2 - Using pre-trained VGG16.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/home/ubuntu/anaconda3/envs/tensorflow_p27/lib/python2.7/site-packages/h5py/__init__.py:36: FutureWarning: Conversion of the second argument of issubdtype from `float` to `np.floating` is deprecated. In future, it will be treated as `np.float64 == np.dtype(float).type`.\n",
      "  from ._conv import register_converters as _register_converters\n",
      "Using TensorFlow backend.\n",
      "/home/ubuntu/anaconda3/envs/tensorflow_p27/lib/python2.7/site-packages/matplotlib/__init__.py:962: UserWarning: Duplicate key in file \"/home/ubuntu/.config/matplotlib/matplotlibrc\", line #2\n",
      "  (fname, cnt))\n",
      "/home/ubuntu/anaconda3/envs/tensorflow_p27/lib/python2.7/site-packages/matplotlib/__init__.py:962: UserWarning: Duplicate key in file \"/home/ubuntu/.config/matplotlib/matplotlibrc\", line #3\n",
      "  (fname, cnt))\n"
     ]
    }
   ],
   "source": [
    "'''This script goes along the blog post\n",
    "\"Building powerful image classification models using very little data\"\n",
    "from blog.keras.io.\n",
    "It uses data that can be downloaded at:\n",
    "https://www.kaggle.com/c/dogs-vs-cats/data\n",
    "In our setup, we:\n",
    "- created a data/ folder\n",
    "- created train/ and validation/ subfolders inside data/\n",
    "- created cats/ and dogs/ subfolders inside train/ and validation/\n",
    "- put the cat pictures index 0-999 in data/train/cats\n",
    "- put the cat pictures index 1000-1400 in data/validation/cats\n",
    "- put the dogs pictures index 12500-13499 in data/train/dogs\n",
    "- put the dog pictures index 13500-13900 in data/validation/dogs\n",
    "So that we have 1000 training examples for each class, and 400 validation examples for each class.\n",
    "In summary, this is our directory structure:\n",
    "```\n",
    "data/\n",
    "    train/\n",
    "        dogs/\n",
    "            dog001.jpg\n",
    "            dog002.jpg\n",
    "            ...\n",
    "        cats/\n",
    "            cat001.jpg\n",
    "            cat002.jpg\n",
    "            ...\n",
    "    validation/\n",
    "        dogs/\n",
    "            dog001.jpg\n",
    "            dog002.jpg\n",
    "            ...\n",
    "        cats/\n",
    "            cat001.jpg\n",
    "            cat002.jpg\n",
    "            ...\n",
    "```\n",
    "'''\n",
    "import numpy as np\n",
    "from keras.preprocessing.image import ImageDataGenerator\n",
    "from keras.models import Sequential\n",
    "from keras.layers import Dropout, Flatten, Dense\n",
    "from keras import applications\n",
    "from keras.applications import VGG16\n",
    "\n",
    "# dimensions of our images.\n",
    "img_width, img_height = 150, 150\n",
    "\n",
    "top_model_weights_path = 'bottleneck_fc_model.h5'\n",
    "train_data_dir = 'data/train'\n",
    "validation_data_dir = 'data/validation'\n",
    "nb_train_samples = 2000\n",
    "nb_validation_samples = 800\n",
    "epochs = 50\n",
    "batch_size = 16"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "input_2 (InputLayer)         (None, None, None, 3)     0         \n",
      "_________________________________________________________________\n",
      "block1_conv1 (Conv2D)        (None, None, None, 64)    1792      \n",
      "_________________________________________________________________\n",
      "block1_conv2 (Conv2D)        (None, None, None, 64)    36928     \n",
      "_________________________________________________________________\n",
