 4619/5000, loss: 0.6138\n",
            "step 4620/5000, loss: 0.6304\n",
            "step 4621/5000, loss: 0.6188\n",
            "step 4622/5000, loss: 0.6049\n",
            "step 4623/5000, loss: 0.7441\n",
            "step 4624/5000, loss: 0.6452\n",
            "step 4625/5000, loss: 0.6574\n",
            "step 4626/5000, loss: 0.9263\n",
            "step 4627/5000, loss: 0.6605\n",
            "step 4628/5000, loss: 0.7544\n",
            "step 4629/5000, loss: 0.6134\n",
            "step 4630/5000, loss: 0.6445\n",
            "step 4631/5000, loss: 0.6645\n",
            "step 4632/5000, loss: 0.6835\n",
            "step 4633/5000, loss: 0.6343\n",
            "step 4634/5000, loss: 0.6719\n",
            "step 4635/5000, loss: 0.5932\n",
            "step 4636/5000, loss: 0.6245\n",
            "step 4637/5000, loss: 0.6047\n",
            "step 4638/5000, loss: 0.6247\n",
            "step 4639/5000, loss: 0.8278\n",
            "step 4640/5000, loss: 0.8095\n",
            "step 4641/5000, loss: 0.6228\n",
            "step 4642/5000, loss: 0.6806\n",
            "step 4643/5000, loss: 0.7256\n",
            "step 4644/5000, loss: 0.6188\n",
            "step 4645/5000, loss: 0.6557\n",
            "step 4646/5000, loss: 0.7082\n",
            "step 4647/5000, loss: 0.7872\n",
            "step 4648/5000, loss: 0.6930\n",
            "step 4649/5000, loss: 0.6477\n",
            "step 4650/5000, loss: 0.7260\n",
            "step 4651/5000, loss: 0.7216\n",
            "step 4652/5000, loss: 0.6207\n",
            "step 4653/5000, loss: 0.6854\n",
            "step 4654/5000, loss: 0.7524\n",
            "step 4655/5000, loss: 0.6965\n",
            "step 4656/5000, loss: 0.6452\n",
            "step 4657/5000, loss: 0.7598\n",
            "step 4658/5000, loss: 1.3378\n",
            "step 4659/5000, loss: 0.6929\n",
            "step 4660/5000, loss: 0.6662\n",
            "step 4661/5000, loss: 0.7112\n",
            "step 4662/5000, loss: 0.5898\n",
            "step 4663/5000, loss: 0.5679\n",
            "step 4664/5000, loss: 0.5666\n",
            "step 4665/5000, loss: 0.5423\n",
            "step 4666/5000, loss: 0.7443\n",
            "step 4667/5000, loss: 0.5922\n",
            "step 4668/5000, loss: 0.6799\n",
            "step 4669/5000, loss: 0.6093\n",
            "step 4670/5000, loss: 0.7275\n",
            "step 4671/5000, loss: 0.6521\n",
            "step 4672/5000, loss: 0.8556\n",
            "step 4673/5000, loss: 0.6150\n",
            "step 4674/5000, loss: 0.6316\n",
            "step 4675/5000, loss: 0.6693\n",
            "step 4676/5000, loss: 0.8353\n",
            "step 4677/5000, loss: 0.6626\n",
            "step 4678/5000, loss: 0.6433\n",
            "step 4679/5000, loss: 0.6590\n",
            "step 4680/5000, loss: 0.7599\n",
            "step 4681/5000, loss: 0.7400\n",
            "step 4682/5000, loss: 0.6018\n",
            "step 4683/5000, loss: 0.6938\n",
            "step 4684/5000, loss: 0.6917\n",
            "step 4685/5000, loss: 0.7619\n",
            "step 4686/5000, loss: 0.7160\n",
            "step 4687/5000, loss: 0.6823\n",
            "step 4688/5000, loss: 0.6373\n",
            "step 4689/5000, loss: 0.7722\n",
            "step 4690/5000, loss: 0.6481\n",
            "step 4691/5000, loss: 0.7490\n",
            "step 4692/5000, loss: 0.7570\n",
            "step 4693/5000, loss: 0.7197\n",
            "step 4694/5000, loss: 0.7336\n",
            "step 4695/5000, loss: 0.6305\n",
            "step 4696/5000, loss: 0.7260\n",
            "step 4697/5000, loss: 0.5647\n",
            "step 4698/5000, loss: 0.7702\n",
            "step 4699/5000, loss: 0.7434\n",
            "step 4700/5000, loss: 0.6448\n",
            "step 4701/5000, loss: 0.5692\n",
            "step 4702/5000, loss: 0.6880\n",
