SimGAN-Captcha.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# SimGAN-Captcha\n",
    "\n",
    "With this method, breaking captchas has never been easier. You don't need to labeling any data. However, a hand tuned captcha generator is required to save training time.\n",
    "\n",
    "Its underlying model is SimGAN developed by Apple ([pdf](https://arxiv.org/pdf/1612.07828v1.pdf)). The model uses a refiner-discriminator architecture. \n",
    "\n",
    "![SimGAN](http://www.fudzilla.com/images/stories/2016/December/apple-simgan-generative-adversarial-networks.jpg)\n",
    "\n",
    "# The task\n",
    "\n",
    "[HackMIT Puzzle #5](https://captcha.delorean.codes/u/rickyhan/).\n",
    "\n",
    "Correctly label 10000 out of 15000 captcha or 90% per character."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Preprocessing"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Download target captcha files\n",
    "\n",
    "Here we download some captchas from the contest website. Each batch has 1000 captchas. We'll use 20000 so 20 batches."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "import requests\n",
    "import threading\n",
    "URL = \"https://captcha.delorean.codes/u/rickyhan/challenge\"\n",
    "DIR = \"challenges/\"\n",
    "NUM_CHALLENGES = 20\n",
    "lock = threading.Lock()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Done\n"
     ]
    }
   ],
   "source": [
    "def download_file(url, fname):\n",
    "    # NOTE the stream=True parameter\n",
    "    r = requests.get(url, stream=True)\n",
    "    with open(fname, 'wb') as f:\n",
    "        for chunk in r.iter_content(chunk_size=1024): \n",
    "            if chunk: # filter out keep-alive new chunks\n",
    "                f.write(chunk)\n",
    "                #f.flush() commented by recommendation from J.F.Sebastian\n",
    "    with lock:\n",
    "        pass\n",
    "        # print fname\n",
    "\n",
    "\n",
    "ts = []\n",
    "for i in range(NUM_CHALLENGES):\n",
    "    fname = DIR + \"challenge-{}\".format(i)\n",
    "    t = threading.Thread(target=download_file, args=(URL, fname))\n",
    "    ts.append(t)\n",
    "    t.start()\n",
    "for t in ts:\n",
    "    t.join()\n",
    "print \"Done\""
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Decompression\n",
    "\n",
    "Each challenge file is actually a json object containing 1000 base64 encoded jpg image file. So for each of these challenge files, we decompress each base64 strs into a jpeg and put that under a seprate folder."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 91,
   "metadata": {
    "scrolled": true
   },
   "outputs": [],
   "source": [
    "import json, base64, os\n",
    "IMG_DIR = \"./orig\"\n",
    "fnames = [\"{}/challenge-{}\".format(DIR, i) for i in range(NUM_CHALLENGES)]\n",
    "if not os.path.exists(IMG_DIR):\n",
    "    os.mkdir(IMG_DIR)\n",
    "def save_imgs(fname):\n",
    "    with open(fname) as f:\n",
    "        l = json.loads(f.read())\n",
    "\n",
    "    for image in l['images']:\n",
    "        b = base64.decodestring(image['jpg_base64'])\n",
    "        name = image['name']\n",
    "        with open(IMG_DIR+\"/{}.jpg\".format(name), 'w') as f:\n",
    "            f.write(b)\n",
    "\n",
    "for fname in fnames:\n",
    "    save_imgs(fname)\n",
    "assert len(os.listdir(IMG_DIR)) == 1000 * NUM_CHALLENGES"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 87,
   "metadata": {},
   "outputs": [],
   "source": [
    "from PIL import Image\n",
    "imgpath = IMG_DIR + \"/\"+ os.listdir(IMG_DIR)[0]\n",
    "imgpath2 = IMG_DIR + \"/\"+ os.listdir(IMG_DIR)[3]\n",
    "im = Image.open(example_image_path)\n",
    "im2 = Image.open(example_image_path2)\n",
    "IMG_FNAMES = [IMG_DIR + '/' + p for p in os.listdir(IMG_DIR)]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 88,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAGQAAAAyCAIAAAAlV+npAAAkmklEQVR4nNV7aZRdR3XuV8MZ77lz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\n/C9Z0mDAcSXC3AAAAABJRU5ErkJggg==\n",
      "text/plain": [
       "<PIL.JpegImagePlugin.JpegImageFile image mode=RGB size=100x50 at 0x7F16C4831350>"
      ]
     },
     "execution_count": 88,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "im"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 89,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAGQAAAAyCAIAAAAlV+npAAAisklEQVR4nM2beZRcR33vv1V3v7dv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\n",
      "text/plain": [
       "<PIL.JpegImagePlugin.JpegImageFile image mode=RGB size=100x50 at 0x7F16C4831190>"
      ]
     },
     "execution_count": 89,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "im2"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Convert to black and white\n",
    "Instead of RGB, binarized image saves significant compute. Here we hardcode a threshold and iterate over each pixel to obtain a binary image."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 92,
   "metadata": {},
   "outputs": [],
   "source": [
    "def gray(img_path):\n",
    "    # convert to grayscale, then binarize\n",
    "    img = Image.open(img_path).convert(\"L\")\n",
    "    img = img.point(lambda x: 255 if x > 200 or x == 0 else x) # value found through T&E\n",
    "    img = img.point(lambda x: 0 if x < 255 else 255, \"1\")\n",
    "    img.save(img_path)\n",
    "\n",
    "for img_path in IMG_FNAMES:\n",
    "    gray(img_path)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 93,
   "metadata": {
    "scrolled": false
   },
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAGQAAAAyCAAAAACPXiFiAAAGgUlEQVR4nLWYWW8j1xWEv7vf3riJ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\nAAAAAElFTkSuQmCC\n",
      "text/plain": [
       "<PIL.JpegImagePlugin.JpegImageFile image mode=L size=100x50 at 0x7F16C483A550>"
      ]
     },
     "execution_count": 93,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "im = Image.open(example_image_path)\n",
    "im"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Find mask\n",
    "\n",
    "As you may have noticed, all the captchas share the same horizontal lines. Since this is a contest, it was a function of participant's username. In the real world, these noises can be filtered out using morphological transformation with OpenCV.\n",
    "\n",
    "We will extract and save the lines(noise) for later use. Here we average all 20000 captchas and set a threshold as above. Another method is using a bit mask (&=) to iteratively filter out surrounding black pixels i.e.\n",
    "\n",
    "```\n",
    "mask = np.ones((height, width))\n",
    "for im in ims:\n",
    "    mask &= im\n",
    "```\n",
    "\n",
    "The effectiveness of bit mask depends on how clean the binarized data is. With the averaging method, some error is allowed."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 116,
   "metadata": {},
   "outputs": [],
   "source": [
    "import numpy as np\n",
    "WIDTH, HEIGHT = im.size\n",
    "MASK_DIR = \"avg.png\""
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 108,
   "metadata": {},
   "outputs": [],
   "source": [
    "def generateMask():\n",
    "    N=1000*NUM_CHALLENGES\n",
    "    arr=np.zeros((HEIGHT, WIDTH),np.float)\n",
    "    for fname in IMG_FNAMES:\n",
    "        imarr=np.array(Image.open(fname),dtype=np.float)\n",
    "        arr=arr+imarr/N\n",
    "    arr=np.array(np.round(arr),dtype=np.uint8)\n",
    "    out=Image.fromarray(arr,mode=\"L\")\n",
    "    out.save(MASK_DIR)\n",
    "\n",
    "generateMask()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 109,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAGQAAAAyCAAAAACPXiFiAAAJrUlEQVR4nLVXTatt2VUd82Ot/XHO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\nW4RaBPBDf9cpWOyCDg4AAAAASUVORK5CYII=\n",
      "text/plain": [
       "<PIL.PngImagePlugin.PngImageFile image mode=L size=100x50 at 0x7F16C483AF90>"
      ]
     },
     "execution_count": 109,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "im = Image.open(MASK_DIR) # ok this can be done with binary mask: &=\n",
    "im"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 162,
   "metadata": {},
   "outputs": [],
   "source": [
    "im = Image.open(MASK_DIR)\n",
    "im = im.point(lambda x:255 if x > 230 else x)\n",
    "im = im.point(lambda x:0 if x<255 else 255, \"1\")\n",
    "im.save(MASK_DIR)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 163,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAGQAAAAyAQAAAACCTkMTAAAA4UlEQVR4nLWRMW7DMAxFHy0DCtDC\nXjMYaQ/QIUOGDAZ6kR6iR1Bv0iNk9Oi9l9CWoUuWAAmQmh1k1dTQMZr49PkpkoL7n2cTV6wLrSnI\nhRytAOIM/gKwTSA3qOEx0SRQgdMJeP2YU0anqtnNO3jVaaYOd4MqlybfA4gGQ0/RwMNp/1/jK9s4\nrWkcJIIx7gujL4wSYVzE/m9iAB/ozOsnvKEOScYK4Bj0UKRa46YwujEZ0zZ+Bv00YmMbR8ZmyUSH\nczBiayYGicWq+uKPvQtQZ7rGaIi33fIz8PVi6fptyyCBX5hiNIEm2TyfAAAAAElFTkSuQmCC\n",
      "text/plain": [
       "<PIL.Image.Image image mode=1 size=100x50 at 0x7F16596FAE90>"
      ]
     },
     "execution_count": 163,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "im"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Generator for real captchas\n",
    "\n",
    "Using a Keras built in generator function `flow_from_directory` to automatically import and preprocess real captchas from a folder."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 130,
   "metadata": {},
   "outputs": [],
   "source": [
    "from keras import models\n",
    "from keras import layers\n",
    "from keras import optimizers\n",
    "from keras import applications\n",
    "from keras.preprocessing import image\n",
    "import tensorflow as tf"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "# Real data generator\n",
    "\n",
