3 - Fine Tuning CNN.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": [
    "from keras import applications\n",
    "from keras.applications import VGG16\n",
    "from keras.preprocessing.image import ImageDataGenerator\n",
    "from keras import optimizers\n",
    "from keras.models import Sequential\n",
    "from keras.models import Model\n",
    "from keras.layers import Dropout, Flatten, Dense"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "# path to the model weights files.\n",
    "weights_path = 'vgg16_weights.h5' # https://gist.github.com/baraldilorenzo/07d7802847aaad0a35d3\n",
    "top_model_weights_path = 'bottleneck_fc_model.h5'"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "# dimensions of our images.\n",
    "img_width, img_height = 150, 150\n",
    "\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": "markdown",
   "metadata": {},
   "source": [
    "**Load VGG16, we will use that in a second**"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Model loaded.\n"
     ]
    }
   ],
   "source": [
    "# build the VGG16 network\n",
    "base_model = VGG16(weights='imagenet', include_top=False, input_shape=(150,150,3))\n",
    "print('Model loaded.')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "**Let's build a neural network and load it with the weights from Part 2**"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "flatten_1 (Flatten)          (None, 8192)              0         \n",
      "_________________________________________________________________\n",
      "dense_1 (Dense)              (None, 256)               2097408   \n",
      "_________________________________________________________________\n",
      "dropout_1 (Dropout)          (None, 256)               0         \n",
      "_________________________________________________________________\n",
      "dense_2 (Dense)              (None, 1)                 257       \n",
      "=================================================================\n",
      "Total params: 2,097,665\n",
      "Trainable params: 2,097,665\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n",
      "None\n"
     ]
    }
   ],
   "source": [
    "# build a classifier model to put on top of the convolutional model\n",
    "top_model = Sequential()\n",
    "top_model.add(Flatten(input_shape=base_model.output_shape[1:]))\n",
    "top_model.add(Dense(256, activation='relu'))\n",
    "top_model.add(Dropout(0.5))\n",
    "top_model.add(Dense(1, activation='sigmoid'))\n",
    "\n",
    "# note that it is necessary to start with a fully-trained\n",
    "# classifier, including the top classifier,\n",
    "# in order to successfully do fine-tuning\n",
    "top_model.load_weights(top_model_weights_path)\n",
    "\n",
    "print top_model.summary()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "input_1 (InputLayer)         (None, 150, 150, 3)       0         \n",
      "_________________________________________________________________\n",
      "block1_conv1 (Conv2D)        (None, 150, 150, 64)      1792      \n",
      "_________________________________________________________________\n",
      "block1_conv2 (Conv2D)        (None, 150, 150, 64)      36928     \n",
      "_________________________________________________________________\n",
      "block1_pool (MaxPooling2D)   (None, 75, 75, 64)        0         \n",
      "_________________________________________________________________\n",
      "block2_conv1 (Conv2D)        (None, 75, 75, 128)       73856     \n",
      "_________________________________________________________________\n",
      "block2_conv2 (Conv2D)        (None, 75, 75, 128)       147584    \n",
      "_________________________________________________________________\n",
      "block2_pool (MaxPooling2D)   (None, 37, 37, 128)       0         \n",
      "_________________________________________________________________\n",
      "block3_conv1 (Conv2D)        (None, 37, 37, 256)       295168    \n",
      "_________________________________________________________________\n",
      "block3_conv2 (Conv2D)        (None, 37, 37, 256)       590080    \n",
      "_________________________________________________________________\n",
      "block3_conv3 (Conv2D)        (None, 37, 37, 256)       590080    \n",
      "_________________________________________________________________\n",
      "block3_pool (MaxPooling2D)   (None, 18, 18, 256)       0         \n",
      "_________________________________________________________________\n",
      "block4_conv1 (Conv2D)        (None, 18, 18, 512)       1180160   \n",
      "_________________________________________________________________\n",
      "block4_conv2 (Conv2D)        (None, 18, 18, 512)       2359808   \n",
      "_________________________________________________________________\n",
      "block4_conv3 (Conv2D)        (None, 18, 18, 512)       2359808   \n",
      "_________________________________________________________________\n",
      "block4_pool (MaxPooling2D)   (None, 9, 9, 512)         0         \n",