            "step 4703/5000, loss: 0.8233\n",
            "step 4704/5000, loss: 0.7069\n",
            "step 4705/5000, loss: 0.6710\n",
            "step 4706/5000, loss: 0.6331\n",
            "step 4707/5000, loss: 0.8826\n",
            "step 4708/5000, loss: 0.5496\n",
            "step 4709/5000, loss: 0.7849\n",
            "step 4710/5000, loss: 0.7939\n",
            "step 4711/5000, loss: 0.7583\n",
            "step 4712/5000, loss: 0.7522\n",
            "step 4713/5000, loss: 0.5486\n",
            "step 4714/5000, loss: 0.7155\n",
            "step 4715/5000, loss: 0.7240\n",
            "step 4716/5000, loss: 0.7856\n",
            "step 4717/5000, loss: 0.5949\n",
            "step 4718/5000, loss: 0.8020\n",
            "step 4719/5000, loss: 0.7239\n",
            "step 4720/5000, loss: 0.6236\n",
            "step 4721/5000, loss: 0.6160\n",
            "step 4722/5000, loss: 0.6683\n",
            "step 4723/5000, loss: 0.6537\n",
            "step 4724/5000, loss: 0.5909\n",
            "step 4725/5000, loss: 0.7219\n",
            "step 4726/5000, loss: 1.1612\n",
            "step 4727/5000, loss: 0.8831\n",
            "step 4728/5000, loss: 0.6800\n",
            "step 4729/5000, loss: 0.6665\n",
            "step 4730/5000, loss: 0.6167\n",
            "step 4731/5000, loss: 0.6446\n",
            "step 4732/5000, loss: 0.8717\n",
            "step 4733/5000, loss: 0.6816\n",
            "step 4734/5000, loss: 0.6117\n",
            "step 4735/5000, loss: 0.6497\n",
            "step 4736/5000, loss: 0.6521\n",
            "step 4737/5000, loss: 0.6286\n",
            "step 4738/5000, loss: 0.6656\n",
            "step 4739/5000, loss: 0.6248\n",
            "step 4740/5000, loss: 0.6557\n",
            "step 4741/5000, loss: 0.5618\n",
            "step 4742/5000, loss: 0.5954\n",
            "step 4743/5000, loss: 0.7184\n",
            "step 4744/5000, loss: 0.7356\n",
            "step 4745/5000, loss: 0.6929\n",
            "step 4746/5000, loss: 0.7589\n",
            "step 4747/5000, loss: 0.6880\n",
            "step 4748/5000, loss: 0.6826\n",
            "step 4749/5000, loss: 0.7146\n",
            "step 4750/5000, loss: 0.8307\n",
            "step 4751/5000, loss: 0.6763\n",
            "step 4752/5000, loss: 0.7508\n",
            "step 4753/5000, loss: 0.6095\n",
            "step 4754/5000, loss: 0.7199\n",
            "step 4755/5000, loss: 1.9919\n",
            "step 4756/5000, loss: 0.6335\n",
            "step 4757/5000, loss: 0.7931\n",
            "step 4758/5000, loss: 0.6399\n",
            "step 4759/5000, loss: 0.9279\n",
            "step 4760/5000, loss: 0.6701\n",
            "step 4761/5000, loss: 0.6124\n",
            "step 4762/5000, loss: 0.6170\n",
            "step 4763/5000, loss: 0.6434\n",
            "step 4764/5000, loss: 0.6954\n",
            "step 4765/5000, loss: 0.6467\n",
            "step 4766/5000, loss: 0.6218\n",
            "step 4767/5000, loss: 0.8490\n",
            "step 4768/5000, loss: 0.6764\n",
            "step 4769/5000, loss: 0.6896\n",
            "step 4770/5000, loss: 0.7954\n",
            "step 4771/5000, loss: 1.4891\n",
            "step 4772/5000, loss: 0.6756\n",
            "step 4773/5000, loss: 0.7854\n",
            "step 4774/5000, loss: 0.6159\n",
            "step 4775/5000, loss: 0.9263\n",
            "step 4776/5000, loss: 0.6177\n",
            "step 4777/5000, loss: 0.5812\n",
            "step 4778/5000, loss: 0.6781\n",
            "step 4779/5000, loss: 0.6803\n",
            "step 4780/5000, loss: 0.5900\n",
            "step 4781/5000, loss: 0.5920\n",
            "step 4782/5000, loss: 0.7629\n",
            "step 4783/5000, loss: 0.6752\n",
            "step 4784/5000, loss: 0.6395\n",
            "step 4785/5000, loss: 0.6004\n",
