DenseNet Galaxy Zoo.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Using TensorFlow backend.\n"
     ]
    }
   ],
   "source": [
    "%matplotlib inline\n",
    "import os\n",
    "import matplotlib.pyplot as plt\n",
    "import numpy as np\n",
    "import pandas as pd\n",
    "import random\n",
    "import tensorflow as tf\n",
    "import threading\n",
    "\n",
    "from keras import backend as K\n",
    "from keras import losses\n",
    "from keras.callbacks import Callback, LearningRateScheduler\n",
    "from keras.layers import Lambda, Dense, Input, concatenate, Activation, Conv3D, MaxPool3D, AvgPool3D, BatchNormalization, Flatten, Dropout, GlobalAveragePooling3D\n",
    "from keras.models import Model\n",
    "from keras.optimizers import Adam\n",
    "from keras.preprocessing.image import Iterator, ImageDataGenerator\n",
    "from keras.utils.data_utils import Sequence\n",
    "from skimage import img_as_float, io\n",
    "from skimage.transform import resize"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Prepare Data"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "TRAIN = './data/images_training_rev1/'\n",
    "TEST = './data/images_test_rev1/'\n",
    "LABELS = './data/training_solutions_rev1.csv'"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "55578\n",
      "6000\n",
      "79975\n"
     ]
    }
   ],
   "source": [
    "files = sorted(os.listdir(TRAIN))\n",
    "train_files = files[6000:]\n",
    "val_files = files[:6000]\n",
    "test_files = sorted(os.listdir(TEST))\n",
    "\n",
    "print(len(train_files))\n",
    "print(len(val_files))\n",
    "print(len(test_files))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "61578\n"
     ]
    }
   ],
   "source": [
    "train_labels = pd.read_csv(LABELS).set_index('GalaxyID').T.to_dict('list')\n",
    "print(len(train_labels))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Construct Model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [],
   "source": [
    "def DenseLayer(inputs, k, bn=None, drop=None):\n",
    "    x = BatchNormalization()(inputs)\n",
    "    if bn: # Bottleneck\n",
    "        x = Activation('relu')(x)\n",
    "        x = Conv3D(k*bn, (1,1,1))(x)\n",
    "        x = BatchNormalization()(x)\n",
    "    x = Activation('relu')(x)\n",
    "    x = Conv3D(k, (1,3,3), padding='same')(x)\n",
    "    if drop:\n",
    "        x = Dropout(drop)(x)\n",
    "    return concatenate([inputs, x], axis=4)\n",
    "\n",
    "def DenseBlock(x, k, L, bn=None, drop=None):\n",
    "    for l in range(L):\n",
    "        x = DenseLayer(x, k, bn, drop)\n",
    "    return x\n",
    "\n",
    "def TransitionLayer(x, k):\n",
    "    x = BatchNormalization()(x)\n",
    "    x = Activation('relu')(x)\n",
    "    x = Conv3D(k, (1,1,1))(x)\n",
    "    x = AvgPool3D((1,2,2), strides=(1,2,2), padding='same')(x)\n",
    "    return x\n",
    "    \n",
    "def DenseNet(shape, k, bn, theta, drop, B, L, outs):\n",
    "    img_input = Input(shape=(4,shape,shape,3))\n",
    "    x = Conv3D(k*2, (1,7,7), strides=(1,2,2), padding='same')(img_input)\n",
    "    x = BatchNormalization()(x)\n",
    "    x = Activation('relu')(x)\n",
    "    x = MaxPool3D((1,3,3), strides=(1,2,2), padding='same')(x)\n",
    "    for i in range(B):\n",
    "        x = DenseBlock(x, k, L, bn, drop)\n",
    "        if i != B-1:\n",
    "            k = int(k*theta) # compression\n",
    "            x = TransitionLayer(x, k)\n",
    "    x = BatchNormalization()(x)\n",
    "    x = Activation('relu')(x)\n",
    "    x = Conv3D(37, (1,7,7))(x)\n",
    "    x = GlobalAveragePooling3D()(x)\n",
    "    x = Activation('sigmoid')(x)\n",
    "    return Model(inputs=img_input, outputs=x)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "____________________________________________________________________________________________________\n",
      "Layer (type)                     Output Shape          Param #     Connected to                     \n",
      "====================================================================================================\n",
      "input_2 (InputLayer)             (None, 4, 112, 112, 3 0                                            \n",
      "____________________________________________________________________________________________________\n",
      "conv3d_17 (Conv3D)               (None, 4, 56, 56, 80) 11840       input_2[0][0]                    \n",
      "____________________________________________________________________________________________________\n",
      "batch_normalization_17 (BatchNor (None, 4, 56, 56, 80) 320         conv3d_17[0][0]                  \n",
      "____________________________________________________________________________________________________\n",