    "datagen = image.ImageDataGenerator(\n",
    "    preprocessing_function=applications.xception.preprocess_input\n",
    ")\n",
    "\n",
    "flow_from_directory_params = {'target_size': (HEIGHT, WIDTH),\n",
    "                              'color_mode': 'grayscale',\n",
    "                              'class_mode': None,\n",
    "                              'batch_size': BATCH_SIZE}\n",
    "\n",
    "real_generator = datagen.flow_from_directory(\n",
    "        directory=\".\",\n",
    "        **flow_from_directory_params\n",
    ")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# (Dumb) Generator\n",
    "\n",
    "Now that we have processed all the real captchas, we need to define a generator that outputs (captcha, label) pairs where the captchas should look almost like the real ones.\n",
    "\n",
    "We filter out the outliers that contain overlapping characters."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 256,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Synthetic captcha generator\n",
    "from PIL import ImageFont, ImageDraw\n",
    "from random import choice, random\n",
    "from string import ascii_lowercase, digits\n",
    "alphanumeric = ascii_lowercase + digits\n",
    "\n",
    "\n",
    "def fuzzy_loc(locs):\n",
    "    acc = []\n",
    "    for i,loc in enumerate(locs[:-1]):\n",
    "        if locs[i+1] - loc < 8:\n",
    "            continue\n",
    "        else:\n",
    "            acc.append(loc)\n",
    "    return acc\n",
    "\n",
    "def seg(img):\n",
    "    arr = np.array(img, dtype=np.float)\n",
    "    arr = arr.transpose()\n",
    "    # arr = np.mean(arr, axis=2)\n",
    "    arr = np.sum(arr, axis=1)\n",
    "    locs = np.where(arr < arr.min() + 2)[0].tolist()\n",
    "    locs = fuzzy_loc(locs)\n",
    "    return locs\n",
    "\n",
    "def is_well_formed(img_path):\n",
    "    original_img = Image.open(img_path)\n",
    "    img = original_img.convert('1')\n",
    "    return len(seg(img)) == 4\n",
    "\n",
    "noiseimg = np.array(Image.open(\"avg.png\").convert(\"1\"))\n",
    "# noiseimg = np.bitwise_not(noiseimg)\n",
    "fnt = ImageFont.truetype('./arial-extra.otf', 26)\n",
    "def gen_one():\n",
    "    og = Image.new(\"1\", (100,50))\n",
    "    text = ''.join([choice(alphanumeric) for _ in range(4)])\n",
    "    draw = ImageDraw.Draw(og)\n",
    "    for i, t in enumerate(text):\n",
    "        txt=Image.new('L', (40,40))\n",
    "        d = ImageDraw.Draw(txt)\n",
    "        d.text( (0, 0), t,  font=fnt, fill=255)\n",
    "        if random() > 0.5:\n",
    "            w=txt.rotate(-20*(random()-1),  expand=1)\n",
    "            og.paste( w, (i*20 + int(25*random()), int(25+30*(random()-1))),  w)\n",
    "        else:\n",
    "            w=txt.rotate(20*(random()-1),  expand=1)\n",
    "            og.paste( w, (i*20 + int(25*random()), int(20*random())),  w)\n",
    "    segments = seg(og)\n",
    "    if len(segments) != 4:\n",
    "        return gen_one()\n",
    "    ogarr = np.array(og)\n",
    "    ogarr = np.bitwise_or(noiseimg, ogarr)\n",
    "    ogarr = np.expand_dims(ogarr, axis=2).astype(float)\n",
    "    ogarr = np.random.random(size=(50,100,1)) * ogarr\n",
    "    ogarr = (ogarr > 0.0).astype(float) # add noise\n",
    "    return ogarr, text\n",
    "\n",
    "def synth_generator():\n",
    "    arrs = []\n",
    "    while True:\n",
    "        for _ in range(BATCH_SIZE):\n",
    "            arrs.append(gen_one()[0])\n",
    "        yield np.array(arrs)\n",
    "        arrs = []"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 204,
   "metadata": {},
   "outputs": [],
   "source": [
    "def get_image_batch(generator):\n",
    "    \"\"\"keras generators may generate an incomplete batch for the last batch\"\"\"\n",
    "    img_batch = generator.next()\n",
    "    if len(img_batch) != BATCH_SIZE:\n",
    "        img_batch = generator.next()\n",
    "\n",
    "    assert len(img_batch) == BATCH_SIZE\n",
    "\n",
    "    return img_batch"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 306,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "<matplotlib.image.AxesImage at 0x7f160fda74d0>"
      ]
     },
     "execution_count": 306,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAXQAAADKCAYAAAC11LviAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nRU5ErkJggg==\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7f160fdd4510>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "import matplotlib.pyplot as plt\n",
    "imarr = get_image_batch(real_generator)[0, :, :, 0]\n",
    "plt.imshow(imarr)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 307,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(50, 100)\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<matplotlib.image.AxesImage at 0x7f160fdd4390>"
      ]
     },
     "execution_count": 307,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAXQAAADKCAYAAAC11LviAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7f160fcad310>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "imarr = get_image_batch(synth_generator())[0, :, :, 0]\n",
    "print imarr.shape\n",
    "plt.imshow(imarr)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# What happened next?\n",
    "\n",
    "Plug all the data in an MNIST-like classifier and call it a day. Unfortunately, it's not that simple.\n",
    "\n",
    "I actually spent a long time fine-tuning the network but accuracy plateued around 55% sampled. The passing requirement is 10000 out of 15000 submitted or 90% accuracy or 66% per char. I was facing a dilemma: tune the model even further or manually label x amount of data: \n",
    "```\n",
    "0.55 * (15000-x) + x = 10000\n",
    "                   x = 3888\n",
    "```\n",
    "\n",
    "Obviously I am not going to label 4000 captchas and break my neck in the process.\n",
    "\n",
    "Meanwhile, there happened a burnt out guy who decided to label all 10000 captchas. This dilligent dude was 2000 in. I asked if he is willing to collaborate on a solution. It's almost like he didn't want to label captchas anymore. He agreed immediately.\n",
    "\n",
    "Using the same model, accuracy immediately shot up to 95% and we both qualified for HackMIT.\n",
    "\n",
    "/aside\n",
    "\n",
    "After the contest, I perfected the model and got 95% without labelling a single image. Here is the model for SimGAN:\n",
    "\n",
    "![SimGAN](http://www.fudzilla.com/images/stories/2016/December/apple-simgan-generative-adversarial-networks.jpg)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Model Definition\n",
    "\n",
    "There are three components to the network:\n",
    "\n",
    "### Refiner\n",
    "\n",
    "The refiner network, Rθ, is a residual network (ResNet). It modifies the synthetic image on a pixel level, rather than holistically modifying the image content, preserving the global structure and annotations.\n",
    "\n",
    "### Discriminator\n",
    "\n",
    "The discriminator network Dφ, is a simple ConvNet that contains 5 conv layers and 2 max-pooling layers. It's abinary classifier that outputs whether a captcha is synthesized or real.\n",
    "\n",
    "### Combined\n",
    "\n",
    "Pipe the refined image into discriminator."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 264,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/usr/local/lib/python2.7/dist-packages/ipykernel_launcher.py:24: UserWarning: Update your `Conv2D` call to the Keras 2 API: `Conv2D(64, (3, 3), padding=\"same\", activation=\"relu\")`\n",
      "/usr/local/lib/python2.7/dist-packages/ipykernel_launcher.py:16: UserWarning: Update your `Conv2D` call to the Keras 2 API: `Conv2D(64, (3, 3), padding=\"same\")`\n",
      "  app.launch_new_instance()\n",
      "/usr/local/lib/python2.7/dist-packages/ipykernel_launcher.py:18: UserWarning: Update your `Conv2D` call to the Keras 2 API: `Conv2D(64, (3, 3), padding=\"same\")`\n",
      "/usr/local/lib/python2.7/dist-packages/ipykernel_launcher.py:20: UserWarning: The `merge` function is deprecated and will be removed after 08/2017. Use instead layers from `keras.layers.merge`, e.g. `add`, `concatenate`, etc.\n",
      "/usr/local/lib/python2.7/dist-packages/ipykernel_launcher.py:32: UserWarning: Update your `Conv2D` call to the Keras 2 API: `Conv2D(1, (1, 1), padding=\"same\", activation=\"tanh\")`\n",
      "/usr/local/lib/python2.7/dist-packages/ipykernel_launcher.py:55: UserWarning: Update your `Model` call to the Keras 2 API: `Model(outputs=Tensor(\"co..., name=\"refiner\", inputs=Tensor(\"in...)`\n",
      "/usr/local/lib/python2.7/dist-packages/ipykernel_launcher.py:41: UserWarning: Update your `Conv2D` call to the Keras 2 API: `Conv2D(96, (3, 3), padding=\"same\", strides=(2, 2), activation=\"relu\")`\n",
      "/usr/local/lib/python2.7/dist-packages/ipykernel_launcher.py:42: UserWarning: Update your `Conv2D` call to the Keras 2 API: `Conv2D(64, (3, 3), padding=\"same\", strides=(2, 2), activation=\"relu\")`\n",
      "/usr/local/lib/python2.7/dist-packages/ipykernel_launcher.py:43: UserWarning: Update your `MaxPooling2D` call to the Keras 2 API: `MaxPooling2D(padding=\"same\", strides=(1, 1), pool_size=(3, 3))`\n",
      "/usr/local/lib/python2.7/dist-packages/ipykernel_launcher.py:44: UserWarning: Update your `Conv2D` call to the Keras 2 API: `Conv2D(32, (3, 3), padding=\"same\", strides=(1, 1), activation=\"relu\")`\n",
      "/usr/local/lib/python2.7/dist-packages/ipykernel_launcher.py:45: UserWarning: Update your `Conv2D` call to the Keras 2 API: `Conv2D(32, (1, 1), padding=\"same\", strides=(1, 1), activation=\"relu\")`\n",