            "step 4786/5000, loss: 0.6031\n",
            "step 4787/5000, loss: 0.6043\n",
            "step 4788/5000, loss: 0.5255\n",
            "step 4789/5000, loss: 0.7864\n",
            "step 4790/5000, loss: 0.6879\n",
            "step 4791/5000, loss: 0.6646\n",
            "step 4792/5000, loss: 0.7031\n",
            "step 4793/5000, loss: 0.6845\n",
            "step 4794/5000, loss: 0.7161\n",
            "step 4795/5000, loss: 0.7213\n",
            "step 4796/5000, loss: 0.6981\n",
            "step 4797/5000, loss: 0.6936\n",
            "step 4798/5000, loss: 0.6885\n",
            "step 4799/5000, loss: 0.6825\n",
            "step 4800/5000, loss: 0.7558\n",
            "step 4801/5000, loss: 0.8464\n",
            "step 4802/5000, loss: 0.7068\n",
            "step 4803/5000, loss: 0.5770\n",
            "step 4804/5000, loss: 0.6603\n",
            "step 4805/5000, loss: 0.6341\n",
            "step 4806/5000, loss: 0.6771\n",
            "step 4807/5000, loss: 0.5907\n",
            "step 4808/5000, loss: 0.6858\n",
            "step 4809/5000, loss: 0.8585\n",
            "step 4810/5000, loss: 0.6678\n",
            "step 4811/5000, loss: 0.6096\n",
            "step 4812/5000, loss: 0.7119\n",
            "step 4813/5000, loss: 0.7213\n",
            "step 4814/5000, loss: 0.6769\n",
            "step 4815/5000, loss: 0.7130\n",
            "step 4816/5000, loss: 0.6495\n",
            "step 4817/5000, loss: 0.6830\n",
            "step 4818/5000, loss: 0.8107\n",
            "step 4819/5000, loss: 0.6680\n",
            "step 4820/5000, loss: 0.6409\n",
            "step 4821/5000, loss: 0.7233\n",
            "step 4822/5000, loss: 0.7799\n",
            "step 4823/5000, loss: 0.6677\n",
            "step 4824/5000, loss: 0.5796\n",
            "step 4825/5000, loss: 0.6543\n",
            "step 4826/5000, loss: 0.7422\n",
            "step 4827/5000, loss: 0.8845\n",
            "step 4828/5000, loss: 0.6906\n",
            "step 4829/5000, loss: 0.6038\n",
            "step 4830/5000, loss: 0.6426\n",
            "step 4831/5000, loss: 0.7101\n",
            "step 4832/5000, loss: 0.7152\n",
            "step 4833/5000, loss: 0.7263\n",
            "step 4834/5000, loss: 0.6247\n",
            "step 4835/5000, loss: 0.5964\n",
            "step 4836/5000, loss: 0.6536\n",
            "step 4837/5000, loss: 0.6244\n",
            "step 4838/5000, loss: 0.6817\n",
            "step 4839/5000, loss: 0.5790\n",
            "step 4840/5000, loss: 0.7832\n",
            "step 4841/5000, loss: 0.5982\n",
            "step 4842/5000, loss: 0.7883\n",
            "step 4843/5000, loss: 0.4999\n",
            "step 4844/5000, loss: 0.7461\n",
            "step 4845/5000, loss: 0.6240\n",
            "step 4846/5000, loss: 0.6703\n",
            "step 4847/5000, loss: 0.7317\n",
            "step 4848/5000, loss: 0.5645\n",
            "step 4849/5000, loss: 0.6190\n",
            "step 4850/5000, loss: 0.6864\n",
            "step 4851/5000, loss: 0.6854\n",
            "step 4852/5000, loss: 0.5736\n",
            "step 4853/5000, loss: 0.6840\n",
            "step 4854/5000, loss: 0.7076\n",
            "step 4855/5000, loss: 0.8067\n",
            "step 4856/5000, loss: 0.6351\n",
            "step 4857/5000, loss: 0.6353\n",
            "step 4858/5000, loss: 0.6724\n",
            "step 4859/5000, loss: 0.6550\n",
            "step 4860/5000, loss: 0.7091\n",
            "step 4861/5000, loss: 0.7338\n",
            "step 4862/5000, loss: 0.6567\n",
            "step 4863/5000, loss: 0.7638\n",
            "step 4864/5000, loss: 0.6177\n",
            "step 4865/5000, loss: 0.7225\n",
            "step 4866/5000, loss: 0.6436\n",
            "step 4867/5000, loss: 0.7889\n",
            "step 4868/5000, loss: 0.6470\n",
