analyze_layers.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 20,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "{'aeroplane': 0, 'bicycle': 1, 'bus': 2, 'car': 3, 'horse': 4, 'knife': 5, 'motorcycle': 6, 'person': 7, 'plant': 8, 'skateboard': 9, 'train': 10, 'truck': 11}\n",
      "{'aeroplane': 0, 'bicycle': 1, 'bus': 2, 'car': 3, 'horse': 4, 'knife': 5, 'motorcycle': 6, 'person': 7, 'plant': 8, 'skateboard': 9, 'train': 10, 'truck': 11}\n",
      "{'aeroplane': 0, 'bicycle': 1, 'bus': 2, 'car': 3, 'horse': 4, 'knife': 5, 'motorcycle': 6, 'person': 7, 'plant': 8, 'skateboard': 9, 'train': 10, 'truck': 11}\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 71/71 [00:30<00:00,  2.32it/s]\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "adaptive             - 70.02570054717295\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 71/71 [00:25<00:00,  2.82it/s]\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[]                   - 31.188857569225668\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 71/71 [00:25<00:00,  2.82it/s]\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11] - 70.58254573591996\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 71/71 [00:25<00:00,  2.82it/s]\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[4, 8, 9, 10]        - 67.03559387608468\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 71/71 [00:25<00:00,  2.82it/s]\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[4, 8, 10]           - 64.83999336760073\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 71/71 [00:25<00:00,  2.82it/s]\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[8, 9, 10]           - 63.2537445420881\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 71/71 [00:25<00:00,  2.80it/s]\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[10]                 - 46.00950643895429\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 71/71 [00:25<00:00,  2.80it/s]\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[4]                  - 39.34947217155806\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 71/71 [00:25<00:00,  2.80it/s]\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[8]                  - 51.0667108826618\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 71/71 [00:25<00:00,  2.80it/s]\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[9]                  - 46.891062841983086\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 71/71 [00:25<00:00,  2.79it/s]"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[0]                  - 32.0220527275742\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    }
   ],
   "source": [
    "'''\n",
    "This Script analyzes a trained SpotTune model to understand the fine-tuning decisions being made\n",
    "'''\n",
    "\n",
    "import torch\n",
    "from torch.autograd import Variable\n",
    "\n",
    "import os\n",
    "import numpy as np\n",
    "import matplotlib.pyplot as plt\n",
    "import seaborn as sns\n",
    "from tqdm import tqdm\n",
    "\n",
    "# import imdbfolder as imdbfolder\n",
    "from spottune_models import *\n",
    "\n",
    "from utils import *\n",
    "from gumbel_softmax import *\n",
    "\n",
    "from visda17 import get_visda_dataloaders\n",
    "import shutil\n",
    "\n",
    "load_path = 'spottune_visda_v3_best.ckpt'\n",
    "# load_path = 'checkpoints/apricot-river-87_epoch4.ckpt'\n",
    "device = 'cuda:1'\n",
    "\n",
    "# Load the net and agent\n",
    "ckpt = torch.load(load_path, map_location=device)\n",
    "net = ckpt['net']\n",
    "agent = ckpt['agent']\n",
    "\n",
    "train_loader, val_loader, test_loader = get_visda_dataloaders(\n",
    "    train_dir='/export/r32/data/visda17/train',\n",
    "    val_dir='/export/r32/data/visda17/validation',\n",
    "    test_dir='/export/r32/data/visda17/test',\n",
    "    batch_size_test=1024\n",
    ")\n",
    "\n",
    "net.eval()\n",
    "agent.eval()\n",
    "\n",
    "policies = [\n",
    "    'adaptive',\n",
    "    [],\n",
    "    list(range(12)),\n",
    "    [4, 8, 9, 10],\n",
    "    [4, 8, 10],\n",
    "    [8, 9, 10],\n",
    "    [10],\n",
    "    [4],\n",
    "    [8],\n",
    "    [9],\n",
    "    [0]\n",
    "]\n",
    "\n",
    "for k_blocks in policies:\n",
    "    tasks_top1 = AverageMeter()\n",
    "    tasks_losses = AverageMeter() \n",