      "activation_18 (Activation)       (None, 4, 56, 56, 80) 0           batch_normalization_17[0][0]     \n",
      "____________________________________________________________________________________________________\n",
      "max_pooling3d_2 (MaxPooling3D)   (None, 4, 28, 28, 80) 0           activation_18[0][0]              \n",
      "____________________________________________________________________________________________________\n",
      "batch_normalization_18 (BatchNor (None, 4, 28, 28, 80) 320         max_pooling3d_2[0][0]            \n",
      "____________________________________________________________________________________________________\n",
      "activation_19 (Activation)       (None, 4, 28, 28, 80) 0           batch_normalization_18[0][0]     \n",
      "____________________________________________________________________________________________________\n",
      "conv3d_18 (Conv3D)               (None, 4, 28, 28, 40) 28840       activation_19[0][0]              \n",
      "____________________________________________________________________________________________________\n",
      "concatenate_13 (Concatenate)     (None, 4, 28, 28, 120 0           max_pooling3d_2[0][0]            \n",
      "                                                                   conv3d_18[0][0]                  \n",
      "____________________________________________________________________________________________________\n",
      "batch_normalization_19 (BatchNor (None, 4, 28, 28, 120 480         concatenate_13[0][0]             \n",
      "____________________________________________________________________________________________________\n",
      "activation_20 (Activation)       (None, 4, 28, 28, 120 0           batch_normalization_19[0][0]     \n",
      "____________________________________________________________________________________________________\n",
      "conv3d_19 (Conv3D)               (None, 4, 28, 28, 40) 43240       activation_20[0][0]              \n",
      "____________________________________________________________________________________________________\n",
      "concatenate_14 (Concatenate)     (None, 4, 28, 28, 160 0           concatenate_13[0][0]             \n",
      "                                                                   conv3d_19[0][0]                  \n",
      "____________________________________________________________________________________________________\n",
      "batch_normalization_20 (BatchNor (None, 4, 28, 28, 160 640         concatenate_14[0][0]             \n",
      "____________________________________________________________________________________________________\n",
      "activation_21 (Activation)       (None, 4, 28, 28, 160 0           batch_normalization_20[0][0]     \n",
      "____________________________________________________________________________________________________\n",
      "conv3d_20 (Conv3D)               (None, 4, 28, 28, 40) 57640       activation_21[0][0]              \n",
      "____________________________________________________________________________________________________\n",
      "concatenate_15 (Concatenate)     (None, 4, 28, 28, 200 0           concatenate_14[0][0]             \n",
      "                                                                   conv3d_20[0][0]                  \n",
      "____________________________________________________________________________________________________\n",
      "batch_normalization_21 (BatchNor (None, 4, 28, 28, 200 800         concatenate_15[0][0]             \n",
      "____________________________________________________________________________________________________\n",
      "activation_22 (Activation)       (None, 4, 28, 28, 200 0           batch_normalization_21[0][0]     \n",
      "____________________________________________________________________________________________________\n",
      "conv3d_21 (Conv3D)               (None, 4, 28, 28, 40) 72040       activation_22[0][0]              \n",
      "____________________________________________________________________________________________________\n",
      "concatenate_16 (Concatenate)     (None, 4, 28, 28, 240 0           concatenate_15[0][0]             \n",
      "                                                                   conv3d_21[0][0]                  \n",
      "____________________________________________________________________________________________________\n",
      "batch_normalization_22 (BatchNor (None, 4, 28, 28, 240 960         concatenate_16[0][0]             \n",
      "____________________________________________________________________________________________________\n",