            "step 4869/5000, loss: 0.7044\n",
            "step 4870/5000, loss: 0.7987\n",
            "step 4871/5000, loss: 0.7117\n",
            "step 4872/5000, loss: 0.5877\n",
            "step 4873/5000, loss: 0.6302\n",
            "step 4874/5000, loss: 0.7147\n",
            "step 4875/5000, loss: 0.6940\n",
            "step 4876/5000, loss: 0.6007\n",
            "step 4877/5000, loss: 0.8002\n",
            "step 4878/5000, loss: 0.6656\n",
            "step 4879/5000, loss: 0.6177\n",
            "step 4880/5000, loss: 0.6003\n",
            "step 4881/5000, loss: 0.8548\n",
            "step 4882/5000, loss: 0.8507\n",
            "step 4883/5000, loss: 0.6247\n",
            "step 4884/5000, loss: 0.5985\n",
            "step 4885/5000, loss: 0.6838\n",
            "step 4886/5000, loss: 0.8191\n",
            "step 4887/5000, loss: 0.6281\n",
            "step 4888/5000, loss: 0.7334\n",
            "step 4889/5000, loss: 0.6359\n",
            "step 4890/5000, loss: 0.8382\n",
            "step 4891/5000, loss: 0.6761\n",
            "step 4892/5000, loss: 0.5952\n",
            "step 4893/5000, loss: 0.6932\n",
            "step 4894/5000, loss: 0.6345\n",
            "step 4895/5000, loss: 0.9588\n",
            "step 4896/5000, loss: 0.5931\n",
            "step 4897/5000, loss: 0.7634\n",
            "step 4898/5000, loss: 0.6365\n",
            "step 4899/5000, loss: 0.6532\n",
            "step 4900/5000, loss: 0.6260\n",
            "step 4901/5000, loss: 0.6286\n",
            "step 4902/5000, loss: 0.6509\n",
            "step 4903/5000, loss: 0.5369\n",
            "step 4904/5000, loss: 0.6710\n",
            "step 4905/5000, loss: 0.5945\n",
            "step 4906/5000, loss: 0.6493\n",
            "step 4907/5000, loss: 0.6991\n",
            "step 4908/5000, loss: 0.5552\n",
            "step 4909/5000, loss: 0.6446\n",
            "step 4910/5000, loss: 0.6568\n",
            "step 4911/5000, loss: 0.7769\n",
            "step 4912/5000, loss: 0.6648\n",
            "step 4913/5000, loss: 0.7634\n",
            "step 4914/5000, loss: 0.5961\n",
            "step 4915/5000, loss: 0.7547\n",
            "step 4916/5000, loss: 0.6515\n",
            "step 4917/5000, loss: 0.6545\n",
            "step 4918/5000, loss: 0.6023\n",
            "step 4919/5000, loss: 0.5889\n",
            "step 4920/5000, loss: 0.7314\n",
            "step 4921/5000, loss: 0.7877\n",
            "step 4922/5000, loss: 0.7347\n",
            "step 4923/5000, loss: 0.7175\n",
            "step 4924/5000, loss: 0.7046\n",
            "step 4925/5000, loss: 0.6176\n",
            "step 4926/5000, loss: 0.8099\n",
            "step 4927/5000, loss: 0.6337\n",
            "step 4928/5000, loss: 0.8001\n",
            "step 4929/5000, loss: 0.6676\n",
            "step 4930/5000, loss: 0.6859\n",
            "step 4931/5000, loss: 0.7845\n",
            "step 4932/5000, loss: 0.6536\n",
            "step 4933/5000, loss: 0.5732\n",
            "step 4934/5000, loss: 0.5585\n",
            "step 4935/5000, loss: 0.7804\n",
            "step 4936/5000, loss: 0.6002\n",
            "step 4937/5000, loss: 0.6772\n",
            "step 4938/5000, loss: 0.6177\n",
            "step 4939/5000, loss: 0.6549\n",
            "step 4940/5000, loss: 0.6034\n",
            "step 4941/5000, loss: 0.5243\n",
            "step 4942/5000, loss: 0.6874\n",
            "step 4943/5000, loss: 0.6141\n",
            "step 4944/5000, loss: 0.6327\n",
            "step 4945/5000, loss: 0.5382\n",
            "step 4946/5000, loss: 0.5727\n",
            "step 4947/5000, loss: 0.6224\n",
            "step 4948/5000, loss: 0.6490\n",
            "step 4949/5000, loss: 0.7360\n",
            "step 4950/5000, loss: 0.7010\n",
            "step 4951/5000, loss: 0.5954\n",
            "step 4952/5000, loss: 0.6160\n",