    "\n",
    "    with torch.no_grad():\n",
    "        for i, (images, labels) in enumerate(tqdm(test_loader)):\n",
    "            if torch.cuda.is_available():\n",
    "                images, labels = images.to(device=device,non_blocking=True), labels.to(device=device, non_blocking=True)\n",
    "            images, labels = Variable(images), Variable(labels)\n",
    "\n",
    "            probs = agent(images)\n",
    "            action = gumbel_softmax(probs.view(probs.size(0), -1, 2))\n",
    "            policy = action[:,:,1]\n",
    "\n",
    "            # override the policy to simulate top-k spottune variant\n",
    "            # k_blocks = [8]\n",
    "            if k_blocks == 'adaptive':\n",
    "                used_policy = policy\n",
    "            else:\n",
    "                used_policy = torch.zeros_like(policy)\n",
    "                used_policy[:, k_blocks] = 1.0\n",
    "\n",
    "            # ax = sns.heatmap(np.mean(policy.cpu().numpy(), dim=0), linewidth=0.5, cmap = sns.color_palette(\"crest_r\"))\n",
    "            # sns.set(rc = {'figure.figsize':(24,6)})\n",
    "            # # plt.yticks(np.arange(len(model_names))+0.5, [f\"{exp_name}\\n{exp_acc:.2f}%\" for exp_name, exp_acc in zip(model_names, accs)],\n",
    "            # #         rotation=0, fontsize=\"10\", va=\"center\")\n",
    "            # # plt.xticks(np.arange(on.shape[1]), model.getIndicatorLayers(), rotation=45, ha=\"right\", va='center', rotation_mode='anchor')\n",
    "            # plt.title(\"Adapted Layers\", size = 'x-large')\n",
    "            # plt.show()\n",
    "            # break\n",
    "\n",
    "            outputs = net.forward(images, used_policy)\n",
    "\n",
    "            _, predicted = torch.max(outputs.data, 1)\n",
    "            correct = predicted.eq(labels.data).cpu().sum()\n",
    "            tasks_top1.update(correct.item()*100 / (labels.size(0)+0.0), labels.size(0))\n",
    "\n",
    "            # break\n",
    "\n",
    "\n",
    "    # print(f\"Test accuracy: {tasks_top1.avg}\")\n",
    "    print(f\"{str(k_blocks).ljust(20)} - {tasks_top1.avg}\")\n",
    "    # print(f\"Test count: {tasks_top1.count}\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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",
      "text/plain": [
       "<Figure size 2400x600 with 2 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[[0.4888889  0.73333335 0.65555555 0.6333333  0.9222222  0.85555553\n",
      "  0.9222222  0.59444445 0.98333335 0.98333335 0.92777777 0.65      ]]\n"
     ]
    }
   ],
   "source": [
    "ax = sns.heatmap(np.mean(policy.cpu().numpy(), axis=0, keepdims=True), annot=True, fmt='.3f', linewidth=0.5, cmap = sns.color_palette(\"crest_r\"))\n",
    "sns.set(rc = {'figure.figsize':(24,6)})\n",
    "# plt.yticks(np.arange(len(model_names))+0.5, [f\"{exp_name}\\n{exp_acc:.2f}%\" for exp_name, exp_acc in zip(model_names, accs)],\n",
    "#         rotation=0, fontsize=\"10\", va=\"center\")\n",
    "# plt.xticks(np.arange(on.shape[1]), model.getIndicatorLayers(), rotation=45, ha=\"right\", va='center', rotation_mode='anchor')\n",
    "plt.title(\"Adapted Layers\", size = 'x-large')\n",
    "plt.show()\n",
    "\n",
    "\n",
    "print(np.mean(policy.cpu().numpy(), axis=0, keepdims=True))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": []
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "tensor([[0., 1., 1.,  ..., 0., 0., 0.],\n",
      "        [0., 1., 1.,  ..., 0., 0., 0.],\n",
      "        [0., 1., 1.,  ..., 0., 0., 0.],\n",
      "        ...,\n",
      "        [0., 1., 1.,  ..., 0., 0., 0.],\n",
      "        [0., 1., 1.,  ..., 0., 0., 0.],\n",
      "        [0., 1., 1.,  ..., 0., 0., 0.]], device='cuda:0')\n"
     ]
    }
   ],
   "source": [
    "new_policy = torch.zeros_like(policy)\n",
    "new_policy[:,[1,2]] = 1\n",
    "\n",
    "print(new_policy)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "2462648"
      ]
     },
     "execution_count": 15,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "\n",
    "sum([p.numel() for p in agent.parameters() if p.requires_grad])"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "taps",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.10.9"
  },
  "orig_nbformat": 4
 },
 "nbformat": 4,
 "nbformat_minor": 2
}