14.4 Icon RNN.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "from keras.layers import Input, Dense, Conv2D, MaxPooling2D, UpSampling2D, Reshape, Concatenate, LSTM, Dropout\n",
    "from keras.models import Model\n",
    "from keras.optimizers import RMSprop\n",
    "from keras.layers.wrappers import TimeDistributed\n",
    "from keras.preprocessing.image import load_img, img_to_array, ImageDataGenerator\n",
    "from keras import backend as K\n",
    "import random\n",
    "import matplotlib.pyplot as plt\n",
    "import json\n",
    "import keras.callbacks\n",
    "import glob\n",
    "from sklearn.model_selection import train_test_split\n",
    "import numpy as np\n",
    "from io import BytesIO\n",
    "import PIL\n",
    "from IPython.display import clear_output, Image, display, HTML"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "((1290142,), (140030,))"
      ]
     },
     "execution_count": 2,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "def encode_icon(img, icon_size):\n",
    "    size_last_x = 0\n",
    "    encoded = []\n",
    "    for y in range(icon_size):\n",
    "        for x in range(icon_size):\n",
    "            p = img.getpixel((x, y))\n",
    "            if img.getpixel((x, y)) < 192:\n",
    "                encoded.append(x)\n",
    "                size_last_x = len(encoded)\n",
    "        encoded.append(icon_size)\n",
    "    return encoded[:size_last_x]\n",
    "\n",
    "\n",
    "def decode_icon(encoded, icon_size, rows=None, stop_token=None):\n",
    "    y = 0\n",
    "    for idx in encoded:\n",
    "        if idx == icon_size:\n",
    "            y += 1\n",
    "        elif idx == icon_size + 1:\n",
    "            break\n",
    "        else:\n",
    "            x = idx\n",
    "            yield x, y\n",
    "            \n",
    "def make_array(icons):\n",
    "    res = []\n",
    "    for icon in icons:\n",
    "        res.extend(icon)\n",
    "        res.append(33)\n",
    "    return np.asarray(res)\n",
    "\n",
    "def load_icons(train_size=0.90):\n",
    "    icon_index = json.load(open('icons/index.json'))\n",
    "    x = []\n",
    "    img_rows, img_cols = 32, 32\n",
    "    for icon in icon_index:\n",
    "        if icon['name'].endswith('_filled'):\n",
    "            continue\n",
    "        img_path = 'icons/png32/%s.png' % icon['name']\n",
    "        img = PIL.Image.open(img_path)\n",
    "        if img.mode != 'L':\n",
    "            img = img.convert('L')\n",
    "        x.append(encode_icon(img, 32))\n",
    "    x_train, x_val = train_test_split(x, train_size=train_size)\n",
    "    x_train = make_array(x_train)\n",
    "    x_val = make_array(x_val)\n",
    "    return x_train, x_val\n",
    "\n",
    "x_train, x_test = load_icons()\n",
    "x_train.shape, x_test.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "input (InputLayer)           (None, None, 34)          0         \n",
      "_________________________________________________________________\n",
      "lstm_layer_1 (LSTM)          (None, None, 256)         297984    \n",
      "_________________________________________________________________\n",
      "lstm_layer_2 (LSTM)          (None, None, 256)         525312    \n",
      "_________________________________________________________________\n",
      "time_distributed_1 (TimeDist (None, None, 34)          8738      \n",
      "=================================================================\n",
      "Total params: 832,034\n",
      "Trainable params: 832,034\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "def icon_rnn_model(num_chars, num_layers, num_nodes=512, dropout=0.1):\n",
    "    input = Input(shape=(None, num_chars), name='input')\n",
    "    prev = input\n",
    "    for i in range(num_layers):\n",
    "        lstm = LSTM(num_nodes, return_sequences=True, name='lstm_layer_%d' % (i + 1))(prev)\n",
    "        if dropout:\n",
    "            prev = Dropout(dropout)(lstm)\n",
    "        else:\n",
    "            prev = lstm\n",
    "    dense = TimeDistributed(Dense(num_chars, name='dense', activation='softmax'))(prev)\n",
    "    model = Model(inputs=[input], outputs=[dense])\n",
    "    optimizer = RMSprop(lr=0.01)\n",
    "    model.compile(loss='categorical_crossentropy', optimizer=optimizer, metrics=['accuracy'])\n",
    "    return model\n",
    "\n",
    "model = icon_rnn_model(34, num_layers=2, num_nodes=256, dropout=0)\n",
    "model.summary()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(array([[[ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  1.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],\n",
       "         [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  1.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],\n",
       "         [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  1.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],\n",
       "         [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           1.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],\n",
       "         [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  1.,  0.],\n",
       "         [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  1.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],\n",
       "         [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  1.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],\n",
       "         [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  1.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],\n",
       "         [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  1.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],\n",
       "         [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  1.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],\n",
       "         [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  1.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],\n",
       "         [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  1.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],\n",
       "         [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           1.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],\n",
       "         [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  1.,  0.],\n",
       "         [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  1.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],\n",
       "         [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  1.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],\n",
       "         [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           1.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],\n",
       "         [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  1.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],\n",
       "         [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  1.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],\n",
       "         [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  1.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.]]]),\n",
       " array([[[ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  1.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],\n",
       "         [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  1.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],\n",
       "         [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           1.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],\n",
       "         [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  1.,  0.],\n",
       "         [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  1.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],\n",
       "         [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  1.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],\n",
       "         [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  1.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],\n",
       "         [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  1.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],\n",
       "         [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  1.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],\n",
       "         [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  1.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],\n",
       "         [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  1.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],\n",
       "         [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           1.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],\n",
       "         [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  1.,  0.],\n",
       "         [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  1.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],\n",
       "         [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  1.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],\n",
       "         [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           1.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],\n",
       "         [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  1.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],\n",
       "         [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  1.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],\n",
       "         [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  1.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],\n",
       "         [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  1.,  0.,  0.,  0.,  0.,  0.,\n",
       "           0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.]]]))"