            "step 4953/5000, loss: 0.6544\n",
            "step 4954/5000, loss: 0.6942\n",
            "step 4955/5000, loss: 0.6948\n",
            "step 4956/5000, loss: 0.5759\n",
            "step 4957/5000, loss: 0.7012\n",
            "step 4958/5000, loss: 0.6569\n",
            "step 4959/5000, loss: 0.7022\n",
            "step 4960/5000, loss: 0.5846\n",
            "step 4961/5000, loss: 0.6556\n",
            "step 4962/5000, loss: 0.6329\n",
            "step 4963/5000, loss: 0.6628\n",
            "step 4964/5000, loss: 0.6301\n",
            "step 4965/5000, loss: 0.6368\n",
            "step 4966/5000, loss: 0.6369\n",
            "step 4967/5000, loss: 0.6632\n",
            "step 4968/5000, loss: 0.7134\n",
            "step 4969/5000, loss: 0.5949\n",
            "step 4970/5000, loss: 0.6165\n",
            "step 4971/5000, loss: 0.7466\n",
            "step 4972/5000, loss: 0.8042\n",
            "step 4973/5000, loss: 0.7602\n",
            "step 4974/5000, loss: 0.7229\n",
            "step 4975/5000, loss: 0.6402\n",
            "step 4976/5000, loss: 0.7238\n",
            "step 4977/5000, loss: 0.6637\n",
            "step 4978/5000, loss: 0.6544\n",
            "step 4979/5000, loss: 0.5783\n",
            "step 4980/5000, loss: 0.6211\n",
            "step 4981/5000, loss: 0.6460\n",
            "step 4982/5000, loss: 0.7342\n",
            "step 4983/5000, loss: 0.6563\n",
            "step 4984/5000, loss: 0.7122\n",
            "step 4985/5000, loss: 0.7320\n",
            "step 4986/5000, loss: 0.6831\n",
            "step 4987/5000, loss: 0.7786\n",
            "step 4988/5000, loss: 0.6382\n",
            "step 4989/5000, loss: 0.8310\n",
            "step 4990/5000, loss: 0.7142\n",
            "step 4991/5000, loss: 0.8255\n",
            "step 4992/5000, loss: 0.6717\n",
            "step 4993/5000, loss: 0.7174\n",
            "step 4994/5000, loss: 0.6088\n",
            "step 4995/5000, loss: 0.7123\n",
            "step 4996/5000, loss: 0.6147\n",
            "step 4997/5000, loss: 0.7083\n",
            "step 4998/5000, loss: 0.6973\n",
            "step 4999/5000, loss: 0.6355\n",
            "step 5000/5000, loss: 0.6150\n",
            "Finished training!\n",
            "Saved plot to distillation_loss.png\n",
            "Saved distilled student model to ./distilled-student-gpt2\n",
            "\n",
            "-- Sample Generations from Distilled Student --\n",
            "Prompt: The curious scientist decided to test a new hypothesis:\n",
            "Completion: The curious scientist decided to test a new hypothesis: what happens if you add enough calcium and a water soluble sugar to one of the four essential amino acids in your animal's muscle?\n",
            "\n",
            "Scientists studied what happens when a compound takes an amino acid and\n",
            "\n",
            "Prompt: Once upon a time in a land far, far away,\n",
            "Completion: Once upon a time in a land far, far away, some of the inhabitants of the Earth's surface, where all mankind, and most of their creatures, are scattered, are slowly drifting back to their natural place. As the air becomes thick and\n",
            "\n",
            "Prompt: In a shocking turn of events today,\n",
            "Completion: In a shocking turn of events today, the former Prime Minister and former head of the UK's intelligence service, Sir John Kerr, had been accused of using \"excessive political influence\" to gain access to his confidential information and expose \"systemically unacceptable\n",
            "\n"
          ]
        },
        {
          "data": {
            "image/png": "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",
            "text/plain": [
              "<Figure size 800x600 with 1 Axes>"