      "/usr/local/lib/python2.7/dist-packages/ipykernel_launcher.py:46: UserWarning: Update your `Conv2D` call to the Keras 2 API: `Conv2D(2, (1, 1), padding=\"same\", strides=(1, 1), activation=\"relu\")`\n",
      "/usr/local/lib/python2.7/dist-packages/ipykernel_launcher.py:61: UserWarning: Update your `Model` call to the Keras 2 API: `Model(outputs=Tensor(\"re..., name=\"discriminator\", inputs=Tensor(\"in...)`\n",
      "/usr/local/lib/python2.7/dist-packages/ipykernel_launcher.py:67: UserWarning: Update your `Model` call to the Keras 2 API: `Model(outputs=[<tf.Tenso..., name=\"combined\", inputs=Tensor(\"in...)`\n"
     ]
    }
   ],
   "source": [
    "def refiner_network(input_image_tensor):\n",
    "    \"\"\"\n",
    "    :param input_image_tensor: Input tensor that corresponds to a synthetic image.\n",
    "    :return: Output tensor that corresponds to a refined synthetic image.\n",
    "    \"\"\"\n",
    "    def resnet_block(input_features, nb_features=64, nb_kernel_rows=3, nb_kernel_cols=3):\n",
    "        \"\"\"\n",
    "        A ResNet block with two `nb_kernel_rows` x `nb_kernel_cols` convolutional layers,\n",
    "        each with `nb_features` feature maps.\n",
    "        See Figure 6 in https://arxiv.org/pdf/1612.07828v1.pdf.\n",
    "        :param input_features: Input tensor to ResNet block.\n",
    "        :return: Output tensor from ResNet block.\n",
    "        \"\"\"\n",
    "        y = layers.Convolution2D(nb_features, nb_kernel_rows, nb_kernel_cols, border_mode='same')(input_features)\n",
    "        y = layers.Activation('relu')(y)\n",
    "        y = layers.Convolution2D(nb_features, nb_kernel_rows, nb_kernel_cols, border_mode='same')(y)\n",
    "\n",
    "        y = layers.merge([input_features, y], mode='sum')\n",
    "        return layers.Activation('relu')(y)\n",
    "\n",
    "    # an input image of size w × h is convolved with 3 × 3 filters that output 64 feature maps\n",
    "    x = layers.Convolution2D(64, 3, 3, border_mode='same', activation='relu')(input_image_tensor)\n",
    "\n",
    "    # the output is passed through 4 ResNet blocks\n",
    "    for _ in range(4):\n",
    "        x = resnet_block(x)\n",
    "\n",
    "    # the output of the last ResNet block is passed to a 1 × 1 convolutional layer producing 1 feature map\n",
    "    # corresponding to the refined synthetic image\n",
    "    return layers.Convolution2D(1, 1, 1, border_mode='same', activation='tanh')(x)\n",
    "\n",
    "def discriminator_network(input_image_tensor):\n",
    "    \"\"\"\n",
    "    :param input_image_tensor: Input tensor corresponding to an image, either real or refined.\n",
    "    :return: Output tensor that corresponds to the probability of whether an image is real or refined.\n",
    "    \"\"\"\n",
    "    x = layers.Convolution2D(96, 3, 3, border_mode='same', subsample=(2, 2), activation='relu')(input_image_tensor)\n",
    "    x = layers.Convolution2D(64, 3, 3, border_mode='same', subsample=(2, 2), activation='relu')(x)\n",
    "    x = layers.MaxPooling2D(pool_size=(3, 3), border_mode='same', strides=(1, 1))(x)\n",
    "    x = layers.Convolution2D(32, 3, 3, border_mode='same', subsample=(1, 1), activation='relu')(x)\n",
    "    x = layers.Convolution2D(32, 1, 1, border_mode='same', subsample=(1, 1), activation='relu')(x)\n",
    "    x = layers.Convolution2D(2, 1, 1, border_mode='same', subsample=(1, 1), activation='relu')(x)\n",
    "\n",
    "    # here one feature map corresponds to `is_real` and the other to `is_refined`,\n",
    "    # and the custom loss function is then `tf.nn.sparse_softmax_cross_entropy_with_logits`\n",
    "    return layers.Reshape((-1, 2))(x)\n",
    "\n",
    "# Refiner\n",
    "synthetic_image_tensor = layers.Input(shape=(HEIGHT, WIDTH, 1))\n",
    "refined_image_tensor = refiner_network(synthetic_image_tensor)\n",
    "refiner_model = models.Model(input=synthetic_image_tensor, output=refined_image_tensor, name='refiner')\n",
    "\n",
    "# Discriminator\n",
    "refined_or_real_image_tensor = layers.Input(shape=(HEIGHT, WIDTH, 1))\n",
    "discriminator_output = discriminator_network(refined_or_real_image_tensor)\n",
    "discriminator_model = models.Model(input=refined_or_real_image_tensor, output=discriminator_output,\n",
    "                                   name='discriminator')\n",
    "\n",
    "# Combined\n",
    "refiner_model_output = refiner_model(synthetic_image_tensor)\n",
    "combined_output = discriminator_model(refiner_model_output)\n",
    "combined_model = models.Model(input=synthetic_image_tensor, output=[refiner_model_output, combined_output],\n",
    "                              name='combined')\n",
    "\n",
    "def self_regularization_loss(y_true, y_pred):\n",
    "    delta = 0.0001  # FIXME: need to figure out an appropriate value for this\n",
    "    return tf.multiply(delta, tf.reduce_sum(tf.abs(y_pred - y_true)))\n",
    "\n",
    "# define custom local adversarial loss (softmax for each image section) for the discriminator\n",
    "# the adversarial loss function is the sum of the cross-entropy losses over the local patches\n",
    "def local_adversarial_loss(y_true, y_pred):\n",
    "    # y_true and y_pred have shape (batch_size, # of local patches, 2), but really we just want to average over\n",
    "    # the local patches and batch size so we can reshape to (batch_size * # of local patches, 2)\n",
    "    y_true = tf.reshape(y_true, (-1, 2))\n",
    "    y_pred = tf.reshape(y_pred, (-1, 2))\n",
    "    loss = tf.nn.softmax_cross_entropy_with_logits(labels=y_true, logits=y_pred)\n",
    "\n",
    "    return tf.reduce_mean(loss)\n",
    "\n",
    "\n",
    "# compile models\n",
    "BATCH_SIZE = 512\n",
    "sgd = optimizers.RMSprop()\n",
    "\n",
    "refiner_model.compile(optimizer=sgd, loss=self_regularization_loss)\n",
    "discriminator_model.compile(optimizer=sgd, loss=local_adversarial_loss)\n",
    "discriminator_model.trainable = False\n",
    "combined_model.compile(optimizer=sgd, loss=[self_regularization_loss, local_adversarial_loss])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Pre-training\n",
    "\n",
    "It is not necessary to pre-train GANs but it seems pretraining makes GANs converge faster. Here we pre-train both models. For the refiner, we train by supplying the identity. For the discriminator, we train with the correct real, synth labeled pairs."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 234,
   "metadata": {},
   "outputs": [],
   "source": [
    "# the target labels for the cross-entropy loss layer are 0 for every yj (real) and 1 for every xi (refined)\n",
    "\n",
    "y_real = np.array([[[1.0, 0.0]] * discriminator_model.output_shape[1]] * BATCH_SIZE)\n",
    "y_refined = np.array([[[0.0, 1.0]] * discriminator_model.output_shape[1]] * BATCH_SIZE)\n",
    "assert y_real.shape == (BATCH_SIZE, discriminator_model.output_shape[1], 2)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 266,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "pre-training the refiner network...\n",
      "Saving batch of refined images during pre-training at step: 0.\n",
      "Refiner model self regularization loss: [ 0.05277019].\n",
      "Saving batch of refined images during pre-training at step: 10.\n",
      "Refiner model self regularization loss: [ 4.2269813].\n",
      "Saving batch of refined images during pre-training at step: 20.\n",
      "Refiner model self regularization loss: [ 0.76108101].\n",
      "Saving batch of refined images during pre-training at step: 30.\n",
      "Refiner model self regularization loss: [ 0.28633648].\n",
      "Saving batch of refined images during pre-training at step: 40.\n",
      "Refiner model self regularization loss: [ 0.19448772].\n",
      "Saving batch of refined images during pre-training at step: 50.\n",
      "Refiner model self regularization loss: [ 0.16131182].\n",
      "Saving batch of refined images during pre-training at step: 60.\n",
      "Refiner model self regularization loss: [ 0.11931724].\n",
      "Saving batch of refined images during pre-training at step: 70.\n",
      "Refiner model self regularization loss: [ 0.11075923].\n",
      "Saving batch of refined images during pre-training at step: 80.\n",
      "Refiner model self regularization loss: [ 0.10888441].\n",
      "Saving batch of refined images during pre-training at step: 90.\n",
      "Refiner model self regularization loss: [ 0.10765313].\n"
     ]
    }
   ],
   "source": [
    "LOG_INTERVAL = 10\n",
    "MODEL_DIR = \"./model/\"\n",
    "print('pre-training the refiner network...')\n",
    "gen_loss = np.zeros(shape=len(refiner_model.metrics_names))\n",
    "\n",
    "for i in range(100):\n",
    "    synthetic_image_batch = get_image_batch(synth_generator())\n",
    "    gen_loss = np.add(refiner_model.train_on_batch(synthetic_image_batch, synthetic_image_batch), gen_loss)\n",
    "\n",
    "    # log every `log_interval` steps\n",
    "    if not i % LOG_INTERVAL:\n",
    "        print('Refiner model self regularization loss: {}.'.format(gen_loss / LOG_INTERVAL))\n",
    "        gen_loss = np.zeros(shape=len(refiner_model.metrics_names))\n",
    "\n",
    "refiner_model.save(os.path.join(MODEL_DIR, 'refiner_model_pre_trained.h5'))\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 269,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "pre-training the discriminator network...\n",
      "Discriminator model loss: [ 0.04783788].\n"
     ]
    }
   ],
   "source": [
    "from tqdm import tqdm\n",
    "print('pre-training the discriminator network...')\n",
    "disc_loss = np.zeros(shape=len(discriminator_model.metrics_names))\n",
    "\n",
    "for _ in tqdm(range(100)):\n",