      ]
     },
     "execution_count": 4,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "CHUNK_SIZE = 160\n",
    "\n",
    "def data_generator(corpus, icon_size, batch_size, chunk_size):\n",
    "    X = np.zeros((batch_size, chunk_size, icon_size + 2))\n",
    "    y = np.zeros((batch_size, chunk_size, icon_size + 2))\n",
    "    while True:\n",
    "        for row in range(batch_size):\n",
    "            idx = random.randrange(len(corpus) - chunk_size - 1)\n",
    "            chunk = np.zeros((chunk_size + 1, icon_size + 2))\n",
    "            for i in range(chunk_size + 1):\n",
    "                chunk[i, corpus[idx + i]] = 1\n",
    "            X[row, :, :] = chunk[:chunk_size]\n",
    "            y[row, :, :] = chunk[1:]\n",
    "        yield X, y\n",
    "\n",
    "next(data_generator(x_train, 32, 1, chunk_size=20))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 1/100\n",
      "18s - loss: 3.6500 - acc: 0.0710 - val_loss: 3.5513 - val_acc: 0.0438\n",
      "Epoch 2/100\n",
      "17s - loss: 3.0590 - acc: 0.1909 - val_loss: 2.1740 - val_acc: 0.4109\n",
      "Epoch 3/100\n",
      "17s - loss: 1.3228 - acc: 0.6636 - val_loss: 1.1567 - val_acc: 0.7170\n",
      "Epoch 4/100\n",
      "17s - loss: 0.9252 - acc: 0.7492 - val_loss: 0.9690 - val_acc: 0.7545\n",
      "Epoch 5/100\n",
      "17s - loss: 0.7579 - acc: 0.7835 - val_loss: 0.9126 - val_acc: 0.7684\n",
      "Epoch 6/100\n",
      "17s - loss: 0.6835 - acc: 0.8021 - val_loss: 0.8794 - val_acc: 0.7841\n",
      "Epoch 7/100\n",
      "17s - loss: 0.6432 - acc: 0.8139 - val_loss: 0.8219 - val_acc: 0.7964\n",
      "Epoch 8/100\n",
      "17s - loss: 0.5926 - acc: 0.8247 - val_loss: 0.8131 - val_acc: 0.7989\n",
      "Epoch 9/100\n",
      "17s - loss: 0.5480 - acc: 0.8350 - val_loss: 0.8180 - val_acc: 0.8006\n",
      "Epoch 10/100\n",
      "17s - loss: 0.5352 - acc: 0.8413 - val_loss: 0.8576 - val_acc: 0.7984\n",
      "Epoch 11/100\n",
      "17s - loss: 0.5153 - acc: 0.8464 - val_loss: 0.8278 - val_acc: 0.8015\n",
      "Epoch 12/100\n",
      "17s - loss: 0.5114 - acc: 0.8499 - val_loss: 0.8119 - val_acc: 0.8071\n",
      "Epoch 13/100\n",
      "17s - loss: 0.4793 - acc: 0.8563 - val_loss: 0.8380 - val_acc: 0.8040\n",
      "Epoch 14/100\n",
      "18s - loss: 0.4661 - acc: 0.8613 - val_loss: 0.8506 - val_acc: 0.8007\n",
      "Epoch 15/100\n",
      "17s - loss: 0.4605 - acc: 0.8648 - val_loss: 0.8366 - val_acc: 0.8056\n",
      "Epoch 16/100\n",
      "18s - loss: 0.4636 - acc: 0.8655 - val_loss: 0.8438 - val_acc: 0.8032\n",
      "Epoch 17/100\n",
      "17s - loss: 0.4395 - acc: 0.8698 - val_loss: 0.8754 - val_acc: 0.8015\n",
      "Epoch 18/100\n",
      "17s - loss: 0.4179 - acc: 0.8732 - val_loss: 0.8906 - val_acc: 0.8000\n",
      "Epoch 19/100\n",
      "17s - loss: 0.4019 - acc: 0.8769 - val_loss: 0.9157 - val_acc: 0.7976\n",
      "Epoch 20/100\n",
      "18s - loss: 0.4001 - acc: 0.8788 - val_loss: 1.0108 - val_acc: 0.7921\n",
      "Epoch 21/100\n",
      "17s - loss: 0.3963 - acc: 0.8807 - val_loss: 0.8964 - val_acc: 0.8005\n",
      "Epoch 22/100\n",
      "17s - loss: 0.4129 - acc: 0.8811 - val_loss: 0.9044 - val_acc: 0.8024\n",
      "Epoch 23/100\n",
      "17s - loss: 0.3784 - acc: 0.8851 - val_loss: 0.9083 - val_acc: 0.8034\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<keras.callbacks.History at 0x7fea57193780>"
      ]
     },
     "execution_count": 6,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "early = keras.callbacks.EarlyStopping(monitor='loss',\n",