            ]
          },
          "metadata": {},
          "output_type": "display_data"
        }
      ],
      "source": [
        "import torch\n",
        "import torch.nn.functional as F\n",
        "from torch.utils.data import DataLoader\n",
        "import matplotlib.pyplot as plt\n",
        "from datasets import load_dataset\n",
        "\n",
        "from transformers import (\n",
        "    GPT2LMHeadModel,\n",
        "    GPT2Tokenizer,\n",
        "    AdamW,\n",
        "    get_linear_schedule_with_warmup,\n",
        ")\n",
        "\n",
        "# -----------------------------\n",
        "# Hyperparameters\n",
        "# -----------------------------\n",
        "MODEL_NAME_TEACHER = \"gpt2-medium\"  # ~345M params\n",
        "MODEL_NAME_STUDENT = \"gpt2\"         # ~124M params\n",
        "BATCH_SIZE = 2                      # For GPU-poor environments; increase if possible\n",
        "MAX_STEPS = 5000                    # Number of training steps\n",
        "WARMUP_STEPS = 20\n",
        "LR = 5e-5                           # Reduced learning rate for stability\n",
        "TEMPERATURE = 2.0                   # Temperature for softening distributions\n",
        "MAX_SEQ_LEN = 128                   # Sequence length\n",
        "LOG_EVERY = 1                       # Print loss every X steps\n",
        "GRAD_CLIP = 1.0                     # Gradient clipping norm\n",
        "DEVICE = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n",
        "\n",
        "# -----------------------------------------------------\n",
        "# Load Teacher & Student Models + Tokenizer\n",
        "# -----------------------------------------------------\n",
        "print(\"Loading teacher and student models...\")\n",
        "teacher_model = GPT2LMHeadModel.from_pretrained(MODEL_NAME_TEACHER).to(DEVICE)\n",
        "teacher_model.eval()  # Teacher is fixed/frozen\n",
        "\n",
        "student_model = GPT2LMHeadModel.from_pretrained(MODEL_NAME_STUDENT).to(DEVICE)\n",
        "student_model.train()  # We'll train the student\n",
        "\n",
        "tokenizer = GPT2Tokenizer.from_pretrained(MODEL_NAME_STUDENT)\n",
        "# GPT-2 does not have a pad token by default; set pad_token to EOS\n",
        "tokenizer.pad_token = tokenizer.eos_token\n",
        "\n",
        "# -----------------------------------------------------\n",
        "# Load a Small Chunk of OpenWebText\n",
        "# -----------------------------------------------------\n",
        "print(\"Loading a small subset of OpenWebText dataset...\")\n",
        "owt = load_dataset(\"openwebtext\", split=\"train[:1%]\")  # ~a few thousand samples\n",
        "owt = owt.select(range(min(len(owt), 1000)))          # Subset (e.g., 1000 samples)\n",
        "\n",
        "# -----------------------------------------------------\n",
        "# Preprocess / Tokenize\n",
        "# -----------------------------------------------------\n",
        "def tokenize_function(examples):\n",
        "    return tokenizer(\n",
        "        examples[\"text\"],\n",
        "        truncation=True,\n",
        "        max_length=MAX_SEQ_LEN,\n",
        "        padding=\"max_length\",\n",
        "    )\n",
        "\n",
        "print(\"Tokenizing dataset...\")\n",
        "owt_tokenized = owt.map(tokenize_function, batched=True, num_proc=1)\n",
        "owt_tokenized.set_format(type=\"torch\", columns=[\"input_ids\", \"attention_mask\"])\n",
        "\n",
        "# Split into train/eval\n",
        "train_size = int(0.9 * len(owt_tokenized))\n",
        "train_dataset = owt_tokenized.select(range(train_size))\n",
        "eval_dataset = owt_tokenized.select(range(train_size, len(owt_tokenized)))\n",
        "\n",
        "train_loader = DataLoader(train_dataset, batch_size=BATCH_SIZE, shuffle=True)\n",