    "    real_image_batch = get_image_batch(real_generator)\n",
    "    disc_loss = np.add(discriminator_model.train_on_batch(real_image_batch, y_real), disc_loss)\n",
    "\n",
    "    synthetic_image_batch = get_image_batch(synth_generator())\n",
    "    refined_image_batch = refiner_model.predict_on_batch(synthetic_image_batch)\n",
    "    disc_loss = np.add(discriminator_model.train_on_batch(refined_image_batch, y_refined), disc_loss)\n",
    "\n",
    "discriminator_model.save(os.path.join(MODEL_DIR, 'discriminator_model_pre_trained.h5'))\n",
    "\n",
    "# hard-coded for now\n",
    "print('Discriminator model loss: {}.'.format(disc_loss / (100 * 2)))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Training\n",
    "\n",
    "This is the most important training step in which we refine a synthesized captcha, then pass it through the discriminator and backprop gradients."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 277,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Step: 0 of 1000.\n",
      "Saving batch of refined images at adversarial step: 0.\n",
      "Refiner model loss: [ 2.46834831  0.01272553  2.45562277].\n",
      "Discriminator model loss real: [  2.27849432e-07].\n",
      "Discriminator model loss refined: [  1.63936726e-05].\n",
      "Step: 1 of 1000.\n",
      "Step: 2 of 1000.\n",
      "Step: 3 of 1000.\n",
      "Step: 4 of 1000.\n",
      "Step: 5 of 1000.\n",
      "Step: 6 of 1000.\n",
      "Step: 7 of 1000.\n",
      "Step: 8 of 1000.\n",
      "Step: 9 of 1000.\n",
      "Step: 10 of 1000.\n",
      "Saving batch of refined images at adversarial step: 10.\n",
      "Refiner model loss: [ 27.00968537   0.11238954  26.8972959 ].\n",
      "Discriminator model loss real: [  1.26835085e-10].\n",
      "Discriminator model loss refined: [  4.44882481e-08].\n",
      "Step: 11 of 1000.\n",
      "Step: 12 of 1000.\n",
      "Step: 13 of 1000.\n",
      "Step: 14 of 1000.\n",
      "Step: 15 of 1000.\n",
      "Step: 16 of 1000.\n",
      "Step: 17 of 1000.\n",
      "Step: 18 of 1000.\n",
      "Step: 19 of 1000.\n",
      "Step: 20 of 1000.\n",
      "Saving batch of refined images at adversarial step: 20.\n",
      "Refiner model loss: [ 26.89902883   0.10987803  26.78915081].\n",
      "Discriminator model loss real: [  1.48619811e-07].\n",
      "Discriminator model loss refined: [  4.60907181e-08].\n",
      "Step: 21 of 1000.\n",
      "Step: 22 of 1000.\n",
      "Step: 23 of 1000.\n",
      "Step: 24 of 1000.\n",
      "Step: 25 of 1000.\n",
      "Step: 26 of 1000.\n",
      "Step: 27 of 1000.\n",
      "Step: 28 of 1000.\n",
      "Step: 29 of 1000.\n",
      "Step: 30 of 1000.\n",
      "Saving batch of refined images at adversarial step: 30.\n",
      "Refiner model loss: [ 25.93090506   0.10890296  25.82200208].\n",
      "Discriminator model loss real: [  3.96611703e-09].\n",
      "Discriminator model loss refined: [  5.07067440e-08].\n",
      "Step: 31 of 1000.\n",
      "Step: 32 of 1000.\n",
      "Step: 33 of 1000.\n",
      "Step: 34 of 1000.\n",
      "Step: 35 of 1000.\n",
      "Step: 36 of 1000.\n",
      "Step: 37 of 1000.\n",
      "Step: 38 of 1000.\n",
      "Step: 39 of 1000.\n",
      "Step: 40 of 1000.\n",
      "Saving batch of refined images at adversarial step: 40.\n",
      "Refiner model loss: [ 28.67232819   2.33041485  26.34191332].\n",
      "Discriminator model loss real: [  1.10259192e-09].\n",
      "Discriminator model loss refined: [  3.45548763e-08].\n",
      "Step: 41 of 1000.\n",
      "Step: 42 of 1000.\n",
      "Step: 43 of 1000.\n",
      "Step: 44 of 1000.\n",
      "Step: 45 of 1000.\n",
      "Step: 46 of 1000.\n",
      "Step: 47 of 1000.\n",
      "Step: 48 of 1000.\n",
      "Step: 49 of 1000.\n",
      "Step: 50 of 1000.\n",
      "Saving batch of refined images at adversarial step: 50.\n",
      "Refiner model loss: [ 27.58684621   0.10730441  27.47954178].\n",
      "Discriminator model loss real: [  1.50006033e-10].\n",
      "Discriminator model loss refined: [  1.93523770e-08].\n",
      "Step: 51 of 1000.\n",
      "Step: 52 of 1000.\n",
      "Step: 53 of 1000.\n",
      "Step: 54 of 1000.\n",
      "Step: 55 of 1000.\n",
      "Step: 56 of 1000.\n",
      "Step: 57 of 1000.\n",
      "Step: 58 of 1000.\n",
      "Step: 59 of 1000.\n",
      "Step: 60 of 1000.\n",
      "Saving batch of refined images at adversarial step: 60.\n",
      "Refiner model loss: [ 28.97681231   0.10622904  28.87058334].\n",
      "Discriminator model loss real: [  8.48147695e-11].\n",
      "Discriminator model loss refined: [  8.90565564e-09].\n",
      "Step: 61 of 1000.\n",
      "Step: 62 of 1000.\n",
      "Step: 63 of 1000.\n",
      "Step: 64 of 1000.\n",
      "Step: 65 of 1000.\n",
      "Step: 66 of 1000.\n",
      "Step: 67 of 1000.\n",
      "Step: 68 of 1000.\n",
      "Step: 69 of 1000.\n",
      "Step: 70 of 1000.\n",
      "Saving batch of refined images at adversarial step: 70.\n",
      "Refiner model loss: [ 28.66822872   0.1088825   28.55934606].\n",
      "Discriminator model loss real: [  5.22255202e-09].\n",
      "Discriminator model loss refined: [  6.25894856e-09].\n",
      "Step: 71 of 1000.\n",
      "Step: 72 of 1000.\n",
      "Step: 73 of 1000.\n",
      "Step: 74 of 1000.\n",
      "Step: 75 of 1000.\n",
      "Step: 76 of 1000.\n",
      "Step: 77 of 1000.\n",
      "Step: 78 of 1000.\n",
      "Step: 79 of 1000.\n",
      "Step: 80 of 1000.\n",
      "Saving batch of refined images at adversarial step: 80.\n",
      "Refiner model loss: [ 34.46133389   4.25902364  30.20231042].\n",
      "Discriminator model loss real: [  7.93363459e-11].\n",
      "Discriminator model loss refined: [  2.00646231e-09].\n",
      "Step: 81 of 1000.\n",
      "Step: 82 of 1000.\n",
      "Step: 83 of 1000.\n",
      "Step: 84 of 1000.\n",
      "Step: 85 of 1000.\n",
      "Step: 86 of 1000.\n",
      "Step: 87 of 1000.\n",
      "Step: 88 of 1000.\n",
      "Step: 89 of 1000.\n",
      "Step: 90 of 1000.\n",
      "Saving batch of refined images at adversarial step: 90.\n",
      "Refiner model loss: [ 31.02199268   0.10858321  30.91340938].\n",
      "Discriminator model loss real: [  2.88720957e-10].\n",
      "Discriminator model loss refined: [  1.10082318e-09].\n",
      "Step: 91 of 1000.\n",
      "Step: 92 of 1000.\n",
      "Step: 93 of 1000.\n",
      "Step: 94 of 1000.\n",
      "Step: 95 of 1000.\n",
      "Step: 96 of 1000.\n",
      "Step: 97 of 1000.\n",
      "Step: 98 of 1000.\n",
      "Step: 99 of 1000.\n",
      "Step: 100 of 1000.\n",
      "Saving batch of refined images at adversarial step: 100.\n",
      "Refiner model loss: [ 33.74043455   1.45508248  32.2853519 ].\n",
      "Discriminator model loss real: [  5.83133614e-11].\n",
      "Discriminator model loss refined: [  3.08375181e-10].\n",
      "Step: 101 of 1000.\n",
      "Step: 102 of 1000.\n",
      "Step: 103 of 1000.\n",
      "Step: 104 of 1000.\n",
      "Step: 105 of 1000.\n",
      "Step: 106 of 1000.\n",
      "Step: 107 of 1000.\n",
      "Step: 108 of 1000.\n",
      "Step: 109 of 1000.\n",
      "Step: 110 of 1000.\n",
      "Saving batch of refined images at adversarial step: 110.\n",
      "Refiner model loss: [ 33.19685087   0.1087612   33.08808966].\n",
      "Discriminator model loss real: [  8.33836509e-11].\n",
      "Discriminator model loss refined: [  1.60044156e-10].\n",
      "Step: 111 of 1000.\n",
      "Step: 112 of 1000.\n",
      "Step: 113 of 1000.\n",
      "Step: 114 of 1000.\n",
      "Step: 115 of 1000.\n",
      "Step: 116 of 1000.\n",
      "Step: 117 of 1000.\n",
      "Step: 118 of 1000.\n",
      "Step: 119 of 1000.\n",
      "Step: 120 of 1000.\n",
      "Saving batch of refined images at adversarial step: 120.\n",
      "Refiner model loss: [ 33.29263887   0.10645921  33.18617992].\n",
      "Discriminator model loss real: [  1.21886291e-10].\n",
      "Discriminator model loss refined: [  1.08857246e-10].\n",
      "Step: 121 of 1000.\n",
      "Step: 122 of 1000.\n",
      "Step: 123 of 1000.\n",
      "Step: 124 of 1000.\n",
      "Step: 125 of 1000.\n",
      "Step: 126 of 1000.\n",
      "Step: 127 of 1000.\n",
      "Step: 128 of 1000.\n",
      "Step: 129 of 1000.\n",
      "Step: 130 of 1000.\n",
      "Saving batch of refined images at adversarial step: 130.\n",
      "Refiner model loss: [ 33.38862801   0.10835286  33.28027554].\n",
      "Discriminator model loss real: [  1.24294775e-11].\n",
      "Discriminator model loss refined: [  2.17428002e-11].\n",
      "Step: 131 of 1000.\n",
      "Step: 132 of 1000.\n",
      "Step: 133 of 1000.\n",
      "Step: 134 of 1000.\n",
      "Step: 135 of 1000.\n",
      "Step: 136 of 1000.\n",
      "Step: 137 of 1000.\n",
      "Step: 138 of 1000.\n",
      "Step: 139 of 1000.\n",
      "Step: 140 of 1000.\n",
      "Saving batch of refined images at adversarial step: 140.\n",
      "Refiner model loss: [ 35.3840229    2.04447198  33.33955069].\n",
      "Discriminator model loss real: [  4.40586085e-12].\n",
      "Discriminator model loss refined: [  1.83398912e-11].\n",
      "Step: 141 of 1000.\n",
      "Step: 142 of 1000.\n",
      "Step: 143 of 1000.\n",
      "Step: 144 of 1000.\n",
      "Step: 145 of 1000.\n",
      "Step: 146 of 1000.\n",
      "Step: 147 of 1000.\n",
      "Step: 148 of 1000.\n",
      "Step: 149 of 1000.\n",
      "Step: 150 of 1000.\n",
      "Saving batch of refined images at adversarial step: 150.\n",
      "Refiner model loss: [ 33.46622419   0.11144205  33.3547821 ].\n",
      "Discriminator model loss real: [  6.55844760e-11].\n",
      "Discriminator model loss refined: [  1.91637525e-11].\n",
      "Step: 151 of 1000.\n",