      "block1_pool (MaxPooling2D)   (None, None, None, 64)    0         \n",
      "_________________________________________________________________\n",
      "block2_conv1 (Conv2D)        (None, None, None, 128)   73856     \n",
      "_________________________________________________________________\n",
      "block2_conv2 (Conv2D)        (None, None, None, 128)   147584    \n",
      "_________________________________________________________________\n",
      "block2_pool (MaxPooling2D)   (None, None, None, 128)   0         \n",
      "_________________________________________________________________\n",
      "block3_conv1 (Conv2D)        (None, None, None, 256)   295168    \n",
      "_________________________________________________________________\n",
      "block3_conv2 (Conv2D)        (None, None, None, 256)   590080    \n",
      "_________________________________________________________________\n",
      "block3_conv3 (Conv2D)        (None, None, None, 256)   590080    \n",
      "_________________________________________________________________\n",
      "block3_pool (MaxPooling2D)   (None, None, None, 256)   0         \n",
      "_________________________________________________________________\n",
      "block4_conv1 (Conv2D)        (None, None, None, 512)   1180160   \n",
      "_________________________________________________________________\n",
      "block4_conv2 (Conv2D)        (None, None, None, 512)   2359808   \n",
      "_________________________________________________________________\n",
      "block4_conv3 (Conv2D)        (None, None, None, 512)   2359808   \n",
      "_________________________________________________________________\n",
      "block4_pool (MaxPooling2D)   (None, None, None, 512)   0         \n",
      "_________________________________________________________________\n",
      "block5_conv1 (Conv2D)        (None, None, None, 512)   2359808   \n",
      "_________________________________________________________________\n",
      "block5_conv2 (Conv2D)        (None, None, None, 512)   2359808   \n",
      "_________________________________________________________________\n",
      "block5_conv3 (Conv2D)        (None, None, None, 512)   2359808   \n",
      "_________________________________________________________________\n",
      "block5_pool (MaxPooling2D)   (None, None, None, 512)   0         \n",
      "=================================================================\n",
      "Total params: 14,714,688\n",
      "Trainable params: 14,714,688\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n",
      "None\n"
     ]
    }
   ],
   "source": [
    "datagen = ImageDataGenerator(rescale=1. / 255)\n",
    "model = VGG16(include_top=False, weights='imagenet')\n",
    "print model.summary()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "**PART 1 - Generate the output for the Convolutional Neural Network**\n",
    "\n",
    "Notice that we are saving the output from predict_generator into \"bottleneck_features_train.npy\" and \"bottleneck_features_validation.npy\". "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "generator = datagen.flow_from_directory(\n",
    "    train_data_dir,\n",
    "    target_size=(img_width, img_height),\n",
    "    batch_size=batch_size,\n",
    "    class_mode=None,\n",
    "    shuffle=False)\n",
    "\n",
    "bottleneck_features_train = model.predict_generator(generator, nb_train_samples // batch_size)\n",