        "\n",
        "# -----------------------------------------------------\n",
        "# Define the Distillation Loss (Reverse KL)\n",
        "# -----------------------------------------------------\n",
        "def distillation_loss(student_logits, teacher_logits, attention_mask, temperature=TEMPERATURE):\n",
        "    \"\"\"\n",
        "    Distillation loss based on KL(Teacher || Student).\n",
        "    We apply temperature scaling to soften the distributions.\n",
        "    \"\"\"\n",
        "    # Scale logits by temperature\n",
        "    s_logits = student_logits / temperature\n",
        "    t_logits = teacher_logits / temperature\n",
        "\n",
        "    # Compute log probabilities for student and probabilities for teacher\n",
        "    s_log_probs = F.log_softmax(s_logits, dim=-1)   # (B, L, V)\n",
        "    t_probs = F.softmax(t_logits, dim=-1)          # (B, L, V)\n",
        "\n",
        "    # Compute KL divergence (Teacher || Student) per token\n",
        "    kl_div = F.kl_div(\n",
        "        s_log_probs,\n",
        "        t_probs,\n",
        "        reduction=\"none\",  # Keep per-element losses\n",
        "        log_target=False\n",
        "    )  # (B, L, V)\n",
        "\n",
        "    kl_per_token = kl_div.sum(-1)  # Sum over vocabulary: (B, L)\n",
        "\n",
        "    # Apply attention mask to ignore padding tokens\n",
        "    mask = attention_mask.float()   # (B, L)\n",
        "\n",
        "    # Compute average loss over all non-padded tokens\n",
        "    loss = (kl_per_token * mask).sum() / mask.sum()\n",
        "\n",
        "    # Scale the loss by temperature squared as per standard KD practice\n",
        "    loss = loss * (temperature ** 2)\n",
        "\n",
        "    return loss\n",
        "\n",
        "# -----------------------------------------------------\n",
        "# Optimizer & Scheduler\n",
        "# -----------------------------------------------------\n",
        "optimizer = AdamW(student_model.parameters(), lr=LR)\n",
        "scheduler = get_linear_schedule_with_warmup(\n",
        "    optimizer,\n",
        "    num_warmup_steps=WARMUP_STEPS,\n",
        "    num_training_steps=MAX_STEPS,\n",
        ")\n",
        "\n",
        "# -----------------------------------------------------\n",
        "# Train\n",
        "# -----------------------------------------------------\n",
        "print(\"Starting knowledge distillation training...\")\n",
        "global_step = 0\n",
        "losses = []\n",
        "train_iter = iter(train_loader)\n",
        "\n",
        "while global_step < MAX_STEPS:\n",
        "    try:\n",
        "        batch = next(train_iter)\n",
        "    except StopIteration:\n",
        "        train_iter = iter(train_loader)\n",
        "        batch = next(train_iter)\n",
        "\n",
        "    input_ids = batch[\"input_ids\"].to(DEVICE)\n",
        "    attention_mask = batch[\"attention_mask\"].to(DEVICE)\n",
        "\n",
        "    # Forward teacher (no gradient)\n",
        "    with torch.no_grad():\n",
        "        teacher_outputs = teacher_model(input_ids, attention_mask=attention_mask)\n",
        "        teacher_logits = teacher_outputs.logits  # (B, L, V)\n",
        "\n",
        "    # Forward student\n",
        "    student_outputs = student_model(input_ids, attention_mask=attention_mask)\n",
        "    student_logits = student_outputs.logits      # (B, L, V)\n",
        "\n",
        "    # Compute distillation loss (Reverse KL)\n",
        "    loss = distillation_loss(\n",
        "        student_logits=student_logits,\n",
        "        teacher_logits=teacher_logits,\n",
        "        attention_mask=attention_mask,\n",
        "        temperature=TEMPERATURE,\n",