      "Step: 152 of 1000.\n",
      "Step: 153 of 1000.\n",
      "Step: 154 of 1000.\n",
      "Step: 155 of 1000.\n",
      "Step: 156 of 1000.\n",
      "Step: 157 of 1000.\n",
      "Step: 158 of 1000.\n",
      "Step: 159 of 1000.\n",
      "Step: 160 of 1000.\n",
      "Saving batch of refined images at adversarial step: 160.\n",
      "Refiner model loss: [ 33.40007172   0.11607643  33.28399525].\n",
      "Discriminator model loss real: [  4.91435148e-11].\n",
      "Discriminator model loss refined: [  2.27457638e-11].\n",
      "Step: 161 of 1000.\n",
      "Step: 162 of 1000.\n",
      "Step: 163 of 1000.\n",
      "Step: 164 of 1000.\n",
      "Step: 165 of 1000.\n",
      "Step: 166 of 1000.\n",
      "Step: 167 of 1000.\n",
      "Step: 168 of 1000.\n",
      "Step: 169 of 1000.\n",
      "Step: 170 of 1000.\n",
      "Saving batch of refined images at adversarial step: 170.\n",
      "Refiner model loss: [ 35.22831955   1.84048391  33.3878355 ].\n",
      "Discriminator model loss real: [  4.94316164e-12].\n",
      "Discriminator model loss refined: [  2.08114768e-11].\n",
      "Step: 171 of 1000.\n",
      "Step: 172 of 1000.\n",
      "Step: 173 of 1000.\n",
      "Step: 174 of 1000.\n",
      "Step: 175 of 1000.\n",
      "Step: 176 of 1000.\n",
      "Step: 177 of 1000.\n",
      "Step: 178 of 1000.\n",
      "Step: 179 of 1000.\n",
      "Step: 180 of 1000.\n",
      "Saving batch of refined images at adversarial step: 180.\n",
      "Refiner model loss: [ 33.69246931   0.11457105  33.57789841].\n",
      "Discriminator model loss real: [  5.65956004e-12].\n",
      "Discriminator model loss refined: [  1.41489390e-11].\n",
      "Step: 181 of 1000.\n",
      "Step: 182 of 1000.\n",
      "Step: 183 of 1000.\n"
     ]
    },
    {
     "ename": "KeyboardInterrupt",
     "evalue": "",
     "output_type": "error",
     "traceback": [
      "\u001b[0;31m---------------------------------------------------------------------------\u001b[0m",
      "\u001b[0;31mKeyboardInterrupt\u001b[0m                         Traceback (most recent call last)",
      "\u001b[0;32m<ipython-input-277-40ad1fa6caa9>\u001b[0m in \u001b[0;36m<module>\u001b[0;34m()\u001b[0m\n\u001b[1;32m     24\u001b[0m         \u001b[0;31m# update θ by taking an SGD step on mini-batch loss LR(θ)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     25\u001b[0m         combined_loss = np.add(combined_model.train_on_batch(synthetic_image_batch,\n\u001b[0;32m---> 26\u001b[0;31m                                                              [synthetic_image_batch, y_real]), combined_loss)\n\u001b[0m\u001b[1;32m     27\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     28\u001b[0m     \u001b[0;32mfor\u001b[0m \u001b[0m_\u001b[0m \u001b[0;32min\u001b[0m \u001b[0mrange\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mk_d\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
      "\u001b[0;32m/usr/local/lib/python2.7/dist-packages/keras/engine/training.pyc\u001b[0m in \u001b[0;36mtrain_on_batch\u001b[0;34m(self, x, y, sample_weight, class_weight)\u001b[0m\n\u001b[1;32m   1640\u001b[0m             \u001b[0mins\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mx\u001b[0m \u001b[0;34m+\u001b[0m \u001b[0my\u001b[0m \u001b[0;34m+\u001b[0m \u001b[0msample_weights\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m   1641\u001b[0m         \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0m_make_train_function\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m-> 1642\u001b[0;31m         \u001b[0moutputs\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mtrain_function\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mins\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m   1643\u001b[0m         \u001b[0;32mif\u001b[0m \u001b[0mlen\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0moutputs\u001b[0m\u001b[0;34m)\u001b[0m \u001b[0;34m==\u001b[0m \u001b[0;36m1\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m   1644\u001b[0m             \u001b[0;32mreturn\u001b[0m \u001b[0moutputs\u001b[0m\u001b[0;34m[\u001b[0m\u001b[0;36m0\u001b[0m\u001b[0;34m]\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
      "\u001b[0;32m/usr/local/lib/python2.7/dist-packages/keras/backend/tensorflow_backend.pyc\u001b[0m in \u001b[0;36m__call__\u001b[0;34m(self, inputs)\u001b[0m\n\u001b[1;32m   2267\u001b[0m         updated = session.run(self.outputs + [self.updates_op],\n\u001b[1;32m   2268\u001b[0m                               \u001b[0mfeed_dict\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0mfeed_dict\u001b[0m\u001b[0;34m,\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m-> 2269\u001b[0;31m                               **self.session_kwargs)\n\u001b[0m\u001b[1;32m   2270\u001b[0m         \u001b[0;32mreturn\u001b[0m \u001b[0mupdated\u001b[0m\u001b[0;34m[\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0mlen\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0moutputs\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m]\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m   2271\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n",
      "\u001b[0;32m/usr/local/lib/python2.7/dist-packages/tensorflow/python/client/session.pyc\u001b[0m in \u001b[0;36mrun\u001b[0;34m(self, fetches, feed_dict, options, run_metadata)\u001b[0m\n\u001b[1;32m    787\u001b[0m     \u001b[0;32mtry\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m    788\u001b[0m       result = self._run(None, fetches, feed_dict, options_ptr,\n\u001b[0;32m--> 789\u001b[0;31m                          run_metadata_ptr)\n\u001b[0m\u001b[1;32m    790\u001b[0m       \u001b[0;32mif\u001b[0m \u001b[0mrun_metadata\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m    791\u001b[0m         \u001b[0mproto_data\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mtf_session\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mTF_GetBuffer\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mrun_metadata_ptr\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
      "\u001b[0;32m/usr/local/lib/python2.7/dist-packages/tensorflow/python/client/session.pyc\u001b[0m in \u001b[0;36m_run\u001b[0;34m(self, handle, fetches, feed_dict, options, run_metadata)\u001b[0m\n\u001b[1;32m    995\u001b[0m     \u001b[0;32mif\u001b[0m \u001b[0mfinal_fetches\u001b[0m \u001b[0;32mor\u001b[0m \u001b[0mfinal_targets\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m    996\u001b[0m       results = self._do_run(handle, final_targets, final_fetches,\n\u001b[0;32m--> 997\u001b[0;31m                              feed_dict_string, options, run_metadata)\n\u001b[0m\u001b[1;32m    998\u001b[0m     \u001b[0;32melse\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m    999\u001b[0m       \u001b[0mresults\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0;34m[\u001b[0m\u001b[0;34m]\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
      "\u001b[0;32m/usr/local/lib/python2.7/dist-packages/tensorflow/python/client/session.pyc\u001b[0m in \u001b[0;36m_do_run\u001b[0;34m(self, handle, target_list, fetch_list, feed_dict, options, run_metadata)\u001b[0m\n\u001b[1;32m   1130\u001b[0m     \u001b[0;32mif\u001b[0m \u001b[0mhandle\u001b[0m \u001b[0;32mis\u001b[0m \u001b[0mNone\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m   1131\u001b[0m       return self._do_call(_run_fn, self._session, feed_dict, fetch_list,\n\u001b[0;32m-> 1132\u001b[0;31m                            target_list, options, run_metadata)\n\u001b[0m\u001b[1;32m   1133\u001b[0m     \u001b[0;32melse\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m   1134\u001b[0m       return self._do_call(_prun_fn, self._session, handle, feed_dict,\n",
      "\u001b[0;32m/usr/local/lib/python2.7/dist-packages/tensorflow/python/client/session.pyc\u001b[0m in \u001b[0;36m_do_call\u001b[0;34m(self, fn, *args)\u001b[0m\n\u001b[1;32m   1137\u001b[0m   \u001b[0;32mdef\u001b[0m \u001b[0m_do_call\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mself\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mfn\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m*\u001b[0m\u001b[0margs\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m   1138\u001b[0m     \u001b[0;32mtry\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m-> 1139\u001b[0;31m       \u001b[0;32mreturn\u001b[0m \u001b[0mfn\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m*\u001b[0m\u001b[0margs\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m   1140\u001b[0m     \u001b[0;32mexcept\u001b[0m \u001b[0merrors\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mOpError\u001b[0m \u001b[0;32mas\u001b[0m \u001b[0me\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m   1141\u001b[0m       \u001b[0mmessage\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mcompat\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mas_text\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0me\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mmessage\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
      "\u001b[0;32m/usr/local/lib/python2.7/dist-packages/tensorflow/python/client/session.pyc\u001b[0m in \u001b[0;36m_run_fn\u001b[0;34m(session, feed_dict, fetch_list, target_list, options, run_metadata)\u001b[0m\n\u001b[1;32m   1119\u001b[0m         return tf_session.TF_Run(session, options,\n\u001b[1;32m   1120\u001b[0m                                  \u001b[0mfeed_dict\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mfetch_list\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mtarget_list\u001b[0m\u001b[0;34m,\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m-> 1121\u001b[0;31m                                  status, run_metadata)\n\u001b[0m\u001b[1;32m   1122\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m   1123\u001b[0m     \u001b[0;32mdef\u001b[0m \u001b[0m_prun_fn\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0msession\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mhandle\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mfeed_dict\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mfetch_list\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