      "activation_23 (Activation)       (None, 4, 28, 28, 240 0           batch_normalization_22[0][0]     \n",
      "____________________________________________________________________________________________________\n",
      "conv3d_22 (Conv3D)               (None, 4, 28, 28, 20) 4820        activation_23[0][0]              \n",
      "____________________________________________________________________________________________________\n",
      "average_pooling3d_3 (AveragePool (None, 4, 14, 14, 20) 0           conv3d_22[0][0]                  \n",
      "____________________________________________________________________________________________________\n",
      "batch_normalization_23 (BatchNor (None, 4, 14, 14, 20) 80          average_pooling3d_3[0][0]        \n",
      "____________________________________________________________________________________________________\n",
      "activation_24 (Activation)       (None, 4, 14, 14, 20) 0           batch_normalization_23[0][0]     \n",
      "____________________________________________________________________________________________________\n",
      "conv3d_23 (Conv3D)               (None, 4, 14, 14, 20) 3620        activation_24[0][0]              \n",
      "____________________________________________________________________________________________________\n",
      "concatenate_17 (Concatenate)     (None, 4, 14, 14, 40) 0           average_pooling3d_3[0][0]        \n",
      "                                                                   conv3d_23[0][0]                  \n",
      "____________________________________________________________________________________________________\n",
      "batch_normalization_24 (BatchNor (None, 4, 14, 14, 40) 160         concatenate_17[0][0]             \n",
      "____________________________________________________________________________________________________\n",
      "activation_25 (Activation)       (None, 4, 14, 14, 40) 0           batch_normalization_24[0][0]     \n",
      "____________________________________________________________________________________________________\n",
      "conv3d_24 (Conv3D)               (None, 4, 14, 14, 20) 7220        activation_25[0][0]              \n",
      "____________________________________________________________________________________________________\n",
      "concatenate_18 (Concatenate)     (None, 4, 14, 14, 60) 0           concatenate_17[0][0]             \n",
      "                                                                   conv3d_24[0][0]                  \n",
      "____________________________________________________________________________________________________\n",
      "batch_normalization_25 (BatchNor (None, 4, 14, 14, 60) 240         concatenate_18[0][0]             \n",
      "____________________________________________________________________________________________________\n",
      "activation_26 (Activation)       (None, 4, 14, 14, 60) 0           batch_normalization_25[0][0]     \n",
      "____________________________________________________________________________________________________\n",
      "conv3d_25 (Conv3D)               (None, 4, 14, 14, 20) 10820       activation_26[0][0]              \n",
      "____________________________________________________________________________________________________\n",
      "concatenate_19 (Concatenate)     (None, 4, 14, 14, 80) 0           concatenate_18[0][0]             \n",
      "                                                                   conv3d_25[0][0]                  \n",
      "____________________________________________________________________________________________________\n",
      "batch_normalization_26 (BatchNor (None, 4, 14, 14, 80) 320         concatenate_19[0][0]             \n",
      "____________________________________________________________________________________________________\n",
      "activation_27 (Activation)       (None, 4, 14, 14, 80) 0           batch_normalization_26[0][0]     \n",
      "____________________________________________________________________________________________________\n",
      "conv3d_26 (Conv3D)               (None, 4, 14, 14, 20) 14420       activation_27[0][0]              \n",
      "____________________________________________________________________________________________________\n",
      "concatenate_20 (Concatenate)     (None, 4, 14, 14, 100 0           concatenate_19[0][0]             \n",
      "                                                                   conv3d_26[0][0]                  \n",
      "____________________________________________________________________________________________________\n",
      "batch_normalization_27 (BatchNor (None, 4, 14, 14, 100 400         concatenate_20[0][0]             \n",