    "np.save(open('bottleneck_features_train.npy', 'w'), bottleneck_features_train)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "generator = datagen.flow_from_directory(\n",
    "    validation_data_dir,\n",
    "    target_size=(img_width, img_height),\n",
    "    batch_size=batch_size,\n",
    "    class_mode=None,\n",
    "    shuffle=False)\n",
    "\n",
    "bottleneck_features_validation = model.predict_generator(generator, nb_validation_samples // batch_size)\n",
    "np.save(open('bottleneck_features_validation.npy', 'w'), bottleneck_features_validation)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "**PART 2 - Train the fully connected network**"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(2000, 4, 4, 512)\n",
      "(800, 4, 4, 512)\n"
     ]
    }
   ],
   "source": [
    "# Train top model\n",
    "train_data = np.load(open('bottleneck_features_train.npy'))\n",
    "train_labels = np.array([0] * (nb_train_samples / 2) + [1] * (nb_train_samples / 2))\n",
    "\n",
    "validation_data = np.load(open('bottleneck_features_validation.npy'))\n",
    "validation_labels = np.array([0] * (nb_validation_samples / 2) + [1] * (nb_validation_samples / 2))\n",
    "\n",
    "print train_data.shape\n",
    "print validation_data.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Train on 2000 samples, validate on 800 samples\n",
      "Epoch 1/50\n",
      "2000/2000 [==============================] - 1s 593us/step - loss: 0.6727 - acc: 0.7705 - val_loss: 0.2750 - val_acc: 0.8912\n",
      "Epoch 2/50\n",
      "2000/2000 [==============================] - 1s 444us/step - loss: 0.3633 - acc: 0.8520 - val_loss: 0.2807 - val_acc: 0.8812\n",
      "Epoch 3/50\n",
      "2000/2000 [==============================] - 1s 444us/step - loss: 0.2972 - acc: 0.8795 - val_loss: 0.2415 - val_acc: 0.9087\n",
      "Epoch 4/50\n",
      "2000/2000 [==============================] - 1s 440us/step - loss: 0.2607 - acc: 0.8970 - val_loss: 0.2614 - val_acc: 0.8988\n",
      "Epoch 5/50\n",
      "2000/2000 [==============================] - 1s 497us/step - loss: 0.2137 - acc: 0.9190 - val_loss: 0.4860 - val_acc: 0.8237\n",
      "Epoch 6/50\n",
      "2000/2000 [==============================] - 1s 487us/step - loss: 0.2084 - acc: 0.9260 - val_loss: 0.3062 - val_acc: 0.8950\n",
      "Epoch 7/50\n",
      "2000/2000 [==============================] - 1s 461us/step - loss: 0.1931 - acc: 0.9315 - val_loss: 0.3913 - val_acc: 0.8650\n",
      "Epoch 8/50\n",
      "2000/2000 [==============================] - 1s 452us/step - loss: 0.1534 - acc: 0.9380 - val_loss: 0.3031 - val_acc: 0.9050\n",
      "Epoch 9/50\n",
      "2000/2000 [==============================] - 1s 432us/step - loss: 0.1524 - acc: 0.9470 - val_loss: 0.3807 - val_acc: 0.8862\n",
      "Epoch 10/50\n",
      "2000/2000 [==============================] - 1s 431us/step - loss: 0.1234 - acc: 0.9495 - val_loss: 0.4787 - val_acc: 0.8675\n",
      "Epoch 11/50\n",
      "2000/2000 [==============================] - 1s 430us/step - loss: 0.1097 - acc: 0.9580 - val_loss: 0.3980 - val_acc: 0.8925\n",
      "Epoch 12/50\n",
      "2000/2000 [==============================] - 1s 427us/step - loss: 0.1118 - acc: 0.9570 - val_loss: 0.4078 - val_acc: 0.9000\n",
      "Epoch 13/50\n",
      "2000/2000 [==============================] - 1s 427us/step - loss: 0.0819 - acc: 0.9680 - val_loss: 0.4123 - val_acc: 0.8938\n",
      "Epoch 14/50\n",
      "2000/2000 [==============================] - 1s 434us/step - loss: 0.0790 - acc: 0.9695 - val_loss: 0.4803 - val_acc: 0.8975\n",