      "\u001b[0;31mKeyboardInterrupt\u001b[0m: "
     ]
    }
   ],
   "source": [
    "from image_history_buffer import ImageHistoryBuffer\n",
    "\n",
    "\n",
    "k_d = 1  # number of discriminator updates per step\n",
    "k_g = 2  # number of generative network updates per step\n",
    "nb_steps = 1000\n",
    "\n",
    "# TODO: what is an appropriate size for the image history buffer?\n",
    "image_history_buffer = ImageHistoryBuffer((0, HEIGHT, WIDTH, 1), BATCH_SIZE * 100, BATCH_SIZE)\n",
    "\n",
    "combined_loss = np.zeros(shape=len(combined_model.metrics_names))\n",
    "disc_loss_real = np.zeros(shape=len(discriminator_model.metrics_names))\n",
    "disc_loss_refined = np.zeros(shape=len(discriminator_model.metrics_names))\n",
    "\n",
    "# see Algorithm 1 in https://arxiv.org/pdf/1612.07828v1.pdf\n",
    "for i in range(nb_steps):\n",
    "    print('Step: {} of {}.'.format(i, nb_steps))\n",
    "\n",
    "    # train the refiner\n",
    "    for _ in range(k_g * 2):\n",
    "        # sample a mini-batch of synthetic images\n",
    "        synthetic_image_batch = get_image_batch(synth_generator())\n",
    "\n",
    "        # update θ by taking an SGD step on mini-batch loss LR(θ)\n",
    "        combined_loss = np.add(combined_model.train_on_batch(synthetic_image_batch,\n",
    "                                                             [synthetic_image_batch, y_real]), combined_loss)\n",
    "\n",
    "    for _ in range(k_d):\n",
    "        # sample a mini-batch of synthetic and real images\n",
    "        synthetic_image_batch = get_image_batch(synth_generator())\n",
    "        real_image_batch = get_image_batch(real_generator)\n",
    "\n",
    "        # refine the synthetic images w/ the current refiner\n",
    "        refined_image_batch = refiner_model.predict_on_batch(synthetic_image_batch)\n",
    "\n",
    "        # use a history of refined images\n",
    "        half_batch_from_image_history = image_history_buffer.get_from_image_history_buffer()\n",
    "        image_history_buffer.add_to_image_history_buffer(refined_image_batch)\n",
    "\n",
    "        if len(half_batch_from_image_history):\n",
    "            refined_image_batch[:batch_size // 2] = half_batch_from_image_history\n",
    "\n",
    "        # update φ by taking an SGD step on mini-batch loss LD(φ)\n",
    "        disc_loss_real = np.add(discriminator_model.train_on_batch(real_image_batch, y_real), disc_loss_real)\n",
    "        disc_loss_refined = np.add(discriminator_model.train_on_batch(refined_image_batch, y_refined),\n",
    "                                   disc_loss_refined)\n",
    "\n",
    "    if not i % LOG_INTERVAL:\n",
    "        # log loss summary\n",
    "        print('Refiner model loss: {}.'.format(combined_loss / (LOG_INTERVAL * k_g * 2)))\n",
    "        print('Discriminator model loss real: {}.'.format(disc_loss_real / (LOG_INTERVAL * k_d * 2)))\n",
    "        print('Discriminator model loss refined: {}.'.format(disc_loss_refined / (LOG_INTERVAL * k_d * 2)))\n",
    "\n",
    "        combined_loss = np.zeros(shape=len(combined_model.metrics_names))\n",
    "        disc_loss_real = np.zeros(shape=len(discriminator_model.metrics_names))\n",
    "        disc_loss_refined = np.zeros(shape=len(discriminator_model.metrics_names))\n",
    "\n",
    "        # save model checkpoints\n",
    "        model_checkpoint_base_name = os.path.join(MODEL_DIR, '{}_model_step_{}.h5')\n",
    "        refiner_model.save(model_checkpoint_base_name.format('refiner', i))\n",
    "        discriminator_model.save(model_checkpoint_base_name.format('discriminator', i))\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Results of SimGAN:\n",
    "\n",
    "As you can see below, we no longer have the cookie-cutter fonts. There are quite a few artifacts that did not exist before refinement. The edges are blurred and noisy - which is *impossible* to simulate heuristically. And it is exactly these tiny things that renders MNIST-like convnet useless.\n",
    "\n",
    "Now the refined results are basically the original captchas."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 279,
   "metadata": {},
   "outputs": [],
   "source": [
    "synthetic_image_batch = get_image_batch(synth_generator())\n",
    "arr = refiner_model.predict_on_batch(synthetic_image_batch)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 300,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "<matplotlib.image.AxesImage at 0x7f161417fa90>"
      ]
     },
     "execution_count": 300,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAXQAAADKCAYAAAC11LviAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nktQIC7okNeJ/AVH/kaZ3SZoBAAAAAElFTkSuQmCC\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7f1614243290>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "plt.imshow(arr[200, :, :, 0])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 304,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "<matplotlib.image.AxesImage at 0x7f1614381690>"
      ]
     },
     "execution_count": 304,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAXQAAADKCAYAAAC11LviAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz\nAAALEgAACxIB0t1+/AAAIABJREFUeJztnXuQXNV957+/7tH7ARISstBISBiBAZcDBoMcEhZje8FA\nbBO7EtusF++yy67L67WzccUku1vrpJKteGsrdnbLa5cSvGbXlDHBNig8YxPY+IkRCD9AAQkJSSOE\nHkYChB4z0332j27d8z1n5hzd6ZnpEbe/nyqK03NP3z597+2je773+/v9zDkHIYQQr39qUz0AIYQQ\nE4MmdCGEqAia0IUQoiJoQhdCiIqgCV0IISqCJnQhhKgImtCFEKIijGtCN7OrzOwZM9tsZjdP1KCE\nEEKMHes0sMjM6gCeBfBuAAMAHgPwYefc0xM3PCGEEGXpG8d7Lwaw2Tm3BQDM7HYA7wOQnNCn2ww3\nE3PG8ZHjx/rq4R+a/h8012yW20ctXNiUfZ8QQnTCq9i/zzm3+Hj9xjOhLwOwg14PALgk94aZmINL\n7J3j+MgOMSua9ZMXBJvc4SNFu3noUPJ9oJVMbe68oFvz1VcnYJBCCDE633N3bivTbzwTeinM7CYA\nNwHATMye7I8TQoieZTwT+k4Ay+l1f/tvAc65tQDWAsB8W9i1TGC12f4fD77zdoNDQb8Rd+VE/ZSF\nRbux71f+PbojF0KcgIzH5fIYgNVmtsrMpgP4EIB1EzMsIYQQY6XjO3Tn3LCZ/TsADwKoA/iqc+6p\nCRuZEEKIMTEuDd05dx+A+yZoLEIIIcbBpD8UnSpYG6/Pn1+0G6+8EvTrW/qGoj2868VgG+vmqHm7\nY31uaL2M9ymEEFOBQv+FEKIiaEIXQoiKUFnJhWlS8FCMO3I0/UYOLGo2iqYkFiHEiYju0IUQoiJo\nQhdCiIrQE5KLGxos2vWzzwy2NZ7ZnHxfEG362msTPzAhhJhAdIcuhBAVQRO6EEJUBE3oQghREXpC\nQ7cZM4p2rJlbnz8Ebng42Ma6Oe+jRm1ANkYhxImB7tCFEKIiaEIXQoiKUFnJxaZNL9ruqI8GZSsi\nECbxskhKQcNHh/I+Gkcz0aVCCDFF6A5dCCEqgiZ0IYSoCJrQhRCiIlRWQ+dwf9bG79/8o6Dflaed\nn95J3Re1AFka+5adFnQb3vlCh6MUQoiJQ3foQghRETShCyFERais5FKb4+t+csTn/sahoN+DLzyZ\n3AfLMfVTFhbt4V27ow/z0oxNo8jTyN4YZG88FI6DqS9YULQb+/f7DVxwAwCco88lmybZLVsfFr0+\nNp6ZM8Nug0O0LbRwJsdL3701Dv/9OaJW0bTdo37ySUW7ceDlcu/J1N0te93moq75Wmse8QVnuKYv\nMLKub7HvnKWYPivuV5vlP7f0saDfeuNXL5V6DwD09S8r2sMDO0u/byLRHboQQlQETehCCFERzNGy\nfbKZbwvdJfbOrn3eMXg52YxkEJZFcvLL9uGDRftfr/iNZL/U0vJ425L7o+WuGxxMbuNlMi99W3+g\nf7drXraJJaHsOObNG/XvzVdf7ej9zddo6Z6QhMT4YfmOzz0QyglB8rl5c8N++37lt8UyHV/HLL/F\n55S3uSa1M/MPSYxWD6W9WNIpPoakViCdYG/Etc9yZmZMfW9YUrSHX9yd7kdOuOYr4W+k7G+G+Z67\n83Hn3EXH66c7dCGEqAjHndDN7KtmtsfMfkl/W2hm3zWzTe3/L8jtQwghxORT5g79awCuiv52M4CH\nnHOrATzUfi2EEGIKKaWhm9lKAPc4597cfv0MgMudc7vMbCmAR5xzZx9vP93U0FN69Qgtt6SeldPX\nGbY69p2+PNg2vG3HqO+pL14cvLYZ3oKYsz/1Le/3/XYM0A7S9sYcgfVxaDDTMw1/Z/6+vO8R+4+s\nj2Js1E+aH7wOrK4ZUuc79wxmxHWR0M1j+2Aq46lNnxb0a7zin1Xlnq2wRu3o9x3bDFO6eZx1lZ9P\nsT6fe/azbsdPgm3v7b+YdjixzyYnW0Nf4pzb1W6/CGBJrrMQQojJZ9wPRV3rFj/5z5GZ3WRm681s\n/RCUR1wIISaLTiNFd5vZUpJc9qQ6OufWAlgLtCSXDj9vzLDMwpFzbmh0uxMA9K1cEbwefn570c4l\n8bptxw+L9n07nyjajw8+FvT7L+f8pn9BNqzG3r3Jfeei9AKZhd8zN7SdcVQdmt4yFlsn3fAQUtSX\nnEr78KexGS3vh3d0kKhMtsVxkZNY+Nq36LpIyXk2ffqofwdGSi59y730wRJb/DtjeSe4jqPAU5ZK\na4t8FGk8VrZS5uy3D2x9dNS//3wwvPbfMt1/7j/QpstClyYONv3GIxlZpb7olFHHOtl0eoe+DsAN\n7fYNAO6emOEIIYTolDK2xW8A+DGAs81swMxuBPDnAN5tZpsAvKv9WgghxBRS2UjRvjNWFu3hLc8n\n++USCrHM0NjtVaXcE3x2mDy4c0PQb0/DR6398zOvGP398fhoqVo7fVmwrbF5q9+WSEY2glw0X+Jz\ngbTrJXZEpCL4Rr6xXGSeGDvsmmrs2+c35CSCkgmpSruVMpGiNXK21BacHHS79/EHkp/dCYeafnzX\nkQslF/FaP3NV0ebfGBAeJ/daqBdxcjv+/tno2pIoUlQIIXoMTehCCFERNKELIURFqGyBC9bNWecd\nkSifiPVf1s2DfnHhipL69al13+/eLT8u2nUL/11liyRrlLGeF4yJxx5HXqYKXETRckysjTc4kpB1\nfbLFAWkLJmepA6JMdYoUHRf1uWGGQT4HuWIXsbab3D/pxrc9eU+wbUHdX0OckXRFX2iRfLl5uGiv\nP+q3vXNWeG3uo+dM1y+/NDmm3LOvYOxstyV39QhrJkeb0u+Mi1YA+chtnlusjyLVc8+0JhjdoQsh\nREXQhC6EEBWhspILw7JALmqLo7tyfcNlHIDDfrnGS0FePgLAIpJcnh3y7zlrWrj0/dr2HxTtpbR0\n5WUrAJxUm1W0c9FtKXbREhkA5tX82LcNhxa386bPwmgMRPvop/FyVN0LUZ3T31vzgaKdqiMpyhHX\nAGVJ7LZf3F+0WR4BQhst39nxdQqE53hBPZRSchHUKernrC7a/23jpnAj2VnrixcV7cB+ibTEOEJ+\nItmUJab4mKXq+LqoOEVOXg2k2Liub5fQHboQQlQETehCCFERKhspmiJXczAmW4MwtX9a1tnM0FHD\nLgPO5zy8M0xoxUvm/7Pl4aJ9arQUPup8ZNr73+gTf8V1UzlCMLdkDLZFicBqdCxykW6pHNtxPvmt\nQ34Zv2palExMjIkhFy7vayDZghxUsTwSJH47QtdM5IoKpMjBMIEbSxeBbHHgQHrAmTmn7G8uqHfA\n/XL7LumMybquMjnfmbHU6y2DIkWFEKLH0IQuhBAVQRO6EEJUhJ7T0GNyWeZSVqZcxrmydTmDz90f\nRvCxTpdLlM+6NFvL5kSRp9eff63fB33HehzlGUUSMp08T2DiqNRYo/cdQ/2yNod0Xqr/ms36x5kc\no2NRX+iz+wXHM/5cev4Rj5WPe8NRwZCocNc0q14EbM46m+uXeg8/SwGAj5/l54eJ1qFfz0hDF0KI\nHkMTuhBCVITKRormkhIxgQRBMki8jXGZKLDaHL+cbBwIJZdUwYzavHlBP5YWWBaIbVKpKL3YIvi5\n9Q8W7Ytn+OICI96fKTqRWv7G0kfwHpJBatGxZRmDvz9/9/h131KqMRlFl/K5+9IT64p2bIlkaYqj\nWuOo3ia1Y7vo/oYf+4fPuLxo5yS2qsDFM4B0MrY46tod9MeXba/37Hw86MdJsyS5jB3doQshREXQ\nhC6EEBWhJ1wuOVmgxq6HRP5zIO82KetsCd+Uljfq553tP2vjZr8hU6eRt3EUKhBGosZyTIpYjslF\ntgZDysgnQT9yvdy/+UelxvTQYf99L5pxMNmPXRRX9l8YbKsvICkuI7dxvcg4MjYl59Xnzw/6NQ52\nLw9214iuQY4utlmU/CqXBC9Ts3O8bqqqIpeLEEL0GJrQhRCiImhCF0KIitATGnpKax4LgbY3FGVq\nS2RgizXAvpUrivbw89s7GgcTZMtLRV6ifJY5tpDFUY5PDfrIv1SxCyAsajG3Rpqqawb94jqqxxiR\nEZA0+dopPnK3uScseIAmRWxmskGmKGtZBRAWYVhERRgSFr4qER+n5iv+WUZgU81YcRm2ogIqdpJi\nwjR0M1tuZg+b2dNm9pSZfar994Vm9l0z29T+/4Lj7UsIIcTkUUZyGQbw+865cwGsAfAJMzsXwM0A\nHnLOrQbwUPu1EEKIKeK4kaLOuV0AdrXbr5rZRgDLALwPwOXtbrcCeATAZydllB1QNlKU6etfFrxu\n7PXLepZPOGkXECbu4kIQjcHQwpiSWXJJsoKCGdND+2VcF/EYZa2Jo4ykaHE0JACcN93LO5xQ6eNv\nvDzoF0g6LE1ES/DU2ONzMDyws2gHy3a2fQKB9TMnMcVJworxZCSWeEzuoP/+vSCzMLnjFBRIiSSW\nvtOXF+3hbTt8+8XdEzg6MaaHoma2EsAFAB4FsKQ92QPAiwCWTOjIhBBCjInSE7qZzQXwLQCfds4F\nt1eu9WR11KerZnaTma03s/VDUKCAEEJMFqUmdDObhtZkfptz7tvtP+82s6Xt7UsBjBpm6Zxb65y7\nyDl30TSka/AJIYQYH8fV0M3MANwCYKNz7i9o0zoANwD48/b/756UEXZISjePMxbWT/XZ44Z3DATb\nWG/l5Qdr5kA+C2CwP9KR73/m+0U7tuqxBr6HsgA2Iovp0r7Riysfaoba/XX9Fxft0oVyIzi9wQPb\nfurb29cn3/OeM9YU7ebhtJWQjzNr5q0PtlH7xQWuc6kZmMByGlkpg4+l7ztiTAlGPFt5efTnBK9r\n4gLSJZ9VsW6eKhwDlD+PYnTKpM+9FMBHAfzCzI7NNH+E1kR+h5ndCGAbgN+ZnCEKIYQoQxmXyw8A\nWGLziVVPTgghepieiBTNRlRSFKlNC/99C6yKmdqjDEs6FlnrOIIxm72RZSEqplGbGxZacEf8+Hjf\nufqdgeQSF+qgayEuUBCMMZMp8r9u9XLMhTPSWS6vXHaBH+9cLx3Fdjde0jdf89GqcVbLTmyq/B2b\nr4bZG1PnvtXZf+dYMqg8Ue1Vlq3KSmJ8zWQjckWBsi0KIUSPoQldCCEqQk9ILn1v8DFPuci0OMIy\nKI7AT/ejZafVvASRc45w1GeQQCreX92/Llswo5NEXTG5sackp3h/juSIGhU8uOvZR4J+M2waRoOl\nmNYOSQZK1GQdActo9ejYkqzGx2lEcYpEJGtMJ1JPZWEpLk6+xo4iOqdKzlUOSS5CCNFjaEIXQoiK\noAldCCEqQpnAotclQeHmQ4czPdPct+Oxos0FGbiIAwB8oH8NRqNskv/azDB6NaWB57I8BpkNSWsG\nQr2ZdfJ4fC5T1Lnxkv+sbFFsGgdb16578z9Njv3/7vhh0X5w54bkGDiiljP7AYDRMWSN35UtaFJL\nhVrkCSJgMxkgq0LOEpuFjgU/S4o18+BZENlya9PDZy7JIiZln0dF54ozmeaKU5ctgh7sOypSX7qQ\nfAfoDl0IISqCJnQhhKgIPWFbZPqW9wevs4mXEsembAGJjYPhcvTTK3+9aJdduuWSaZVNZJSy1mUt\nhyVloBFFQcgWWjb5F8sntUVhVOaXv/+Nor0ikYwMGJngrNhfLBEc9vJb/SRvVcxZDuOoWY7kHd75\nQvJ9wpMqcNEpgYxBlki+hgEEduOgAEecLI7rAneYwK5sIrlOkG1RCCF6DE3oQghREXpCcim73OOI\nUiCMKi0bicmRb/c+/kCwrUFLQ3bNxHJB38oVfgxUhzTnXsk6T3jfJDmNyP9O7gNEEZa8bOyoXmuH\nEYEsTTmSS+I87C83/baTarOKdi7yNJdkrL76jKLd2LSl1FhHJK4q67B5PRG5Q/pWnV60h5+n31aH\n350jdh3V5E26WuLhxTIiSSbBWLduC9+X+P3EEcT8289GhSdqHwBpeTCHJBchhOgxNKELIURF0IQu\nhBAVoSc0dLaZxVFguQx+yW1lIwKjfnXS1Tib39e2/yDol6oVmtXeMoU6GP7+gWaOUKfMaZEjvj8R\nZGUsWRMydw76lp1WtHMWQdb1b/vF/cl+C+r+WcjV572jaMd6KFtJ64sXB9sae/f68ZFts2zt0aqS\nzCaKdLbO2PZa9hgG1yc978kVlZkQ6HfG2UQB4P5NPuJ5H9UCjivXfnT5pWP+WGnoQgjRY2hCF0KI\nilBZySUnszAc9TnkQqvVtct8gYucbTG5LSPNcNRavNzn8d4x8OOizXa8mJ8c8WP/3Dm/HmwLZJZM\ntFx9MUWe7tkbblvoE4M1X/H1Nzu2SJa0gSb3nbGYMnGU59c2rCvas80vn+MEa5wILa4bGhxDjgLs\nAdtiLD85Ond8LOIIXZZFspHRid9tLAHy65yskrLY1s8+M+jnBnbRC/87ja9NTiR3Ui1MupUq2hIj\n26IQQojjogldCCEqQmUllxTxUus/3n9H0b50ZvjvG0sw02h5HtQaRZirObf8K5uQK8h7Tnm62R0A\nhO6Q+zb8XXJ/R91Q0eZlYW7pl424yzhPykasBq6cTE1WXuKzuyQmiKKlpfqISFZOoDTXu4nu+se/\nD7rxcRoYPhhs60+4kMrmya8qfO4tzl/OckxKskIo1XCkaEzppFmZaOAU7DpLOc4A4FAzHN9L9Jqv\nkYm4LiS5CCFEj3HcCd3MZprZT83sZ2b2lJn9cfvvq8zsUTPbbGbfNLPpx9uXEEKIyaPMHfpRAFc4\n534NwPkArjKzNQA+D+ALzrkzAewHcOPkDVMIIcTxOG5NUdcS2Y+JiNPa/zkAVwD4SPvvtwL4HIAv\nT/wQOyShnd3x0Nczbwojv1g3Z1sg2/uAMLoxpy+zbs72rNhaV7ZoAn8u6/pxpChHz3312e8V7Vyh\njlj3+901H/Djy0Tz1Zd4zXv4BZ9RMR4TW9IoCeWIWqFBVCZlbBzeHerpTdLKczbVwJJG5+O9y94W\ndONjs6Qe2kVTz1Z6kbKRsnw8t9MziX974XVBv9xzEoZ/P3x926zwXLHltHb+uUW7+eTTyfHtafj9\nxVbmGvy22ZFtkV+zvj63Fs4rk0kpDd3M6mb2JIA9AL4L4DkAB5xzx55MDABYlnq/EEKIyafUhO6c\nazjnzgfQD+BiAG8q+wFmdpOZrTez9UPI3DkJIYQYF8eVXBjn3AEzexjA2wGcbGZ97bv0fgCjrrec\nc2sBrAVatsVxjrc8VECiNtdboWZFz2650ETMLloafu4cb7d0R8MEUpyUiOWSuH7pXT+5u2izlfAD\nK8JkPWWjXPlzXYPqKkbv4YP+L896l/97HKFKr2M55k++/52ifeGM9PPvK0+jF2RNjMfE0opR/dLY\nmsnH4puP3VW042UsL3FnmL+s4/MbWDV5qR4V9OB+OWnqH0iZumxm95bWU0Wc0I1llpyN9qlBX4Dk\nvOne0hdLLHy+mfj6YZklsApHtmG2395+zy1FO466Zonx1Hoo+6XYFdlZ2eLI1yAXX5lsyrhcFpvZ\nye32LADvBrARwMMAPtjudgOAu0ffgxBCiG5Q5g59KYBbzayO1j8Adzjn7jGzpwHcbmZ/CmADgFty\nOxFCCDG5lHG5/BzABaP8fQtaeroQQogTgDFp6K8rKNNdUKwg0lRZO4t1WdbEsla4BPc+ek/0F6/T\n7m54Xe2eHT8NenGWx1zocifJ+8u+J05vwN5C1pu/vvX/Bd1SevN7zkxngIzDv5m7t/gw7D6ktXu2\njO1v+Ax5XNACAP6Kwro//k+uL9px0eDYPsmw7e6ymf4aiS1ufCy44MH1UYGDVGGIuEAxF0UJwuzr\n4TXN+8iF2ScLlWRC5HPXDz8LicPiT+8b/bPi6yWVjiJXjCX3jIN17pNq/lzFunYqk2ncr062xVxa\ngGCe6d6TQ4X+CyFEVdCELoQQFaGykgtnLPzOL9lCFWaB46x62yMbUipdfa64Ai+F42hLlnRSGfvi\n/acKN0w6UXGGVN3Pj64Os2cG9knKtvjgCz8K+qVkkatWXRL0u47tomSr/NOnHgn6sZWS9xcvmVfQ\ncWeZpe/05UG/4W07ijZbTAFgUW106WfrcHi+P3m6l1ZYFoglgueG/HX3iTOvKNpxARImiLzN1HjN\nyVnhDlkiiKIj42IVwQf48/0nj3yraMdRlCxH5bJ8BtJKzY/JovqdbE/MWUzn1fxxapBsGEssLBEd\npPP9sfPeE/RrklR453Oh3MiWaJZccoVpJhrdoQshREXQhC6EEBWhspLLV372t0X7UNP/uzWjHgop\n7DZZEckgVy7zbs3aHL/sjGUQllke2OYdKw0XLjtZ0ok/i+H9s9OBXQ7dhmUWXhbb9PA7NhNjvPq8\nd4T7o6RJvEz+5FM/C/pdMzslO4SfyxLOSyQDvHFa+jizVMESCxDKDHGtyCGMXiuUJRYgTLqWkxn4\n+/P1E0s91731mqLN54PrZsb0nbGyaA9veT7cyG4WkkT4egaA+zd7uSyWsGZSROQMkhxYRgLC88Df\nt3Txh6i2Jydqax4sJyuxDLKnEb7nY299f9EOo1fD6zkopBLxKzo2HG3aSQ3RTtEduhBCVARN6EII\nURE0oQshREWorIae06gZLl4Q62pBMYSM/YvteRsHvdZ3zvTQ7sVj4gi2G6Mk/4DX8LpZxJuJdVQu\n+hsciyhykO2iTdI9WTMHgPrqM4r2e1Z7vbE2Lzxv/4OeJ+S0V7YqLqDEic8OhedtCUVVPrhzQ9G+\n5pJrg37DOwaQgu2JM82fnziasbHvV0U7KHzywq6gX0pjvW/nE8Hru5+4v2gfbHr73IeWh1G4wWeR\nbp6LtuTzdt9TDwf9yj77Yc0/9+zi8aP+93LhjHSGypxOPryLiqckMjQCoVWYj3P8HnfU/+b42Qef\nQyB8dhFnbp1LNku2QeZszhON7tCFEKIiaEIXQoiKUFnJheHl+ZEoCm4R2Ys+9pZw2d2qfd0iZx+s\nn7KwaMcyCxPKMekk/yeCVZFlpNFeF0RRiu6wt25xAi6WAQBgeNOWUfcRS1tcoCBYMkeSENv92DJ3\nVrT056UwSwmxxBInxmLOmjZ64q57ngujYa853ScjDWSWSEbj64dtoFcve2tyDLnI032NHxbtj579\nbr+B7JxAWIDlb3787aL9XHSuWT75+WAodb2h7n9PbNWLpS4+ZhzVG8uc/F34fI+wC1LCPb7O4khj\nN8jfheSxSNpzlHCOZRa+/gDgge3r6VV4P8yWztkkx3Qz2lt36EIIURE0oQshREWorOSSepK+YzCM\nOJtmfpkUOzGYZiYf+hef8FGpRyk6NI4wZDmGl/ux+yCu9TkVxPnAgzE1aem6Mqyb2iAphWWR5iuv\nBv1SOcBHug9Itlm5omgPP7896MfL89q8eUX79o1hbUtOlLTU/DI7l1N7IEraxonVOOlUkMceQN/S\nRX685MqIieuopuBzctWKi4p2fL3wd7lrk08g9cvBUOp5y3T//Ydo04q+dL3Nmy+8OtjW2P9y0eaE\nYQ9sfTTox64uziM+N/qNsGyRS1LHydQ4yndE7drEdRa7V5icyyU1VgAYouRf02rpa2syI0d1hy6E\nEBVBE7oQQlQETehCCFERKquhszXqJ0e8zrlmZtpWGGtdrFOiMXqGPSBtY9saZZxbRfavf3PWu4o2\nF4UAAMc1IRMa4GRTujDC3lBjTOmeSdtjRC3KHMiReaybx7bFGhVAYKtnrrgAR1vOsPCnwAUaZmYK\nSNSoxmRscWPdnKMeYz2dLZJc1CI+ZrX5/tnA8K70+WGNdt3Ox4o2/yaA8NnAQvq+0+jZAhB+x5ze\n747630hcX/VUiuQNn2+FYwr15bTdj3Vztn3GNlrWwDmDZvNI9EyMCrrwe1hPB8pHzbJt87R6Pdlv\notEduhBCVARN6EIIUREqK7lwROCamX5ZF8sgi+v+EMRLzTAqzBMnhtpHcgxHnvK+gXAZGsgnmSU9\n11XsJvFSM1i6kn2uceBlhPjXObkoVSSDJRYgTBrFttJYjmiUlHRYZmD7IV8vQJhois9pvO2I85bB\n+HoJzje8rMaFKoB08ZCovkUg1ZSNJr7uPC/txefqOwM+upYlpn1R9OZM+l3ECcO4aATX7HylGUoa\nnDyNZZZdkSU0kNJq/ncRRnwCRjJG82X//XOWX04WN9Ie689VcH1HtsWyMktKhp1sSs8WZlY3sw1m\ndk/79Soze9TMNpvZN81s9Mq5QgghusJYbv8+BWAjvf48gC84585EK+nJjRM5MCGEEGOjlORiZv0A\nrgHwZwD+g5kZgCsAfKTd5VYAnwPw5UkYY0dc1+8TI7F7JZZBOF9yDl5mx++Z5ZpxdwDAS81w+XfT\nGy8r2tbn33MiRIbG5CLkyjpgcq6cYFumXy56twxxVN63Bn4yar+jLjwH8y19XXAEMEdf/qc3hzVF\n+TjxNfiVx74d9ONlPMt5H1z5G8kxBDJLLNlxHn/KIx5/99mJaz++mvl6jyUSjvpk+WXLcPg7uzBh\n9PjYiug71sgdw0aZKKFZ8JuhqMygDaDGefzpOosjSgOHUkmZk+vYAsDpfaOLFHFtWCYV8QpE8lM5\nRbH0HfoXAfwB/Lk+BcAB54pfwQCAZSX3JYQQYhI47oRuZtcC2OOce7yTDzCzm8xsvZmtH0I6H4oQ\nQojxUUZyuRTAe83sagAzAcwH8JcATjazvvZdej+AnaO92Tm3FsBaAJhvC6emnpoQQvQANpaalWZ2\nOYDPOOeuNbO/AfAt59ztZvYVAD93zv2v3Pvn20J3ib1zXAMuC2fcY8tTrJ2xrvj0UKi/XUjZ6H7r\nvHcU7dj+xfoo2yJXRcUVgoyAbNubE0avls2+J8bHX23/QdE+KdJeOcI01kr3UrZJtqeNyKJXG104\nrkXnu0nFGvhaei6y2HKhCbYIxoUwuF5rc6uPrs09q7l9hy/OwRZDILR0sr0RCDMO8jGLj0VKK44L\nVzTp2l+75ZGiHdsFr+ynzJbNdBR38HyB5roguhTlf3Ns573v5w8F2/j5x+6GP9ZxfdVr3uYzVg7v\nfGH0sUb0xhfuAAAJGklEQVTj/Z6783Hn3EU4DuMxOX8WrQekm9HS1G8Zx76EEEKMkzEFFjnnHgHw\nSLu9BcDFuf5CCCG6x5gkl/HSTcklgJa+nIQfAIyWObUoOjJcDvnFzIMD4fPhlC2pER3b337jbxbt\nWPqJBlU0+2hJ2s3ahFWFEzSBog3vf+b7QT+OlnyhEUonf3jpbxdtvkbiZTwXAsnZL+vnrPb9Nm7y\nG+IleOYaZFIJpGIZpE6J0FhGzBX7iKNIr1/urZr8vkZk5X1y0EsQ//liHyl72xN3B/3mk0WSbZA5\n+SmoNZsrkEJ1bYPfdvS+Gu+jFp6D//mze4v2wsjeGEtVZcgVu+BEdw/s+tKkSy5CCCFOIDShCyFE\nRais5JJKXpRbkuUIcikfCl0PqSXqiCXu4sV+THv3+n1HNUVthY/Rajz7XKnxiYklcGVsHwi21WZ5\nN0d8LZTaN+VGB8KkW5wYivOfx/2y+z9jZdH+7PfuKtqXlQuKHhMpt02cGz6V6C6GHUWzaz7KM86v\nXjbCO1V7NZ4H+PfN8ljcj2W0B7b9NNj0JM0l50z398q7G+Ecw46qD53l58MR55sk1m64XIQQQpxA\naEIXQoiKoAldCCEqQmU1dKZ+5qqi3di8NdyWKRTAOmAuyi6nr6fo6/c6+fDAqFkTxCQT1D+NCmuw\nRTAXicgRyRzx2dpIdscOohlj+Frl6OdsrVnad33hgmATR0dy4Yo9UWQsZ1TkIg5AGCnL+nfOwhcX\nmWHi6OoUbJ9cQBGqh12YlvAD/WuKNuvhZZ+dxfBzsG9sWBdsS9WvjS2cbMfM2Rb52drfHf66NHQh\nhOglNKELIURFqGxN0aAuYCSzMCyz8PIZGGUJ3YaXvvE+ctFoPKZmpoAEJ7avcTQfWR1Fh5AEkYu8\nrc3wljnUQ4scF64IrpEospOjkt2gX3YH0aoAmoe9ZFJfSOc7ThhFUYsss+Sux1TtzRiWAZZm6mbm\namWyzXB7XAij7uWIJfXp9J4w2deVyy4o2g/u3FC042hsrvPKkgZLLADQt7y/aA/vCO2nASSP9Z3m\nbaWxHMq/wVhi4eIfrzp/rj4ZFypJyGq5Or5l0R26EEJUBE3oQghRETShCyFERaisbZEtPzlbVyoc\nHziOJS35wV6LyxUyCPT0kkWXxfhJWkwzxQVG7IPOnRskbTfKzBfQ8LbFskXB4/D5ZGHkjCUy2Ec9\nzBqZsu59ZdsPgtdsJYy18UW1tB5ehti2x2k0eNvXtodjOtD096LnTJ896ntylNWr6wsiq2cma2bK\n5sxZLQHANbzmz3NCUBQaAOjZwHeHbpdtUQgheglN6EIIUREqa1sMbF1UJKIRRQQGWQ/nhJaslEQy\nYsncwXK6NGWjDUUpUpG8tbmhVc8d9rUy43PKElnZKGHul7tGWAooa1uLl/SNV7wsEnxW9LksEdQX\nLyrauWjNuLYnU1bu4PqluWIa4bbwcxeSjZFti7n98fjiY5uSaJsH03LoiOMe1Ro+3t+BUGZxw1Gh\nnA7kcN2hCyFERdCELoQQFaGyLhchepaSyb5Scl7srrmXCjkMI5T9ZtjoRSh+a+Xbkx/7t8//uGhP\ns9B5wxGhvL+4oAVvO9j0bp3rz7826PcNqgHKkZ2HmmESr+v6fb17drbErpa4GA3DUk0gpQwNjtZ9\nTKjAhRBC9Bia0IUQoiJoQhdCiIpQWduiEGLiYc0cAPZQoYlTKQPiFzY9ktzHNPMWTn5/vI8+eH39\nqcHDQb/PvOkdRduoaHdjf5ih8iNvu65oc5Ht2N7Ir9e95sf3lbeHzwI6yYA4lijk8aI7dCGEqAia\n0IUQoiJ01bZoZnsBvAZgX9c+9MRmEXQsjqFj4dGx8OhYtDjdObf4eJ26OqEDgJmtL+On7AV0LDw6\nFh4dC4+OxdiQ5CKEEBVBE7oQQlSEqZjQ107BZ56o6Fh4dCw8OhYeHYsx0HUNXQghxOQgyUUIISpC\nVyd0M7vKzJ4xs81mdnM3P3uqMbPlZvawmT1tZk+Z2afaf19oZt81s03t/y843r6qgpnVzWyDmd3T\nfr3KzB5tXx/fNLOxF6l8HWJmJ5vZnWb2j2a20cze3qvXhZn9Xvv38Usz+4aZzezV66ITujahm1kd\nwJcAvAfAuQA+bGbnduvzTwCGAfy+c+5cAGsAfKL9/W8G8JBzbjWAh9qve4VPAdhIrz8P4AvOuTMB\n7Adw45SMqvv8JYAHnHNvAvBraB2TnrsuzGwZgH8P4CLn3JsB1AF8CL17XYyZbt6hXwxgs3Nui3Nu\nEMDtAN7Xxc+fUpxzu5xzT7Tbr6L1o12G1jG4td3tVgDvn5oRdhcz6wdwDYC/br82AFcAuLPdpSeO\nhZmdBOAyALcAgHNu0Dl3AD16XaCVX2qWmfUBmA1gF3rwuuiUbk7oywDsoNcD7b/1HGa2EsAFAB4F\nsMQ5t6u96UUAS6ZoWN3miwD+AMCxgpCnADjgnDtW+LJXro9VAPYC+N9t+emvzWwOevC6cM7tBPDf\nAWxHayJ/GcDj6M3roiP0ULTLmNlcAN8C8Gnn3Cu8zbUsR5W3HZnZtQD2OOcen+qxnAD0AXgrgC87\n5y5AKzVGIK/00HWxAK2VySoApwGYA+CqKR3U64xuTug7ASyn1/3tv/UMZjYNrcn8Nufct9t/3m1m\nS9vblwLYM1Xj6yKXAnivmT2PlvR2BVo68sntpTbQO9fHAIAB59yj7dd3ojXB9+J18S4AW51ze51z\nQwC+jda10ovXRUd0c0J/DMDq9hPr6Wg97FjXxc+fUtoa8S0ANjrn/oI2rQNwQ7t9A4C7uz22buOc\n+0PnXL9zbiVa18HfO+euB/AwgA+2u/XKsXgRwA4zO7v9p3cCeBo9eF2gJbWsMbPZ7d/LsWPRc9dF\np3Q72+LVaGmndQBfdc79Wdc+fIoxs98A8H0Av4DXjf8ILR39DgArAGwD8DvOuZdG3UkFMbPLAXzG\nOXetmZ2B1h37QgAbAPwz59zR3PurgJmdj9bD4ekAtgD4F2jdbPXcdWFmfwzgd9FyhW0A8K/Q0sx7\n7rroBEWKCiFERdBDUSGEqAia0IUQoiJoQhdCiIqgCV0IISqCJnQhhKgImtCFEKIiaEIXQoiKoAld\nCCEqwv8HfNTyEe39zSYAAAAASUVORK5CYII=\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7f1614040490>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "plt.imshow(get_image_batch(real_generator)[2,:,:,0])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# MNIST for Captcha\n",