      "____________________________________________________________________________________________________\n",
      "activation_28 (Activation)       (None, 4, 14, 14, 100 0           batch_normalization_27[0][0]     \n",
      "____________________________________________________________________________________________________\n",
      "conv3d_27 (Conv3D)               (None, 4, 14, 14, 10) 1010        activation_28[0][0]              \n",
      "____________________________________________________________________________________________________\n",
      "average_pooling3d_4 (AveragePool (None, 4, 7, 7, 10)   0           conv3d_27[0][0]                  \n",
      "____________________________________________________________________________________________________\n",
      "batch_normalization_28 (BatchNor (None, 4, 7, 7, 10)   40          average_pooling3d_4[0][0]        \n",
      "____________________________________________________________________________________________________\n",
      "activation_29 (Activation)       (None, 4, 7, 7, 10)   0           batch_normalization_28[0][0]     \n",
      "____________________________________________________________________________________________________\n",
      "conv3d_28 (Conv3D)               (None, 4, 7, 7, 10)   910         activation_29[0][0]              \n",
      "____________________________________________________________________________________________________\n",
      "concatenate_21 (Concatenate)     (None, 4, 7, 7, 20)   0           average_pooling3d_4[0][0]        \n",
      "                                                                   conv3d_28[0][0]                  \n",
      "____________________________________________________________________________________________________\n",
      "batch_normalization_29 (BatchNor (None, 4, 7, 7, 20)   80          concatenate_21[0][0]             \n",
      "____________________________________________________________________________________________________\n",
      "activation_30 (Activation)       (None, 4, 7, 7, 20)   0           batch_normalization_29[0][0]     \n",
      "____________________________________________________________________________________________________\n",
      "conv3d_29 (Conv3D)               (None, 4, 7, 7, 10)   1810        activation_30[0][0]              \n",
      "____________________________________________________________________________________________________\n",
      "concatenate_22 (Concatenate)     (None, 4, 7, 7, 30)   0           concatenate_21[0][0]             \n",
      "                                                                   conv3d_29[0][0]                  \n",
      "____________________________________________________________________________________________________\n",
      "batch_normalization_30 (BatchNor (None, 4, 7, 7, 30)   120         concatenate_22[0][0]             \n",
      "____________________________________________________________________________________________________\n",
      "activation_31 (Activation)       (None, 4, 7, 7, 30)   0           batch_normalization_30[0][0]     \n",
      "____________________________________________________________________________________________________\n",
      "conv3d_30 (Conv3D)               (None, 4, 7, 7, 10)   2710        activation_31[0][0]              \n",
      "____________________________________________________________________________________________________\n",
      "concatenate_23 (Concatenate)     (None, 4, 7, 7, 40)   0           concatenate_22[0][0]             \n",
      "                                                                   conv3d_30[0][0]                  \n",
      "____________________________________________________________________________________________________\n",
      "batch_normalization_31 (BatchNor (None, 4, 7, 7, 40)   160         concatenate_23[0][0]             \n",
      "____________________________________________________________________________________________________\n",
      "activation_32 (Activation)       (None, 4, 7, 7, 40)   0           batch_normalization_31[0][0]     \n",
      "____________________________________________________________________________________________________\n",
      "conv3d_31 (Conv3D)               (None, 4, 7, 7, 10)   3610        activation_32[0][0]              \n",
      "____________________________________________________________________________________________________\n",
      "concatenate_24 (Concatenate)     (None, 4, 7, 7, 50)   0           concatenate_23[0][0]             \n",
      "                                                                   conv3d_31[0][0]                  \n",