      "Epoch 15/50\n",
      "2000/2000 [==============================] - 1s 437us/step - loss: 0.0721 - acc: 0.9720 - val_loss: 0.4371 - val_acc: 0.9038\n",
      "Epoch 16/50\n",
      "2000/2000 [==============================] - 1s 432us/step - loss: 0.0712 - acc: 0.9715 - val_loss: 0.5592 - val_acc: 0.9000\n",
      "Epoch 17/50\n",
      "2000/2000 [==============================] - 1s 442us/step - loss: 0.0598 - acc: 0.9765 - val_loss: 0.5341 - val_acc: 0.8950\n",
      "Epoch 18/50\n",
      "2000/2000 [==============================] - 1s 424us/step - loss: 0.0549 - acc: 0.9825 - val_loss: 0.7652 - val_acc: 0.8675\n",
      "Epoch 19/50\n",
      "2000/2000 [==============================] - 1s 438us/step - loss: 0.0461 - acc: 0.9820 - val_loss: 0.4808 - val_acc: 0.9025\n",
      "Epoch 20/50\n",
      "2000/2000 [==============================] - 1s 438us/step - loss: 0.0427 - acc: 0.9845 - val_loss: 0.5831 - val_acc: 0.8988\n",
      "Epoch 21/50\n",
      "2000/2000 [==============================] - 1s 428us/step - loss: 0.0324 - acc: 0.9890 - val_loss: 0.5708 - val_acc: 0.8962\n",
      "Epoch 22/50\n",
      "2000/2000 [==============================] - 1s 430us/step - loss: 0.0325 - acc: 0.9890 - val_loss: 0.7020 - val_acc: 0.8862\n",
      "Epoch 23/50\n",
      "2000/2000 [==============================] - 1s 428us/step - loss: 0.0471 - acc: 0.9845 - val_loss: 0.6452 - val_acc: 0.8875\n",
      "Epoch 24/50\n",
      "2000/2000 [==============================] - 1s 428us/step - loss: 0.0187 - acc: 0.9930 - val_loss: 0.8209 - val_acc: 0.8888\n",
      "Epoch 25/50\n",
      "2000/2000 [==============================] - 1s 433us/step - loss: 0.0305 - acc: 0.9885 - val_loss: 0.6564 - val_acc: 0.8925\n",
      "Epoch 26/50\n",
      "2000/2000 [==============================] - 1s 449us/step - loss: 0.0225 - acc: 0.9925 - val_loss: 0.6778 - val_acc: 0.8962\n",
      "Epoch 27/50\n",
      "2000/2000 [==============================] - 1s 430us/step - loss: 0.0227 - acc: 0.9925 - val_loss: 0.7210 - val_acc: 0.8900\n",
      "Epoch 28/50\n",
      "2000/2000 [==============================] - 1s 442us/step - loss: 0.0289 - acc: 0.9880 - val_loss: 0.6965 - val_acc: 0.8975\n",
      "Epoch 29/50\n",
      "2000/2000 [==============================] - 1s 444us/step - loss: 0.0329 - acc: 0.9905 - val_loss: 0.7614 - val_acc: 0.8962\n",
      "Epoch 30/50\n",
      "2000/2000 [==============================] - 1s 445us/step - loss: 0.0185 - acc: 0.9930 - val_loss: 0.7955 - val_acc: 0.9000\n",
      "Epoch 31/50\n",
      "2000/2000 [==============================] - 1s 442us/step - loss: 0.0294 - acc: 0.9915 - val_loss: 0.7204 - val_acc: 0.8975\n",
      "Epoch 32/50\n",
      "2000/2000 [==============================] - 1s 436us/step - loss: 0.0277 - acc: 0.9915 - val_loss: 0.7362 - val_acc: 0.8988\n",
      "Epoch 33/50\n",
      "2000/2000 [==============================] - 1s 441us/step - loss: 0.0252 - acc: 0.9925 - val_loss: 0.7711 - val_acc: 0.9012\n",
      "Epoch 34/50\n",
      "2000/2000 [==============================] - 1s 424us/step - loss: 0.0191 - acc: 0.9920 - val_loss: 0.8754 - val_acc: 0.8938\n",
      "Epoch 35/50\n",
      "2000/2000 [==============================] - 1s 451us/step - loss: 0.0219 - acc: 0.9945 - val_loss: 0.8236 - val_acc: 0.9000\n",
      "Epoch 36/50\n",
      "2000/2000 [==============================] - 1s 441us/step - loss: 0.0216 - acc: 0.9925 - val_loss: 0.7772 - val_acc: 0.8950\n",
      "Epoch 37/50\n",