        "    )\n",
        "\n",
        "    # Backprop and update\n",
        "    optimizer.zero_grad()\n",
        "    loss.backward()\n",
        "\n",
        "    # Gradient Clipping\n",
        "    torch.nn.utils.clip_grad_norm_(student_model.parameters(), GRAD_CLIP)\n",
        "\n",
        "    optimizer.step()\n",
        "    scheduler.step()\n",
        "\n",
        "    # Record loss and step\n",
        "    losses.append(loss.item())\n",
        "    global_step += 1\n",
        "\n",
        "    # Logging\n",
        "    if global_step % LOG_EVERY == 0:\n",
        "        print(f\"step {global_step}/{MAX_STEPS}, loss: {loss.item():.4f}\")\n",
        "\n",
        "print(\"Finished training!\")\n",
        "plt.figure(figsize=(8,6))\n",
        "plt.plot(losses, label=\"Distillation Loss\")\n",
        "plt.xlabel(\"Step\")\n",
        "plt.ylabel(\"Loss\")\n",
        "plt.title(\"Distillation Loss (Reverse KL) over Steps\")\n",
        "plt.legend()\n",
        "plt.grid(True)\n",
        "plt.tight_layout()\n",
        "plt.savefig(\"distillation_loss.png\")\n",
        "print(\"Saved plot to distillation_loss.png\")\n",
        "\n",
        "# -----------------------------------------------------\n",
        "# Save Trained Student Model\n",
        "# -----------------------------------------------------\n",
        "output_dir = \"./distilled-student-gpt2\"\n",
        "student_model.save_pretrained(output_dir)\n",
        "tokenizer.save_pretrained(output_dir)\n",
        "print(f\"Saved distilled student model to {output_dir}\")\n",
        "\n",
        "# -----------------------------------------------------\n",
        "# Sample from the Distilled Student Model\n",
        "# -----------------------------------------------------\n",
        "def generate(prompt, max_length=50, temperature=1.0, top_k=50, top_p=0.95):\n",
        "    \"\"\"\n",
        "    Generate text from the distilled GPT-2 student model.\n",
        "    Pass attention_mask to avoid warnings and set pad_token_id = eos_token_id.\n",
        "    \"\"\"\n",
        "    student_model.eval()\n",
        "    inputs = tokenizer(prompt, return_tensors=\"pt\", padding=True, truncation=True, max_length=MAX_SEQ_LEN)\n",
        "    input_ids = inputs[\"input_ids\"].to(DEVICE)\n",
        "    attention_mask = inputs[\"attention_mask\"].to(DEVICE)\n",
        "\n",
        "    with torch.no_grad():\n",
        "        output_ids = student_model.generate(\n",
        "            input_ids=input_ids,\n",
        "            attention_mask=attention_mask,\n",
        "            max_length=max_length,\n",
        "            temperature=temperature,\n",
        "            top_k=top_k,\n",
        "            top_p=top_p,\n",
        "            do_sample=True,\n",
        "            pad_token_id=tokenizer.eos_token_id,  # to avoid warnings\n",
        "        )\n",
        "    return tokenizer.decode(output_ids[0], skip_special_tokens=True)\n",
        "\n",
        "sample_prompts = [\n",
        "    \"The curious scientist decided to test a new hypothesis:\",\n",
        "    \"Once upon a time in a land far, far away,\",\n",
        "    \"In a shocking turn of events today,\"\n",
        "]\n",
        "\n",
        "print(\"\\n-- Sample Generations from Distilled Student --\")\n",
        "for prompt in sample_prompts:\n",
        "    output = generate(prompt)\n",
        "    print(f\"Prompt: {prompt}\\nCompletion: {output}\\n\")\n"
      ]
    }
  ],
  "metadata": {
    "accelerator": "GPU",
    "colab": {
      "gpuType": "T4",
      "provenance": []
    },
    "kernelspec": {
      "display_name": "Python 3",
      "name": "python3"
    },
    "language_info": {
      "name": "python"
    }
  },
  "nbformat": 4,
  "nbformat_minor": 0
}