      "____________________________________________________________________________________________________\n",
      "batch_normalization_32 (BatchNor (None, 4, 7, 7, 50)   200         concatenate_24[0][0]             \n",
      "____________________________________________________________________________________________________\n",
      "activation_33 (Activation)       (None, 4, 7, 7, 50)   0           batch_normalization_32[0][0]     \n",
      "____________________________________________________________________________________________________\n",
      "conv3d_32 (Conv3D)               (None, 4, 1, 1, 37)   90687       activation_33[0][0]              \n",
      "____________________________________________________________________________________________________\n",
      "global_average_pooling3d_2 (Glob (None, 37)            0           conv3d_32[0][0]                  \n",
      "____________________________________________________________________________________________________\n",
      "activation_34 (Activation)       (None, 37)            0           global_average_pooling3d_2[0][0] \n",
      "====================================================================================================\n",
      "Total params: 360,557\n",
      "Trainable params: 357,897\n",
      "Non-trainable params: 2,660\n",
      "____________________________________________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "IMAGE_SIZE = 112\n",
    "model = DenseNet(IMAGE_SIZE, k=40, bn=None, theta=0.5, drop=None, B=3, L=4, outs=37)\n",
    "model.summary()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Build Data Generator"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {},
   "outputs": [],
   "source": [
    "class StandardIterator(Iterator):\n",
    "    def __init__(self, primary_dir, primary_files, labels=None,\n",
    "                 batch_size=64, shuffle=True, seed=None,\n",
    "                 secondary_dir=None, secondary_files=None):\n",
    "        self.primary_dir = primary_dir\n",
    "        self.file_names = primary_files + secondary_files if secondary_files else primary_files\n",
    "        self.boundary = len(primary_files)\n",
    "        self.secondary_dir = secondary_dir\n",
    "        self.labels = labels\n",
    "        super(StandardIterator, self).__init__(len(self.file_names), batch_size, shuffle, seed)\n",
    "    \n",
    "    def update(self, labels):\n",
    "        with self.lock:\n",
    "            self.labels = labels\n",
    "    \n",
    "    def next(self):\n",
    "        with self.lock:\n",
    "            index_array, current_index, current_batch_size = next(self.index_generator)\n",
    "            batch_x = np.zeros((current_batch_size,4,IMAGE_SIZE,IMAGE_SIZE,3), dtype=K.floatx())\n",
    "            batch_y = np.zeros((current_batch_size,37), dtype=K.floatx()) if self.labels else None\n",
    "        \n",
    "        for i, j in enumerate(index_array):\n",
    "            is_primary = j < self.boundary\n",
    "            fname = self.file_names[j]\n",
    "            file_dir = self.primary_dir if is_primary else self.secondary_dir\n",
    "            fpath = os.path.join(file_dir, fname)\n",
    "            img = io.imread(fpath)\n",
    "            img = img[70:-70,70:-70]\n",
    "            img = resize(img, (IMAGE_SIZE,IMAGE_SIZE), mode='reflect')\n",
    "            x = img_as_float(img)\n",
    "            arrs = [np.rot90(x, a) for a in range(4)]\n",
    "            x = np.stack(arrs,0)\n",
    "            batch_x[i] = x\n",
    "            if not (batch_y is None):\n",
    "                batch_y[i] = np.reshape(self.labels[int(fname[:-4])], (-1,37))\n",
    "        return (batch_x, batch_y) if (not (batch_y is None)) else batch_x"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "class MixedIterator(object):\n",
    "    def __init__(self, train_dir, train_files, validation_files,\n",
    "                 test_dir, test_files, train_labels, batch_split):\n",
    "        self.train_dir = train_dir\n",
    "        self.test_dir = test_dir\n",
    "        self.train_files = train_files\n",
    "        self.validation_files = validation_files\n",
    "        self.test_files = test_files\n",
    "        self.train_labels = train_labels\n",
    "        self.pseudo_labels = {}\n",
    "        self.train_gen = StandardIterator(train_dir, train_files, train_labels, batch_size=batch_split[0])\n",
    "        self.pseudo_gen = StandardIterator(train_dir, validation_files, batch_size=batch_split[1],\n",
    "                                           secondary_dir=self.test_dir, secondary_files=self.test_files)\n",
    "        self.lock = threading.Lock()\n",