    "                              min_delta=0.03,\n",
    "                              patience=3,\n",
    "                              verbose=0, mode='auto')\n",
    "\n",
    "BATCH_SIZE = 256\n",
    "model.fit_generator(\n",
    "    data_generator(x_train, 32, batch_size=BATCH_SIZE, chunk_size=CHUNK_SIZE),\n",
    "    validation_data=data_generator(x_test, 32, batch_size=BATCH_SIZE, chunk_size=CHUNK_SIZE),\n",
    "    epochs=100,\n",
    "    callbacks=[early,],\n",
    "    steps_per_epoch=2 * len(x_train) / (BATCH_SIZE * CHUNK_SIZE),\n",
    "    validation_steps=2 * len(x_test) / (BATCH_SIZE * CHUNK_SIZE),\n",
    "    verbose=2\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(1290142,)"
      ]
     },
     "execution_count": 7,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "x_train.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "def generate_icons(model, num=2, diversity=1.0):\n",
    "    start_index = random.randint(0, len(x_test) - CHUNK_SIZE - 1)\n",
    "    generated = x_test[start_index: start_index + CHUNK_SIZE]\n",
    "    sofar = None\n",
    "    while num > 0:\n",
    "        x = np.zeros((1, len(generated), 34))\n",
    "        for t, char in enumerate(generated):\n",
    "            x[0, t, char] = 1.\n",
    "        preds = model.predict(x, verbose=0)[0]\n",
    "        if diversity is None:\n",
    "            next_index = np.argmax(preds[len(generated) - 1])\n",
    "        else:\n",
    "            preds = np.asarray(preds[len(generated) - 1]).astype('float64')\n",
    "            preds = np.log(preds) / diversity\n",
    "            exp_preds = np.exp(preds)\n",
    "            preds = exp_preds / np.sum(exp_preds)\n",
    "            probas = np.random.multinomial(1, preds, 1)\n",
    "            next_index = np.argmax(probas)\n",
    "        generated[:-1] = generated[1:]\n",
    "        generated[-1] = next_index\n",
    "        if sofar is not None:\n",
    "            if next_index == 33:\n",
    "                icon = PIL.Image.new('L', (32, 32), 'white')\n",
    "                for x, y in decode_icon(sofar, 32):\n",
    "                    if y < 32:\n",
    "                        icon.putpixel((x, y), 0)\n",
    "                yield icon\n",
    "                num -= 1\n",
    "            else:\n",
    "                sofar.append(next_index)\n",
    "        if next_index == 33:\n",
    "            sofar = []\n",
    "    return sofar\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 29,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAWwAAAFsCAIAAABn/RTuAAAy5ElEQVR4nO2d3ZbkLKuAk72+++7u\nK88+qHdcFCACYoJVPAezahKjiIj4k/T5+/t7FEVRePm/pwUoimJvyokURTFFOZGiKKb4H/zPz8/P\n0sL+/v6sZZ3nia5c1zVf1nmeKB9akL44R73c5CzrpT1ZV1DDMOXrOlyby1Ov2LJedRxqSWnhclmm\nemmaTyjrf0K6R+h15l4Ch8ZpEUKhM416D73OOZOVLx9BV81Mh+37zbyUE9igyhLhb0eJWZwIa1ts\nfVBKq92zmhKMeyMPcnjtgPWq+kxeiXsKRA30Shnlmu/sb/O0utPr8Afq2IvqRVUHXZip0OfXRFi1\nClz/sObTEpgUlHnkhP0T6kSvUpiSZuKQ5CS0zIUH3YGP8F9Hbj3hFz3Yw92aeqzxvowqEqGmYK2V\nJqbQPDJ8tmeUgjWjR+h/Y0eDmVoPH0FDijI3GpcN5XRLePQH5JnM5z1IiDAwQ8HSlE3ja02NbCh/\nesVU4jgSQcGVr/kdTof6dfhfq0KRU0CZo37Idsv50UbzlClnNiI7yFDmyOHwujPKwWkPZR7eh30g\nsQ/LiE2flQuySqURxkSTQfihZOBEWh+7/k1oHfTEEtJDfdG6Qal6ObDFma7rE2gQrI1eSdKplMhi\no5q2lBfx1D4Cg5rM0H7hg46acJD2FaedzgidMxAqNCr0+reAx3qZIe4qTFq5PpPWeL6A6ynYdqGg\nqLY9u1Q2B3rZYMUTVqRHkxa1V++6zMCJXO+r7qs1BYuTA4fMJgjxybmjad4vJ7SEydIvskxjHZnc\nj+u5yBaYwzw0pmU12nEkgsLO1aGjUu5dutYLOhFjk8G7O/qRfZnv/2hss7YaCgHodCPQBmgvZqMP\nfYnmcyI3+JGPgZ1t6p8