    "\n",
    "Now we finish the puzzle by building an MNIST like convnet to predict captcha labels."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 382,
   "metadata": {},
   "outputs": [],
   "source": [
    "n_class = len(alphanumeric)\n",
    "\n",
    "def mnist_raw_generator(batch_size=128):\n",
    "    X = np.zeros((batch_size, HEIGHT, WIDTH, 1), dtype=np.uint8)\n",
    "    y = [np.zeros((batch_size, n_class), dtype=np.uint8) for _ in range(4)] # 4 chars\n",
    "    while True:\n",
    "        for i in range(batch_size):\n",
    "            im, random_str = gen_one()\n",
    "            X[i] = im\n",
    "            for j, ch in enumerate(random_str):\n",
    "                y[j][i, :] = 0\n",
    "                y[j][i, alphanumeric.find(ch)] = 1\n",
    "        yield np.array(X), y\n",
    "\n",
    "def mnist_generator(batch_size=128):\n",
    "    X = np.zeros((batch_size, HEIGHT, WIDTH, 1), dtype=np.uint8)\n",
    "    y = [np.zeros((batch_size, n_class), dtype=np.uint8) for _ in range(4)] # 4 chars\n",
    "    while True:\n",
    "        for i in range(batch_size):\n",
    "            im, random_str = gen_one()\n",
    "            X[i] = im\n",
    "            for j, ch in enumerate(random_str):\n",
    "                y[j][i, :] = 0\n",
    "                y[j][i, alphanumeric.find(ch)] = 1\n",
    "        yield refiner_model.predict(np.array(X)), y\n",
    "\n",
    "mg = mnist_generator().next()\n",
    "\n",
    "# plt.imshow(mg[0][0,:,:,0]) # sanity check"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 383,
   "metadata": {
    "scrolled": true
   },
   "outputs": [],
   "source": [
    "from keras.layers import *\n",
    "\n",
    "input_tensor = Input((HEIGHT, WIDTH, 1))\n",
    "x = input_tensor\n",
    "x = Conv2D(32, kernel_size=(3, 3),\n",
    "                 activation='relu')(x)\n",
    "for _ in range(4):\n",
    "    x = Conv2D(128, (3, 3), activation='relu')(x)\n",
    "    x = MaxPooling2D(pool_size=(2, 2))(x)\n",
    "x = Dropout(0.25)(x)\n",
    "x = Flatten()(x)\n",
    "x = Dense(128, activation='relu')(x)\n",
    "x = Dropout(0.5)(x)\n",
    "x = [Dense(n_class, activation='softmax', name='c%d'%(i+1))(x) for i in range(4)]\n",
    "\n",
    "model = models.Model(inputs=input_tensor, outputs=x)\n",
    "model.compile(loss='categorical_crossentropy',\n",
    "              optimizer='rmsprop',\n",
    "              metrics=['accuracy'])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 417,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 1/20\n",
      " 341/1000 [=========>....................] - ETA: 376s - loss: 2.7648 - c1_loss: 0.6493 - c2_loss: 0.6757 - c3_loss: 0.6681 - c4_loss: 0.7717 - c1_acc: 0.8199 - c2_acc: 0.8185 - c3_acc: 0.8197 - c4_acc: 0.7820"
     ]
    },
    {
     "ename": "KeyboardInterrupt",
     "evalue": "",
     "output_type": "error",
     "traceback": [
      "\u001b[0;31m---------------------------------------------------------------------------\u001b[0m",
      "\u001b[0;31mKeyboardInterrupt\u001b[0m                         Traceback (most recent call last)",
      "\u001b[0;32m<ipython-input-417-ad15887877d5>\u001b[0m in \u001b[0;36m<module>\u001b[0;34m()\u001b[0m\n\u001b[1;32m      1\u001b[0m \u001b[0;32mfrom\u001b[0m \u001b[0mkeras\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mcallbacks\u001b[0m \u001b[0;32mimport\u001b[0m \u001b[0mHistory\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m      2\u001b[0m \u001b[0mhistory\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mHistory\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m----> 3\u001b[0;31m \u001b[0mmodel\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mfit_generator\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mmnist_generator\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0msteps_per_epoch\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0;36m1000\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mepochs\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0;36m20\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mcallbacks\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0;34m[\u001b[0m\u001b[0mhistory\u001b[0m\u001b[0;34m]\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m",
      "\u001b[0;32m/usr/local/lib/python2.7/dist-packages/keras/legacy/interfaces.pyc\u001b[0m in \u001b[0;36mwrapper\u001b[0;34m(*args, **kwargs)\u001b[0m\n\u001b[1;32m     86\u001b[0m                 warnings.warn('Update your `' + object_name +\n\u001b[1;32m     87\u001b[0m                               '` call to the Keras 2 API: ' + signature, stacklevel=2)\n\u001b[0;32m---> 88\u001b[0;31m             \u001b[0;32mreturn\u001b[0m \u001b[0mfunc\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m*\u001b[0m\u001b[0margs\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m**\u001b[0m\u001b[0mkwargs\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m     89\u001b[0m         \u001b[0mwrapper\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0m_legacy_support_signature\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0minspect\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mgetargspec\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mfunc\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     90\u001b[0m         \u001b[0;32mreturn\u001b[0m \u001b[0mwrapper\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
      "\u001b[0;32m/usr/local/lib/python2.7/dist-packages/keras/engine/training.pyc\u001b[0m in \u001b[0;36mfit_generator\u001b[0;34m(self, generator, steps_per_epoch, epochs, verbose, callbacks, validation_data, validation_steps, class_weight, max_q_size, workers, pickle_safe, initial_epoch)\u001b[0m\n\u001b[1;32m   1869\u001b[0m                             \u001b[0;32mbreak\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m   1870\u001b[0m                         \u001b[0;32melse\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m-> 1871\u001b[0;31m                             \u001b[0mtime\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0msleep\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mwait_time\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m   1872\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m   1873\u001b[0m                     \u001b[0;32mif\u001b[0m \u001b[0;32mnot\u001b[0m \u001b[0mhasattr\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mgenerator_output\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m'__len__'\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
      "\u001b[0;31mKeyboardInterrupt\u001b[0m: "
     ]
    }
   ],
   "source": [
    "from keras.callbacks import History\n",
    "history = History()\n",
    "model.fit_generator(mnist_generator(), steps_per_epoch=1000, epochs=20, callbacks=[history])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Let's test the trained model\n",
    "\n",
    "## Synthetic"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 442,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(-0.5, 99.5, 49.5, -0.5)"
      ]
     },
     "execution_count": 442,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAXQAAADlCAYAAABDLgwBAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nQaADQBIEOgAkQaADQBIEOgAk8T8omx+G+3g1UQAAAABJRU5ErkJggg==\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7f16046d9f90>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "def decode(y):\n",
    "    y = np.argmax(np.array(y), axis=2)[:,0]\n",
    "    return ''.join([alphanumeric[x] for x in y])\n",
    "\n",
    "X, y = next(mnist_generator(1))\n",
    "plt.title('real: %s\\npred:%s'%(decode(y), decode(y_pred)))\n",
    "plt.imshow(X[0, :, :, 0], cmap='gray')\n",
    "plt.axis('off')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Real"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 441,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(-0.5, 99.5, 49.5, -0.5)"
      ]
     },
     "execution_count": 441,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAXQAAADWCAYAAADBwHkCAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nQQcAJ/4PrJKOcE9V/HcAAAAASUVORK5CYII=\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7f160475bf90>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "\n",
    "X = next(real_generator)\n",
    "X = refiner_model.predict(X)\n",
    "y_pred = model.predict(X)\n",
    "plt.title('pred:%s'%(decode(y_pred)))\n",
    "plt.imshow(X[0,:,:,0], cmap='gray')\n",
    "plt.axis('off')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 2",
   "language": "python",
   "name": "python2"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 2
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython2",
   "version": "2.7.12"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}