    "    \n",
    "    def reset(self):\n",
    "        self.train_gen.reset()\n",
    "        self.pseudo_gen.reset()\n",
    "        \n",
    "    def get_pseudo_labels(self):\n",
    "        return self.pseudo_labels\n",
    "    \n",
    "    def update(self, model):\n",
    "        with self.lock:\n",
    "            pred_gen = StandardIterator(self.train_dir, self.validation_files, shuffle=False, batch_size=60)\n",
    "            new_labels = model.predict_generator(pred_gen, len(self.validation_files)//60, verbose=1)\n",
    "            for i, j in enumerate(self.validation_files):\n",
    "                name = int(j[:-4])\n",
    "                self.pseudo_labels[name] = new_labels[i,:]\n",
    "            pred_gen = StandardIterator(self.test_dir, self.test_files, shuffle=False, batch_size=35)\n",
    "            new_labels = model.predict_generator(pred_gen, len(self.test_files)//35, verbose=1)\n",
    "            for i, j in enumerate(self.test_files):\n",
    "                name = int(j[:-4])\n",
    "                self.pseudo_labels[name] = new_labels[i,:]\n",
    "            self.pseudo_gen.update(self.pseudo_labels)\n",
    "            \n",
    "    def __next__(self):\n",
    "        return self.next()\n",
    "    \n",
    "    def next(self):\n",
    "        with self.lock:\n",
    "            a = next(self.train_gen)\n",
    "            if len(self.pseudo_labels) > 0:\n",
    "                b = next(self.pseudo_gen)\n",
    "                return (np.concatenate((a[0],b[0])), np.concatenate((a[1],b[1])))\n",
    "            else:\n",
    "                return a"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Train Model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "def rmse(y_true, y_pred):\n",
    "    return K.sqrt(losses.mean_squared_error(y_true, y_pred))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {},
   "outputs": [],
   "source": [
    "BATCH_SIZE = 32\n",
    "TRAIN_BATCH_SIZE = 22\n",
    "PSEUDO_BATCH_SIZE = BATCH_SIZE - TRAIN_BATCH_SIZE\n",
    "\n",
    "train_generator = StandardIterator(TRAIN, train_files, train_labels, batch_size=BATCH_SIZE)\n",
    "pseudo_generator = MixedIterator(TRAIN, train_files, val_files, TEST, test_files, train_labels,\n",
    "                                 batch_split=(TRAIN_BATCH_SIZE, PSEUDO_BATCH_SIZE))\n",
    "val_generator = StandardIterator(TRAIN, val_files, train_labels, batch_size=BATCH_SIZE, shuffle=False)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 1/30\n",
      "1736/1736 [==============================] - 404s - loss: 0.0955 - val_loss: 0.1143\n",
      "Epoch 2/30\n",
      "1736/1736 [==============================] - 402s - loss: 0.0916 - val_loss: 0.1265\n",
      "Epoch 3/30\n",
      "1736/1736 [==============================] - 401s - loss: 0.0894 - val_loss: 0.0955\n",
      "Epoch 4/30\n",
      "1736/1736 [==============================] - 403s - loss: 0.0873 - val_loss: 0.0909\n",
      "Epoch 5/30\n",
      "1736/1736 [==============================] - 401s - loss: 0.0861 - val_loss: 0.0891\n",
      "Epoch 6/30\n",
      "1736/1736 [==============================] - 401s - loss: 0.0847 - val_loss: 0.0908\n",
      "Epoch 7/30\n",
      "1736/1736 [==============================] - 401s - loss: 0.0837 - val_loss: 0.0909\n",
      "Epoch 8/30\n",
      "1736/1736 [==============================] - 401s - loss: 0.0826 - val_loss: 0.0856\n",
      "Epoch 9/30\n",
      "1736/1736 [==============================] - 402s - loss: 0.0817 - val_loss: 0.0854\n",
      "Epoch 10/30\n",
      "1736/1736 [==============================] - 400s - loss: 0.0810 - val_loss: 0.0883\n",
      "Epoch 11/30\n",
      "1736/1736 [==============================] - 401s - loss: 0.0801 - val_loss: 0.0851\n",
      "Epoch 12/30\n",
      "1736/1736 [==============================] - 401s - loss: 0.0795 - val_loss: 0.0844\n",
      "Epoch 13/30\n",
      "1736/1736 [==============================] - 401s - loss: 0.0788 - val_loss: 0.0976\n",
      "Epoch 14/30\n",
      "1736/1736 [==============================] - 402s - loss: 0.0784 - val_loss: 0.0822\n",
      "Epoch 15/30\n",
      "1736/1736 [==============================] - 401s - loss: 0.0777 - val_loss: 0.0878\n",
      "Epoch 16/30\n",
      "1736/1736 [==============================] - 402s - loss: 0.0772 - val_loss: 0.0902\n",
      "Epoch 17/30\n",
      "1736/1736 [==============================] - 402s - loss: 0.0767 - val_loss: 0.0843\n",
      "Epoch 18/30\n",
      "1736/1736 [==============================] - 401s - loss: 0.0762 - val_loss: 0.0840\n",
      "Epoch 19/30\n",
      "1736/1736 [==============================] - 402s - loss: 0.0759 - val_loss: 0.0827\n",