qbiOqi05Oe+XHaRjiLotGXihbU0U8h81qYNSgaWOlJisYWco6r21tNdZN\nsB2ejk/DsnreAS57U1VofMrzh80+D7gsyu4csTZhfaSIYpGDtq6C9W71TEVOpr+riZflzLMce/9I\nWC8ubGRQSyrHcQ8PBnotzKQLE/rgwrop9hqcek/JGzeUikRWgRZTHeyyA1VMcgLaFXirXZkJSFtU\n2xPA+lSjIpGF0JmIxh2s3iksUsE2sewsZvZrLvACJBuMODbOy4ksR1jKUqYvINC+6ZZq+Ibos6Bq\nwm6vnMu4y22+Zpi4pjPFZmgWp5+RLAjkMpSVmqk1nbNUJJKCWhMtJhFMyBSACKdC9AIIDqWcSLEZ\n8iCpD8LTghyEdTqshJ0iHa7Br6Yzq2BPfLh5ujZZQAsiiIMsInwAqI60gpoqQxPqmZZgdbIFlhMp\ntgF2IbbPQFfySX6kMeMimyNQPq7fUqzpTLEHvc1IIU3OeY0slVDNmUMiLSt46qTlfHGrquwtlopE\nip1o1s8ad7ueMwwZugON2LTupp0U+DidD0LB9PPociJFcR+0D7dbaPzXOET3rG3oIEwRXE1niuJu\noHcwTdPgg/qFT1mMecqJFHugt/iE6yAsvUBDCC7YyOXx+pYTWc7M/Pxx+yhWM7mIk8FCyokUxZNM\n7tdmoJzIclK1d5EZ2VRybjkdtTtTFM+iHGMcO6+3UZFIUTzMpfurPWk5f39/n5ahKIqNqelMURRT\nlBMpimKKciJFUUzxtrD68/OztLC/vz99WfrvI8yXNUmVdSxuL+t7pXL6tDp8qizTK3y0LCkS6b3k\np0cvEy13Pk1xD9VeMzj6S3h3Qz+sdLd4Q5r8nPigA3yQesoVFnnGvYywQrxspwMOS4AAtzCtJyNg\nEa/eoimL3Td9Voes5OhlPOt5s5nPL9EHTQ3UGKyJ0C+m6THJoREjMEMKVOjksNlaIpBjjWOaAfrc\n61/37gFbUFORRZVlu81SlAIcCiM/33t4U3v7r0kP53vzoQYy6V+1sKrP1Fp8Es73bt8u7liXe4Aa\n0yjqlUbpEHt323Xh8WGax9uUVdfQBUAPQnWo9yDUfSAxkP1r1KVyIkoRNW2szyQ8sZ5yJTJoSHwx\njKRgykPRdjQHFJf1Rn5BANaSZcl9sMWxpQvXX8x4TJhSdh9IHnh32AW0x97len4G5/uM9Eo5iXgc\n6kE0tnG9/x2my/IH1mgpDmvsteZLkljz1oftQmdG8xTqj/SlwKc0j6DE52g9a+xE9CoO7HXDEvVK\nnJfEau4vPtL7CJY0hO0GgmLvVGCsH9EYZ+uZw3LZNFYPMlM72K9ZYbRrIvA3qkBgwG+y0RU+i82t\nQhINvfmFwzyi5hQ9FmlgEkFjPjU2QqosZGJ7i1eenqVtHg3hoU2sNjK4MNnFCwqUH/FZjrWsmfxR\nniHDu/w4vDU5sN0QYZk/BUBlihqoHaHyjBU6cos1d6VIM2UNHzf1h2Eah3K2GHhgu6+brcMilAsZ\ncteDaxmB0F5Q7878B2w5qvdeA2fD3SczV0qg1TfcGbEzIOhzZ0oUuj116/JcJkPDpfso0bNKucAx\nR2Ql0OvHDuCx+MKBwIDuTgS/v6ig9l+froRJH40+UKAhFCqHNqupSOQ4uJM8BxkBWjdDQ1OGoeAL\ngTHIuniEFndDQbSs5M49XSQCvS+d8sl3V8jgZiPn4liN+kJu0I+8/JE2WpQiEXZRmu0b7K2Q2jb7\nXucyZHor8/pHFkkyz0ZujqWZxAc4wa3bYhyJXOTlCCEQmG9OdrmBuqf75xFoNitvCec3aKTnrY24\neBbbuzNwRQCOACHzUs3i0LMLSJCte93j8V0In1GLF8MqZI62nGsibRxeV6s2/aNFLDIXZY22NlaW\nnKZ5M8NmFUJjE9f7a0T0es/gezPrx5vPvzuTf9HYB13HhddR2HUkaEIrn+EEUb+aCUbuX8OCu3u+\nzPN4kEOORJpTVA7RyJVuaqxQ+FddhOpnaEITqYzPjXA8ZGb/gh0zjndjOEh3oOahXH0XdipYaMok\njdh1IjOLlyGWOtzucudslSFJU83zYR5ESBBbQUeHV2Z79CNfX8pHkCIR1FetGlzkQVqCcD/yyKZP\nLIKb+AwP0kARYvu9qPno3lxgzuEpb2awsHqz3Mjl90pf7ZJvi3TCgRtnveD8A2gRIrUEX9sJi5pW\nPkzVGtIdex/OLVd079da6aaOg4L24D+mXsd7XdiVEXdlhYU8dJoBXmHF+yoPciQ89n6Q1SwhzbrS\nBbaYF1zkTcKokfZB6PxFSOmuac8lwZ0gapyf5KmtZHQiL3q2ckM3oMvvSJgtuqJg61vIj+j1UnSd\n1tq06ACf0uzK0bWYHXU7SV4n8uLxJvmMGQFdI39csZMIbh3dMq2gWU8AoYMq30l2J3IPgo/YMQZ5\nQYdoeH2vkER24sOtqN3HgOSUE9HGyceeXY7KTJecktcLLYXAMIqVvHcYrFzJItLtztyMybC2sEI4\nCMsH9lA3+wxQXdIe0PokvjcS0Qcg6G5aW/QtJR7vW5iLZIvlznOPxZDz9/f3aRmKotiYb5/OFEUx\nSTmRoiimKCdSFMUUbwurPz8/Swv7+/ursu4pS7/0KCw6+url2xBJqMMqS1lW6t2Z+1+cUZL57Jl1\n5yJ2X4busCZU0RZstF8W7ER6pyQn82HvPqJfdyfR+J1J3+RQPjobPvnyIXtOtKeiYWUDPXVmp/8B\nhDkR+bV9feNpguF7TlvKJ1b1JyAnB2dT97Zqw3feTOkU6JvEKBOh0EXn33aJjKBzzy9wzMIqOiUJ\nQQlM+fSS3XDaMsonUuR6acRwcAJmMjk6b6P0NCM3020vtuTvh1sT40SgrSBQgh7w0KSyyVvKyb7B\nCqNPoJl5oSsCKI1eJL0M8/25NzzqL+rLOkYaU6IsKwOa9kpF2HQmxNk7Mgl/sUrv+IZFo6yUcm73\nSsuk6xEIbNztXsC7LUyDRfh6cYrdmXNudeOK/kDmjDDDrDQLqy2lUh5ZgfR9VjYHnxo1EsqLIzI1\nE1kKjXocCn/eiUx6kBchfsTkQYZDBLuUGP7IoVagZkxzqNE0wpdHyAMNQJqjP4wtVSdWMfoVmdWS\nLKK3RrBXtB8F1IZyYUW3CPP2iEkeVrCIujKl0LIcJT7sRELCkBe3zSHZcp/CHYbAK4GtsCnKDu/z\nC9mAETfcx/DtqB4ZpjPFCli/oLGP7RYg98KtW8f0WZ8nHIBffmSnSKRYQXmBoscw6nxFW6a4/slI\nJDyKDtym+WyaouDFyeYQ1F5ODRFrnzPqhZK4FwSSTmd6NSnvoAGewXv9Rl4DhqxojjNjkawfWe1B\nasx4nHTTGXntqsY0Frqk3xxEu4KOAtO4w3RWmBZtkvMIGg/mHV8xTzon8uLqcJTFKGDntNCbhOvQ\ncQRmHrQQmN8wFu2C+foF+5RPwlzTmdprNEG3ael