      "2000/2000 [==============================] - 1s 435us/step - loss: 0.0130 - acc: 0.9960 - val_loss: 0.8687 - val_acc: 0.8900\n",
      "Epoch 38/50\n",
      "2000/2000 [==============================] - 1s 438us/step - loss: 0.0361 - acc: 0.9920 - val_loss: 0.8328 - val_acc: 0.9012\n",
      "Epoch 39/50\n",
      "2000/2000 [==============================] - 1s 446us/step - loss: 0.0128 - acc: 0.9960 - val_loss: 1.0353 - val_acc: 0.8825\n",
      "Epoch 40/50\n",
      "2000/2000 [==============================] - 1s 435us/step - loss: 0.0074 - acc: 0.9975 - val_loss: 0.9737 - val_acc: 0.8912\n",
      "Epoch 41/50\n",
      "2000/2000 [==============================] - 1s 435us/step - loss: 0.0139 - acc: 0.9965 - val_loss: 0.9904 - val_acc: 0.8850\n",
      "Epoch 42/50\n",
      "2000/2000 [==============================] - 1s 437us/step - loss: 0.0318 - acc: 0.9925 - val_loss: 0.9131 - val_acc: 0.9038\n",
      "Epoch 43/50\n",
      "2000/2000 [==============================] - 1s 449us/step - loss: 0.0219 - acc: 0.9930 - val_loss: 0.7781 - val_acc: 0.9000\n",
      "Epoch 44/50\n",
      "2000/2000 [==============================] - 1s 457us/step - loss: 0.0145 - acc: 0.9965 - val_loss: 0.8813 - val_acc: 0.8975\n",
      "Epoch 45/50\n",
      "2000/2000 [==============================] - 1s 440us/step - loss: 0.0187 - acc: 0.9945 - val_loss: 0.9299 - val_acc: 0.8938\n",
      "Epoch 46/50\n",
      "2000/2000 [==============================] - 1s 433us/step - loss: 0.0171 - acc: 0.9960 - val_loss: 0.8352 - val_acc: 0.8925\n",
      "Epoch 47/50\n",
      "2000/2000 [==============================] - 1s 443us/step - loss: 0.0103 - acc: 0.9960 - val_loss: 0.9185 - val_acc: 0.8962\n",
      "Epoch 48/50\n",
      "2000/2000 [==============================] - 1s 439us/step - loss: 0.0035 - acc: 0.9980 - val_loss: 0.9365 - val_acc: 0.9012\n",
      "Epoch 49/50\n",
      "2000/2000 [==============================] - 1s 440us/step - loss: 0.0079 - acc: 0.9975 - val_loss: 0.9359 - val_acc: 0.8950\n",
      "Epoch 50/50\n",
      "2000/2000 [==============================] - 1s 438us/step - loss: 0.0148 - acc: 0.9960 - val_loss: 0.9338 - val_acc: 0.8975\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<keras.callbacks.History at 0x7f1d0ada7bd0>"
      ]
     },
     "execution_count": 3,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "model = Sequential()\n",
    "model.add(Flatten(input_shape=train_data.shape[1:]))\n",
    "model.add(Dense(256, activation='relu'))\n",
    "model.add(Dropout(0.5))\n",
    "model.add(Dense(1, activation='sigmoid'))\n",
    "\n",
    "model.compile(optimizer='rmsprop', loss='binary_crossentropy', metrics=['accuracy'])\n",
    "\n",
    "model.fit(train_data, \n",
    "          train_labels,\n",
    "          epochs=epochs,\n",
    "          batch_size=batch_size,\n",
    "          validation_data=(validation_data, validation_labels))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "**By transfering learning from VGG16, after 50 epochs, we have gotten an accuracy of ~89% in the validation set.**"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "model.save_weights(top_model_weights_path)"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Environment (conda_tensorflow_p27)",
   "language": "python",
   "name": "conda_tensorflow_p27"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 2
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython2",
   "version": "2.7.14"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}