      "Epoch 20/30\n",
      "1736/1736 [==============================] - 402s - loss: 0.0749 - val_loss: 0.0828\n",
      "Epoch 21/30\n",
      "1736/1736 [==============================] - 402s - loss: 0.0750 - val_loss: 0.0830\n",
      "Epoch 22/30\n",
      "1736/1736 [==============================] - 402s - loss: 0.0744 - val_loss: 0.0863\n",
      "Epoch 23/30\n",
      "1736/1736 [==============================] - 402s - loss: 0.0740 - val_loss: 0.0826\n",
      "Epoch 24/30\n",
      "1736/1736 [==============================] - 401s - loss: 0.0736 - val_loss: 0.0859\n",
      "Epoch 25/30\n",
      "1736/1736 [==============================] - 401s - loss: 0.0732 - val_loss: 0.0826\n",
      "Epoch 26/30\n",
      "1736/1736 [==============================] - 402s - loss: 0.0729 - val_loss: 0.0863\n",
      "Epoch 27/30\n",
      "1736/1736 [==============================] - 403s - loss: 0.0723 - val_loss: 0.0833\n",
      "Epoch 28/30\n",
      "1736/1736 [==============================] - 401s - loss: 0.0721 - val_loss: 0.0860\n",
      "Epoch 29/30\n",
      "1736/1736 [==============================] - 403s - loss: 0.0718 - val_loss: 0.0836\n",
      "Epoch 30/30\n",
      "1736/1736 [==============================] - 402s - loss: 0.0715 - val_loss: 0.0828\n"
     ]
    }
   ],
   "source": [
    "model.compile(optimizer=Adam(0.001), loss=rmse)\n",
    "training_progress = model.fit_generator(train_generator,\n",
    "                        steps_per_epoch=len(train_files)//BATCH_SIZE,\n",
    "                        epochs=30,\n",
    "                        validation_data=val_generator,\n",
    "                        validation_steps=len(val_files)//BATCH_SIZE,\n",
    "                        max_queue_size=10)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "[<matplotlib.lines.Line2D at 0x7f46d6cdeeb8>]"
      ]
     },
     "execution_count": 24,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAX4AAAD8CAYAAABw1c+bAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nUEqOAAAAAElFTkSuQmCC\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7f46d6cdee80>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "plt.plot(training_progress.epoch, training_progress.history['loss'], label='Train')\n",
    "plt.plot(training_progress.epoch, training_progress.history['val_loss'], label='Validation')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "100/100 [==============================] - 33s    \n",
      "2285/2285 [==============================] - 479s   \n",
      "Epoch 1/5\n",
      "2526/2526 [==============================] - 570s - loss: 0.0566 - val_loss: 0.0829\n",
      "Epoch 2/5\n",
      "2526/2526 [==============================] - 569s - loss: 0.0564 - val_loss: 0.0848\n",
      "Epoch 3/5\n",
      "2526/2526 [==============================] - 568s - loss: 0.0561 - val_loss: 0.0886\n",
      "Epoch 4/5\n",
      "2526/2526 [==============================] - 568s - loss: 0.0560 - val_loss: 0.0819\n",
      "Epoch 5/5\n",
      "2526/2526 [==============================] - 569s - loss: 0.0557 - val_loss: 0.0915\n"
     ]
    }
   ],
   "source": [
    "pseudo_generator.update(model)\n",
    "training_progress = model.fit_generator(pseudo_generator,\n",
    "                        steps_per_epoch=len(train_files)//TRAIN_BATCH_SIZE,\n",
    "                        epochs=5,\n",
    "                        validation_data=val_generator,\n",
    "                        validation_steps=len(val_files)//BATCH_SIZE,\n",
    "                        max_queue_size=10)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 26,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "[<matplotlib.lines.Line2D at 0x7f4646593908>]"
      ]
     },
     "execution_count": 26,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAYQAAAD8CAYAAAB3u9PLAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\n5D7m+3YAAAAASUVORK5CYII=\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7f4646593780>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "plt.plot(training_progress.epoch, training_progress.history['loss'], label='Train')\n",
    "plt.plot(training_progress.epoch, training_progress.history['val_loss'], label='Validation')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 1/10\n",
      "1736/1736 [==============================] - 403s - loss: 0.0668 - val_loss: 0.0808\n",
      "Epoch 2/10\n",
      "1736/1736 [==============================] - 403s - loss: 0.0657 - val_loss: 0.0810\n",
      "Epoch 3/10\n",
      "1736/1736 [==============================] - 402s - loss: 0.0653 - val_loss: 0.0809\n",
      "Epoch 4/10\n",
      "1736/1736 [==============================] - 403s - loss: 0.0651 - val_loss: 0.0809\n",
      "Epoch 5/10\n",