/aQc7/x1OPfpzAescYXiiYV0Pz+87Pp5cTuRx\n9D0nj79jhxTl4dcM8rvZWvhPYjMnwgaugcb0Aat0bmcBJzumY+8m8YoMoRNs4t7k91A3bq41EXqG\nv5em99/bSNJ5WG1AP6s0WXanpvg86J7u8W4wDsNOF4nAWIM15d4mgmn8XBfL3MnVec3Xel4GepA8\n07RPZd0a81PkikSOu9QBPZFvtyLbWM3GohohqQfJyQnQp4+VQb8L/lWki0R6DLuEo3WpH9l3LWDS\nC9DNnWcrKAdTbqc/WanMTnaSmY3/PZzI8HTTvMVncwomoBNkfwhOgT5yWL6cdBv6E25bN+UNwA3+\ndqXdPcEpZ2WGSZ2IyQ6+1mjo6HqCD5TRW5PF3axnwf1ZJUm40DOzkAmb0lcpFHdMxqHp1kSEaefV\n4WYJszHc83ZobMWCQlp6dtXTmzU9LOgguj3twDzZAWMIdGGoXjbdHceRNhIpTPja/nqfv7C/n+WK\n++qlkMOiBSDlnIuGJKzAvVoIpQjBjq++vQzLiWwPaxBWK5mJrpcy6UeUo7TJad4/MRTcivJxfYm9\nlEJx6aYzxZ0MI/A8OLqu5pGlVdZkTjWvaQiN04e3hmGRXKK8CvNpTqRNO58WJDXJ/QWiSatfqYEr\nBcOa6tc4rOhFhVXTVFNZBPIjbtgMGw9PZ9ymjLS8UZcofFzvG5O9Fh9afCp8KxqHepp2xA2rQlnn\n7+/vZO5FUXwznzadKYriZsqJFEUxRTmRoiimeFtY/fn5WVrY39/ffFnKTe+QspRUWVXWN5e1XyTy\n1JJ7bRsXL+jG55dj2+LVbyyt4ykZlpa46OT1B2Dtq3eqMX+rOTzd1TmiJtTU4ETgYZheqbdhdSVR\n48aKyn7YmPaseazu2Kh2+f2IiV5d5DpqncjwVMxtqoSHZ2gTDt8ymDzetsJorg/665+LHKJGM3CQ\nW6RJakiOhluhIqUA1+jVSnpXk7PKiQw7IZwi3j/yoCusANfoL60Mmc9hmPnufgR5cM0B03Bu0ORM\nzoucLFtf1CVbJx0e9pXHacp4YRVmLUQ70HRuiM97wggVRsZt5Xz/yj77ZsEM4UbfeQ0iXvJWHLqC\nrCKwrB4X+CRXwkli68PhHP366rXRcmteWPnsIBIxxRdWBzaPvoimlJnh0WqX1uAwSmN6OUPCBNlI\nLjAMzhekGSQOEI8Ign0JJiX4NCY5EZ+FBRpNLLRrLZLw2THQYTQzelAWd0V8FuRQi4qaIKShb56X\nxTIUG2rMWsfudGZSZfog6hFQvVpfgsaKfrQHhzqRw86lwIBZk14fsgrFHfe+ToomYvJ/19EKyulW\nqBtVJkbV0bQaH4kI2qGdqpf15OATzqX4uxD3z8gCQbVjW0poPnd979ESqpfcSXpKmEdv/8ocHoTW\nRZgDCubBRCJKD6Ick6msT8HWqxdxHOJ+WDbnQkdFdu7GXjxcX/1hizA9eAKGj6D1DnkYQIl9Qmpk\n8M30F9HLnBVVo3boTWTJnR8lytaLlAijtLJLZA5PfC6b+hH9U4fREi7yF0/cUGlva5eZgu43nqGK\naKNc5DSD3F6Gd2d8dpMB6ozpOgX7m16k/31cGzAG0Y8ex3uN0LKCssTJuidRoJLMKyAUVs5e6Efj\nFPRIL8MX5khkFyV+M1u00RZCQrYT+IUmlJgMDPd7i7eACKOKadJLb8mPyz2qPQ6XnGYs9exgTfM9\noPDh6tMeEdLIzV1OZGMe7CG9omG/ZdNYx/OQ8X/TIOIGNLOVIcx05orbmt1rGrkpw/a63peN4X9R\nA7X/+twTXa5mS7FS9rMOd1tDDJGIZnJV3AnbIr2ezF58RapsDtauC1db5QC4DCkV82vb/MKqMLjp\ny6sw5AYC1TvcC1Q+W3wbg0hEM6xpHixuRuPBV/h3drFTmbjYlK4TqQWtzwBFlFFTifkTIpMCFHmQ\nIpFmcCabC1nvLRwIaoctKOwKs/8VynJTFvJJDA6bXZYPcMxMqotAYGNd4LXDnn+fbzi4EXO+n5Uu\nS/h4xidWoUFoxp8ymmxc6i/0VNsVDrRbvJp9oPm9omI18o5vUTiwvTtT1pYWNIVBd8N3gucXVqMO\nNBaPU8fePwHTxmpRxOL8nkiRk/Idxf2UE/kEhN0WdyYzlC/7Kmo684HAAz5yf3ZMfITjataLtSDy\nGVQk8vlo3s1ddEC5QpJv4Pz9/X1ahqIoNqamM0VRTFFOpCiKKcqJFEUxxdvC6s/Pz9LC/v7+qqwq\na8ey3Idrk9fr8L5+Ccuq3ZmPpd6WDCH2u9NuAW74gpS7rLfpzLkYthqBu4DW3EKqEJJtYOkmPZiy\nddfOlH8qTvLhWPQR2dW1O0UXFoLms7gyT66JQAW5DY4+q+8/1rI0Ty21Kk3mLU3P9KnRmNyTT+wd\nnUvzID35V/sR6MLCCzq5oMNX0PPTGXQIakZT8nkqNr0p//PfR1U0X0Uw5awvXZNGXzo823r/1OaR\nQq0Iar9hwFikHyFzxwvWkZ8CcOvU/eqHW8U+Ua1O6mbcZvfgi/k5/Yi1lcM7PMpwRdQsSHu9f8Vq\n+Cz/x6v0oiiLsQYIcqssddKxzIhqelaTGM500C2TH4nSfE6P7JYqp0O8gefPiQizZXkuLdwalnjM\nfUxUObMI6SST+ViVMySwnzwS+OT0XAjWzNJ6qOedyMfjbnv9gz23CEMPdk8BmqnSP/Y8uw9NvSZB\nxVmbQ7MyHQ5tIN8sVRBY2S6aWo+dCCq4l4yao6+SIWhKX0qz10lhWg5DzdOCZBmardCGC+m6k0Dx\nNO5GzgRdDJBvJahBz9BZkikrZeLBwirbBj17Rf/N0JnDUVYttu6Ovq1vAquosVVjK/WyMdaoJqHZ\nDtMIfRhlFaIZKl54PwrPUIpEJpstv8t/oRnr9IPhulCIzXk+RKe3jmRtd39MCp+9uK2K1xWYZpEH\n2YJuJNKs01exa8PPeQs1TVuLnp5715XLEGnra6X5RNYXDJ+Fj/RmRlG62tSDHL6F1RVx5oMIo9bk\ngHYPbOzQmyYcwOjdaw2ruf794Yso5U9mSJ9aZxXJjY3F8Mer0A9ogvRuZuReJ1Rhl1G6VxfW9O/0\nI+c7t5XrBqqraQ9d7D07X0HWUBNiiESgvl7xIZrsUJOdl28RrXmaNUPniBKzyfLQE2woausVT5lp\n/u4xXOa8ZwqcXFELz4kkr/mw1zm65SP0HEFvzKfhAHKgcjWjmjWnMrMhjG15eP4FvGIex6ohRAgn\nF5HQg0RFmr58hDWsyZxv4Fv+Fm9IP0nbikfE/Hl1vaBsmcfV+5HX41DLJtzxDJvOZKuYA+Wu00m4\nRToVVwdrmm8DDg+Tber25mh3QliOSWVyx5e8OyMvnaI2Q+vHbQm5OtunwgaYw446XFLVdHVoYDRn\nuHfRNoYSbtl8hRNpsH6EBiDUnmjcUQ4lhIQ6hCJpdrjcBcnOi3Ur8FaeKNjpRFDotREaP3IofESe\nJtwFVp8Jp4SIyUXrHvL62lLnFY7ZiWiWkZPD7t2y0Ycvtw9mMpbeRVf3TBkEbcBlVPnxDF3vu6Yz\nL6je5eMScDq6SzdYh3XtQH48OfQozWrgqoe8azO8dRvSFi81l2GEjx7P4CYpbDeYWUj7QpoxuI04\nz0Aqs6iXytkqC83gQQ7rORGhyaHXzG8ZEE1LUItP0n4PMtnKaY3kBK90VCtrcL47Q299pLrp7u9H\nVtPKB8/sWliqjE/hg2nd4mrenMgwwqTb1OzgfHEfYvhIm/tyrjiersp/wDkaXZugwN2lPLW4mUEk\nwqpPP5v9WrUWm4I2/oV1H+r7KhLB+Pr/p3oNdNjsU6tZsAz3WY/vNolv3OKd5JvN5Xto4bY826pB\n5TiO8/f392kZiqLYmIpEiqKYopxIURRTlBMpimKKtxOrr/URzSqR5uwmXHN6/YbrLz8/P0oRfedE\n//7+kpflo8qqsrKVxRx7jzrt2zJx7J/3Hllx8PzOsr6ED9uz6FnIfAXZD4Wwd8PLEvIUToexME5k\nKC4KMYZO57K8TSO/xQf/O2+pchPCBLG9ItAo9Vn1qrCuh0Qxf4gr0EjQrVgtwQxRube9yOPQdq6v\nvaN3n9HhLtabuJVres96sixU4oxIygx7eutdn0TzhvdMM7mn2CFGIozYsebXMlwXBetHdGgksgyJ\nFlapB4H/PbioBD1oLUv/1kZL6et48P0LzSsk8vsarGxsPj2BkRrZx601HXq0payIp4b9x20S8mhB\nY5AhynIdcwKNAj1OhA0TJrEOwiixqSHdYk/2Lo3PQq7EKqGQLfoRi6CZ2JkgC/WJUXlq2utwmZ+c\nuaYKw6FC/zi6brXAFN9YZWW1qlhZ4UmbthqNNeRppcwEPjQ3lM+KLj3ZJUIEuKegSZCQvZmLHLFe\n/94M9JmK0jkqsX0KACVWljScWB7v6oDXlWIoJQ8ZFfVautlhaTJcPbOgofjS4rbmhhiNlmjtKZoW\nfHhNRBARxlT6ud9wZh7SZpqWgMVN9qX5rgjHrsmsErJLDNLozRfuWTwa2qS1p4y/J9IrzHeLRZie\nTSZ+CupBHMtgk25ImNkuHQNP8HnBFfk/iG85c5gnumJtF9PyzfV+SIIFCTbMWRWJyCX1bmmyHYqI\n7moSswKgspD8pt/DsoZ16UnYCmo/UnnGISh4RBeLHq3R3c8qH7/e/56enFIvWKIt3h5oMckBHdLb\nb7ieAotoYym9zmaoqQX7W59gC0xOX09vzNQwWS696FvL7GUiFNQTab520MUjrPIcPScCRaTFaECZ\nJEcj5BYVoVDnO+OOlSWiHyG5PZXJ0nZndYVKRJ2oVxdlHVGyoc/VWMsGkcg8w3GDTXCOAr9hApMk\njtzSEuunfMMYGtJM9NoC5XZFrFULtod6NRrX4ePPwjuRph23iO3ZYQ7uMJVFI9jxvoN1cnMTdPF8\nX9OebznWaIokQOvVjPzzkQ50CsgUX1eEEW44RlLJY/1v8IeaV2QSzmSPnXkcDSzCxeKrYF0JG3Gg\nYS8D/hfwzugvBoSwVyfcS9riWVA8EtUB53E6kTZni6rGfHdyT32F0he1WS+f8ilfywW2AuVkrx+p\nTMWzsJqqAm4ctbhnHvQZ6i2smJY24GrdOfq7Fqvx787Eym1a7HGvAFnrEl7HwGTF52H1I3Sh5BH4\n6YxGIP1WSE7YBoPrWOwtR1MpI1VItpWzYjUn+JOdVmt5HByJBI69yl5n3cSVGQqD1kHYBXD035YG\nPmuKIU0OqDzI99BMYr65001noqYVso7Cq918uXD38B5bpk5nKAnMv+eqehLScq2irkhcxALnI/Su\naZKSbjoTgsbLLvIj+rvKuBH1ZKUfoUYwfND3VLEvF/haMLyOQmZlJo8Qf+w9cEK0lBuE7M2PVjxV\n7Mt8l3k2nFz17kxvgpMK5aRmZpHCpwflU3r5e9Hy0YnO2OvFUswLB4SnJP+KF/CGCP1wuyUD6AKE\nSXV5iiKKb3ciw5UtlGw7Lm6xeTvPWGTm253I0fcjm3oQVB1aqfIgRSy5/gLeU1zk7+lt6kFewLV6\nKP/WlSrSUpHIf7D9at/OJki+b6WKnFQk8uGgA281lynCOX9/f5+WoSiKjanpTFEUU5QTKYpiinIi\nRVFM8baw+vPzg26z76EKyOn//v6EsmKpsp4qy2oz7rLcrynm1+HNZTmOL8Oy1r7Fe9tuYp2AQNAD\nL5RwXdGy1h2uv7OsFURtk2Wob9eJtCbx1fa66wNN7DHTDJrVsFTapS8WCmXR423h1RSO0rlNjlbh\nTg84k5Vbwqgj2oZI5M7IopX4+rGLUzABPxixqIJytlGmLDQT7I2r7cc94EH5rR8Bou9MDOsY65sc\nQzWtIwparf1O+sZqe3fL9H0U9GCIynpq2vfoFB0ENnWUmvalbxXElh4C+37A5yGsWFGHqGwvZncm\n0FNGWczQRvd6SxVpeLh48QHIb0tP5sn+15TPZBOc4IO7adsRjujy90dgAs0bm6otXvYLKAKWqnVB\nQWZvIe3g6pl5VGd9dHL762Eab2K7WYt2odVNtjttFE2GgmX6OAkzuSGRTCpSun48nQlfTLoWr7DS\nKFQoaNLXCKGgKQf22XWK6s2BQ7LtZdVT9Wp7mGG+879+vOrYq+Cw4r5+rhRMn3N7ZDgV9R82i/KR\nUQxtmh0xlFWgKa2jhGZ5SB9A6mGzCm+43si5YhazDtQ0GplhOMxmkoeZ6Z6cIO9bvFB0Wg3apdlk\nQp40E01WNPDRoLcwjePX06tL+ACAnAXtWqxsM9HcipyPTkVk87vIns4WHrMxH1/vfewdzrGjRoDe\nvH1muUc/IIcPaL2pU0jmw5zhmkXs3D4kmQBcaukZQ88kNlrhCnF/Wf4WrwnoNcIXcXp6nLQJ6keE\nH0s9SGwRytJRHedXQFEflnv1PFt4BD10xQo1lqm+n7MmEsLSGsGRjV1hackmCzJN7sKr3Buc0y4W\nTBI+jMXiFo9Oc3qmsvd0xkR4M/sylKODtLaIUC4Ss9fhus/u4xDUwy6zGDYEngmH93YipgbTJIbh\ncWzpqBQWX26sVMrcfHavKULOM7+vpDtNtEZID8lDEj1Wq9jYieR3+fcjLOjQXqF5cBG+4u6ZpqGp\n5dDL032cvbwJDKOQ6pQ6zLvFK7BL89yM0kH0dkBNO6NLI7UMaCq4kct4NS4bWEEnyEZb12hnbUsn\nUlCaKTdrgGbNbpTA35f6IGmUaxia5lNY3QFNn9ChIJFMElKDQY97nMgw07SwVjt5MRvsgHO+f6OI\ndTGmXm1VxToL0bu/LwR1z94Yo0GYJPojkV3Wor8WOosR/mvKM2F3TShSTlYEgE4nQn3bFsSa2kYV\n1zCsTrgHiQoiyoOw3ObxzU6kF+dXQyaBmg5tMvluWkwjVhmkRl1w3ZRqTKntjbd4iyHQCNAmpXwC\nIjPKiKk45jwp2uc++qENjkSUESZaqBuKstGwsJ3ASciz1FroGUYrmqN05ukMXOClJRVJ6LXOiolM\njfxbg3o0G3TII6tnYRV5r/IgqeiNLYId3N+CkyWW23LDbtLJ8cgwNp/dnSmSE7Kza0IIdubNRrm2\nWvYp0zve7tNbnVgttkR+k6VwgCYvh3rXlXEicG0V5mIaAeSto8zsJa3APa+30FgDTXWrV+dBM6Wl\nb0IMsx1v8dKdnuF/t7YbpeIKynDBZbViq+H0yC92mN61GUxn2txJOT5PTq5Ws+7dkFTox5AjrqXg\nIWZ9EbF2UoGPDDQM9h2rlgymH2arPWxGvQM005l3BIsV6Jvg/McwGU1v6rGniD6TQ+GVypUMOf/9\n1T46pTBn9fv7GylaURRfRh17L4piinIiRVFMUU6kKIop3nZnfn5+lhb29/e3S1n6DSlU1u/v79J1\n5Rt02BYvlWXpVzEFzWxkGyvKQvbme63kqXp9y4nVnqG7v6HQw30s4qktLXbDr/bXIEPNmIYcFmEz\n9YZ2mTwUanAiW7+zILyWFiuzuz3CJdGUGJjb8NsR1gxXbNNaNUxl6LXvTPPRU7+9BCuOgFsPuVMG\nTsTakJlHM+hH0GmXcFF9HeBOP0Jbqk5YINhjUPCADDqU5Ws+zflMdBY80E5Qc/ty5hdW0RGgywjK\nxHqgaCiPG3q0hhbUkrmLcAp3o9tFLXtPoSZgQ0RxWFxkr2/3LJymtFZTk95REVO2PqTdmWGXGz4I\nH3dLCfUV4kpWI9uxkOwe8eYdZSrCTQI6hZ6KkGH7lOl2PVF+hMZTvmzfnEjUgI9wC0ejoXY9ULwV\nyBI+2IFlw82v2NVA/Qy1MdOZhw0hh/DzfoQVwJdtNxKZCRzazM1dSfqs223LrAgOX/TavpU4r6UQ\nkeDdjdz0BzDpyhe5MEe22InM99JXP2lCODoJTA/lWWrZMzPbmUJvi0pgTIeGuNfvCyyvFkpglwsZ\nD24IvZudR4U53d2ZELkDg67WTiuCkVj3oc/t5nUQWmJFHJP4LKen9lPx17MDud6/DODOBzuRFT4v\nhNV+JDbDVCVSD1KuRIlSMyab1CxLmUbfmR5xcZ96t8KsidBdg3n0AvXmMu4afjlQn2cHmtJXkMBU\nHT6IwJiXtqwvHzRvcsQmuY69U+lh9az+qKAxyIr+rIzAt2s7a3ARSLN2oT+bopUhKDdTY+VyIjIr\nJjKfB9JSL5o7yV7MjFFWu5hgw220/HcsmPigx2nOvnbcyYkUQ5AxCTYhm0s5hdvozd+hK6HN4Z4W\nrQhFUzgRjUaUMXNhBU2tj/Igy0CxBtLzazGiXZRDjKXzUyspnMihMNwMysqMbzZbXuN+ejq/LN/o\nTwXjRHasxhZMKja8q7OT6nIoNxCoZHeEHijD2xYv6x2jtnXl9SHhCgrz9qWpRVCXcCvWubPj4Qco\neUeGG+GaCEXZdr3uOUPw90Qobk95kOHxM0y8VwuhdhodmvSMPEjFnvfQayP2+kaNwjuRO7vrxU0F\ne0vWGxHif63pHVHbdhMZ/cbnRkCB9VtsSXj+a+9UZfK0KD8hMvsyuRSnBuh2jLu4dWgqYsLhXJ6K\nBdD09hEZTDy/O7OFmqw8XqlePNILp9dLFMnk1G++lALyvBMpYpFN/wNWUocy71gpyka1KCeyBxuZ\n1M2UZh6HdyIbrQwXRfEsFYkUxd182CB9xxbvh6msuAffjjWLb9/3wc8mOLINz1OPPxKhuqjZabGO\nycPdvkM31lLCE2+B04lsdx6m2JFF4/YQ0/GZTd+aCyTgsFl5kKL4ZpyRyBX9hfSiyAB8n0j/SttK\niTbg/P39fVqGoig25vl3Z4qi2JpyIkVRTFFOpCiKKd4WVn9+fpYW9vf3t6gsuuW8riwKLasty02e\nlaKr12xZ7UrsIp9Qryhaho72cp8z6JXV2ysYfi3ZUVaPO8uaAZb1CZFIzq9CteMD5wQoKxbhezYf\nCVLO8Pp8Wb67joKGZeVs34WfR7x56yv8MzYbcZG/eDSfoSm9vN8feBpADrua13YUJwxFcr3Ozp8B\nsxY3fBDWTvmIptAhwyIGTkRWXy/Nd/ZkxG3vH+XR9mn5o/anfWY07NVwIDHlT3PWuINWXCsLfetT\n031QbnJxrKdz+ywlQ1UERCJu+RxmpJHkqbPSD/L4C5PX+x+XZxMMryil0nc2Uweb0SF8Vh8RwxOb\n0BMN5XTbua+Ow1IeWxNpzTwPm7NQrjI3R9GN8Klc/mORqCPRfjXDqfYgKJlsDEffg8AoQ3i2d1dw\nJfDiS1FDIdHjSL36Og7tFj0FlSM8NZ7OKDNCafR2w7a3YASsJDCYFMoSbLHdYnOmydp/H+/eeRwW\nq//7PQgSxt1Gyu7tzhz9btKiu+gKetDkgBxotKd9d0Yjq68+yERQw1PHRGuFVNmbOsIEVAzqm3rJ\nVjfb1lA1RvlZRybzjaVZEzEhDLFodVwevdBT+vRWkDtjJbdNZy5A74pPvibl0enPsES5JYbllhe4\nhxDfIbS1ZnapX6G4ATlI14Twwl05gqYJlEDF9kofO5EQ78CCHJt1HnR0TMS9VNEE0M/dChnf6NKQ\nPYjyYio/ImhjGEpY/Qua+OgHezp+yw+qpjNXf+19fg6M2lgzAb7eN9JM5crhH4p3hskKltv0Qwch\nxIqJ1QzKRVDKjOSsltiOBpdg9flrpzMaHzYz7KBnNWVNlovMa/jjSDOafQnDMATdYiMO91x7BW4r\njepWrDCsqZtKtH2UaF1jzPRP5WSyRRbI0bIuGVokGsHgrSTWmY1STg/9bOIG7bEzDEe5Kd6dgd1Y\nP26gxA431OZEvULh+kiFIQlhR9HtENYEoHnLC8k+DUDLd4dsjzkR2D+PTgVYxbEaH2Z1gAjCuphC\nl0VqpGVhLXtF90ZxIhxON20U2TXIOvTNkmiJci8TSPrHqxYNLOyKkduFR0hUTDFcWLVmEoh+efIi\nB51oAnQlVeSVzolYw4TjfTY0TGxdee5lMpnDjigHfOUUAyaz6pOV4dK9gSJIkgF2r0RONllKD71a\nnl8TERbVj36Yx2IKwy4X9voVPMpZZG8+O5yxKh/JQBOPlbA33YDW/qxxPu9EGrSlY/1lEYgwK6Q+\nl02saTh5eav3iD4flKA3ZszcnQEqsKcu3ypGLCmcCNW4YJE0GWrOlZIWXctWdmw0fipLNA0tKLFy\nKOr1w2G90FOxXVrp+0LKcpPCiSDYtnlx9ufbSlVW5BJCs12TPuW4XVMierA3Dut7l5ChvmqCGMoc\naNHshAWNqUmMOaMTOcRzimyCizsIL5OkATYFag9FFqgDXNyJpoTKZ2NeR9TjcJHKWIl96tmJzIss\nTiTKsw5zeBUkzI8ytEpOYBs1Nb5uoQET/b7Ajpis/1ToZ1uxhbIco+h7hbTKxOm2eCHU8miA97rS\nfutzpgpCQ2V+K78ZWb2CxrZwylR+t9jWB5Gds1zvG9gn9yqGqUT5EVMVEjmR6/2jLAcJ2DSxJVS0\nEjpLguF3uRKWphyqMZSmsUiT8zEsO+Fy4KuvxoO0BCFh8rCbWC0/kRN5QXVKg+GQJu8VR0spP8LS\n09gwsZyyR8+y14U5+nafjFn0BUXN+gU/4hg7s6yJoHnKoVvdoOtYpnmNPAjACT+KhraIz1dzEvSJ\nHcX12ojaQCvxULcUWsdxDBvowWEO8zGFPkJngWE+ahprhs87ESh6awZUQ5hSGJGUNoocgTwRZUX9\nQpZWXN/VaRP0rsAxQJk5TSy3OLVGfSeEady69fk7aPnocXhFL9XzTuRF1JRBOZKYAsiazjSai4c/\nKC2x5q6yaOU4GTI4+3KwPkh7LI0LWI5ps4QOgs3H5NfSrYkMCe/SpmEwtuhNOcnHcX2Non9wxjU4\noBXUPOIr67Iv68bqYd6q8zqR2zqtpqA28N5szTmhGhu6A2ESOixOr/P51qFdWshNn3JYqO9Bd3Fw\niQDegu2ol+phJyJPOI/3ftuzVGg6jpASpUc20Qzlyx3HC9pe8Mr5/rd4qdKWqjHKvzdLswYIaUHt\n0i72EtMEgxHi9/d3VsaiKL6YLAurRVFsSjmRoiimKCdSFMUUbwurPz8/QlLToR2Wv7+/9vu1FmNa\nBpMFQFnBsuR60fyti3Omsky1mCxrkiqrytKUpd2dWbRGbT0pEJLVMP9F+7ix25mZ2asW87a9+uDG\nUk3qd2F6SNMZdEju6J+TQ2fv0FMsNP0x2nbqnYBEx/4coPyhYL4Maf5Ih8OK7Luz2DuD8CCCQSrb\nRW6vWDn114dZDR9U7vvKeRrOiQxVds0dD+udyBDOHRwdLehblx5zaP+9vH+CoFeELPz1/tIQlWcL\noP5DtBcijzKlVdRALymHA3CIdeiTPihIzpqrXNNuJEI9tBCVtFvu0KD3lOApmzDuMYEGOFCYkBFV\nLyGrW/3IMMT9iKOyrRbDKqzDah7WMDCwXmyfQrI5+hR6lrasHF6h9IIa35yIbEBolBYusnn2qod6\nL3yKlZi9e767g96DSl75z7QckqqX4Hp/n03IZB6fXzDlf5DKPuVHYD+RJUH++tSNHIs8iJzSZI20\nIdimCXl2sMXb84vWNMcogqLesfcIm9LNBSIs2XVa7WboQQ5guM1nsWIIRUf5TZqtKb1Q2fv9CBRG\nWW5TPs1HyD9Q7cqs5q19xSPMmggKB4a9CDpvlHhyJEfyHBa3LdMbcDSdWVO0PuXBvfRF46Cz/95Q\nC2ccRcvymxKjEuHFK2h1yQSsgrLo6/1FLX0Rbp07HpxUprJqLbEm2fiwGbLvQ7QwmtgN6kIoWpnJ\nmY4kdObCRnRLB9XejOz+ucDuQB2if5WupP07NHX5Sk6UVdP3Mml3phWGei87sRT+64b1mlEeZJjm\nNlgHDR1cZutUGtwjwchk9M7KLLeFr4L3ty+Mtecj/XEkYvLisaAqUacWW1BUHX3BLYpBYB2XRkAz\nDFcN9I/kRBg12ViVTTzMH+Yg0HvQQWwrqN6dGVYAjaV0AuKAjs/zNe+t1Fyjj0csMn3Wbd3vrH0I\no5Zcr438iGAP6NbMOPR6quXZ/kX/decvlHu8d1hfPsEfJWrhnHupCUJd+4oxWRj8GzMqVj4II+H8\nfsQX90JnvUsdlRMxX3WaGpXTdti/Jkts0b1g9koCnAgKyWakuRloH5e49DAziipVEbtksLSX+jxI\ne2QjP/Liet+tYyPZmczhf1kDQxdlWx0WR3PzZdVI/SkANmhcEYaExE0haGxIZvWsgU4ze3d7V9hJ\nfkLYsJcdMleUu6iUFaae9EPNMJKE1yfHsWFoOuPjJxmuyyRBGbJBrOm/DVNEkzCCSx2JIEIGgRWr\nKlHsNUQfwCm7I8ScDSFAA7Go9tLnk01p2IncE6EpQY20QnfZ2uMF7KW+gCtaIica9eZsghfskIO2\nS+4RA602hnSoKPB0JkmL9mLmcC2ggnrVn4kh51Xqm2TlcSUye8kZ7kHYVSR5xzCbxpKuiRz9o2U3\nbzpEeVWr2DPl3uN/EdksOwS6yxu7aoYcE4oyYJPFbt7FwkxnFhEi7golCpbRWjEwDGFXEHrLCo8E\nhiZ3o2nfpSZxP6zkIfuGw6A4yUQBgZ3IuQyffPeY2tmPQdYJ0AoVlBNY+qVeUdannCFnf5Bplhzr\nBGFuqL9cZApj0ts93ee+3Zm0I88pTpesHnCR67mnY+uZlyetPVDuUTsN0C4wkWETCNxpKm9rIhu1\nazi07oH9FlkGcjT0Srt4Zp0Ga9SyOpR7itX9E/qOLbS30zmRLWBdj8Ps2BySmBQ7+2u39JVNUp0e\neeK+5JQTCYZGGe26poPBZDB2zTadOUQXAKtwo0QPs6Ky82qcXJREsHZYTmQ5aB7LmgVtm7TdT2/W\naasQCB0twvMPyXlpW5QTeQC2RfN0OWHUio0vUgVWlAf3oaP0HJIPDGTYJvt2J0K1jGK/qHkEjUdg\nw6ASJ1udbjDB68O6zJS+9QEQB8n94As0o/HNboTt7bwnVm/AuncbfmqD/g5B2LT+qh4+z1M+IraZ\nrolTtpoHv9eJUO3AqISNA6Ev9zXzFfqBXJaWOf0xfGSy6A/mIl8A2yIGObgX9sIl/14nchDNsud8\n0GlC+qCj0BO8xNWTZwWCZ9yoVyylpwSkJWgJ7UoqLwwbdGayTL0nTXP+/v46xSyKoqiF1aIoJikn\nUhTFFOVEiqKY4v8BhoflHZ6wtlEAAAAASUVORK5CYII=\n",
      "text/plain": [
       "<PIL.Image.Image image mode=RGB size=364x364 at 0x7FEA549B8B70>"
      ]
     },
     "execution_count": 29,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "cols = 10\n",
    "rows = 10\n",
    "overview = PIL.Image.new('RGB', (cols * 36 + 4, rows * 36 + 4), (128, 128, 128))\n",
    "for idx, icon in enumerate(generate_icons(model, num=cols * rows)):\n",
    "    x = idx % cols\n",
    "    y = idx // cols\n",
    "    overview.paste(icon, (x * 36 + 4, y * 36 + 4))\n",
    "overview\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "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": "3.5.2"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}