      "1736/1736 [==============================] - 402s - loss: 0.0648 - val_loss: 0.0810\n",
      "Epoch 6/10\n",
      "1736/1736 [==============================] - 402s - loss: 0.0646 - val_loss: 0.0811\n",
      "Epoch 7/10\n",
      "1736/1736 [==============================] - 401s - loss: 0.0645 - val_loss: 0.0812\n",
      "Epoch 8/10\n",
      "1736/1736 [==============================] - 402s - loss: 0.0643 - val_loss: 0.0813\n",
      "Epoch 9/10\n",
      "1736/1736 [==============================] - 401s - loss: 0.0641 - val_loss: 0.0813\n",
      "Epoch 10/10\n",
      "1736/1736 [==============================] - 403s - loss: 0.0641 - val_loss: 0.0818\n"
     ]
    }
   ],
   "source": [
    "model.compile(optimizer=Adam(0.0001), loss=rmse)\n",
    "training_progress = model.fit_generator(train_generator,\n",
    "                        steps_per_epoch=len(train_files)//BATCH_SIZE,\n",
    "                        epochs=10,\n",
    "                        validation_data=val_generator,\n",
    "                        validation_steps=len(val_files)//BATCH_SIZE,\n",
    "                        max_queue_size=10)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 28,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "[<matplotlib.lines.Line2D at 0x7f46458ce4a8>]"
      ]
     },
     "execution_count": 28,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAYoAAAD8CAYAAABpcuN4AAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nmJlZLgeFmZnlclCYmVmu/w/1mQZf6I2tpwAAAABJRU5ErkJggg==\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7f46458ce860>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "plt.plot(training_progress.epoch, training_progress.history['loss'], label='Train')\n",
    "plt.plot(training_progress.epoch, training_progress.history['val_loss'], label='Validation')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 29,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "100/100 [==============================] - 33s    \n",
      "2285/2285 [==============================] - 443s   \n",
      "Epoch 1/5\n",
      "2526/2526 [==============================] - 569s - loss: 0.0478 - val_loss: 0.0819\n",
      "Epoch 2/5\n",
      "2526/2526 [==============================] - 569s - loss: 0.0477 - val_loss: 0.0819\n",
      "Epoch 3/5\n",
      "2526/2526 [==============================] - 569s - loss: 0.0476 - val_loss: 0.0817\n",
      "Epoch 4/5\n",
      "2526/2526 [==============================] - 570s - loss: 0.0476 - val_loss: 0.0815\n",
      "Epoch 5/5\n",
      "2526/2526 [==============================] - 569s - loss: 0.0475 - val_loss: 0.0818\n"
     ]
    }
   ],
   "source": [
    "pseudo_generator.reset()\n",
    "pseudo_generator.update(model)\n",
    "training_progress = model.fit_generator(pseudo_generator,\n",
    "                        steps_per_epoch=len(train_files)//TRAIN_BATCH_SIZE,\n",
    "                        epochs=5,\n",
    "                        validation_data=val_generator,\n",
    "                        validation_steps=len(val_files)//BATCH_SIZE,\n",
    "                        max_queue_size=10)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 30,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "[<matplotlib.lines.Line2D at 0x7f4646598550>]"
      ]
     },
     "execution_count": 30,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAYQAAAD8CAYAAAB3u9PLAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nSUVORK5CYII=\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7f4646598438>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "plt.plot(training_progress.epoch, training_progress.history['loss'], label='Train')\n",
    "plt.plot(training_progress.epoch, training_progress.history['val_loss'], label='Validation')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Predict Outputs"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "test_size = len(test_files)\n",
    "preds = model.predict_generator(StandardIterator(TEST, test_files, batch_size=35, shuffle=False), steps=test_size//35, verbose=1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "train_headers = pd.read_csv(LABELS)\n",
    "\n",
    "files2 = sorted(os.listdir(TEST))\n",
    "files2 = list(map(lambda x: int(x[:-4]), files2))\n",
    "\n",
    "outputcsv = pd.DataFrame(data=np.c_[files2, preds],\n",
    "                         index=np.arange(len(files2)),\n",
    "                         columns=train_headers.columns)\n",
    "outputcsv[['GalaxyID']] = outputcsv[['GalaxyID']].astype(int)\n",
    "outputcsv.to_csv('./predictions.csv', index=False)"
   ]
  },
  {
   "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": "3.6.1"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}