      "_________________________________________________________________\n",
      "block5_conv1 (Conv2D)        (None, 9, 9, 512)         2359808   \n",
      "_________________________________________________________________\n",
      "block5_conv2 (Conv2D)        (None, 9, 9, 512)         2359808   \n",
      "_________________________________________________________________\n",
      "block5_conv3 (Conv2D)        (None, 9, 9, 512)         2359808   \n",
      "_________________________________________________________________\n",
      "block5_pool (MaxPooling2D)   (None, 4, 4, 512)         0         \n",
      "_________________________________________________________________\n",
      "sequential_1 (Sequential)    (None, 1)                 2097665   \n",
      "=================================================================\n",
      "Total params: 16,812,353\n",
      "Trainable params: 9,177,089\n",
      "Non-trainable params: 7,635,264\n",
      "_________________________________________________________________\n",
      "None\n"
     ]
    }
   ],
   "source": [
    "# add the model on top of the convolutional base\n",
    "# model.add(top_model)\n",
    "model = Model(inputs=base_model.input, outputs=top_model(base_model.output))\n",
    "\n",
    "# set the first 25 layers (up to the last conv block)\n",
    "# to non-trainable (weights will not be updated)\n",
    "for layer in model.layers[:15]:\n",
    "    layer.trainable = False\n",
    "\n",
    "# compile the model with a SGD/momentum optimizer\n",
    "# and a very slow learning rate.\n",
    "model.compile(loss='binary_crossentropy',\n",
    "              optimizer=optimizers.SGD(lr=1e-4, momentum=0.9),\n",
    "              metrics=['accuracy'])\n",
    "\n",
    "print model.summary()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Found 2000 images belonging to 2 classes.\n",
      "Found 800 images belonging to 2 classes.\n"
     ]
    }
   ],
   "source": [
    "# prepare data augmentation configuration\n",
    "train_datagen = ImageDataGenerator(\n",
    "    rescale=1. / 255,\n",
    "    shear_range=0.2,\n",
    "    zoom_range=0.2,\n",
    "    horizontal_flip=True)\n",
    "\n",
    "test_datagen = ImageDataGenerator(rescale=1. / 255)\n",
    "\n",
    "train_generator = train_datagen.flow_from_directory(\n",
    "    train_data_dir,\n",
    "    target_size=(img_height, img_width),\n",
    "    batch_size=batch_size,\n",
    "    class_mode='binary')\n",
    "\n",
    "validation_generator = test_datagen.flow_from_directory(\n",
    "    validation_data_dir,\n",
    "    target_size=(img_height, img_width),\n",
    "    batch_size=batch_size,\n",
    "    class_mode='binary')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "** Finally train the model! **"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 1/50\n",
      "125/125 [==============================] - 22s 179ms/step - loss: 0.5183 - acc: 0.8350 - val_loss: 0.2734 - val_acc: 0.9100\n",
      "Epoch 2/50\n",
      "125/125 [==============================] - 20s 163ms/step - loss: 0.2378 - acc: 0.9095 - val_loss: 0.2529 - val_acc: 0.9050\n",
      "Epoch 3/50\n",
      "125/125 [==============================] - 20s 164ms/step - loss: 0.1763 - acc: 0.9410 - val_loss: 0.3169 - val_acc: 0.9075\n",
      "Epoch 4/50\n",
      "125/125 [==============================] - 20s 164ms/step - loss: 0.1411 - acc: 0.9500 - val_loss: 0.3348 - val_acc: 0.9075\n",
      "Epoch 5/50\n",
      "125/125 [==============================] - 20s 164ms/step - loss: 0.1235 - acc: 0.9520 - val_loss: 0.2526 - val_acc: 0.9137\n",
      "Epoch 6/50\n",
      "125/125 [==============================] - 20s 164ms/step - loss: 0.1013 - acc: 0.9620 - val_loss: 0.3191 - val_acc: 0.9075\n",
      "Epoch 7/50\n",
      "125/125 [==============================] - 20s 164ms/step - loss: 0.0806 - acc: 0.9695 - val_loss: 0.3303 - val_acc: 0.9150\n",
      "Epoch 8/50\n",
      "125/125 [==============================] - 20s 164ms/step - loss: 0.0605 - acc: 0.9785 - val_loss: 0.3676 - val_acc: 0.9187\n",
      "Epoch 9/50\n",
      "125/125 [==============================] - 20s 164ms/step - loss: 0.0623 - acc: 0.9760 - val_loss: 0.3317 - val_acc: 0.9250\n",
      "Epoch 10/50\n",
      "125/125 [==============================] - 20s 164ms/step - loss: 0.0761 - acc: 0.9745 - val_loss: 0.3928 - val_acc: 0.9237\n",
      "Epoch 11/50\n",
      "125/125 [==============================] - 20s 164ms/step - loss: 0.0536 - acc: 0.9815 - val_loss: 0.2517 - val_acc: 0.9225\n",
      "Epoch 12/50\n",
      "125/125 [==============================] - 20s 164ms/step - loss: 0.0484 - acc: 0.9845 - val_loss: 0.4273 - val_acc: 0.9225\n",
      "Epoch 13/50\n",
      "125/125 [==============================] - 20s 164ms/step - loss: 0.0409 - acc: 0.9860 - val_loss: 0.3266 - val_acc: 0.9213\n",
      "Epoch 14/50\n",
      "125/125 [==============================] - 20s 164ms/step - loss: 0.0404 - acc: 0.9855 - val_loss: 0.4019 - val_acc: 0.9150\n",
      "Epoch 15/50\n",
      "125/125 [==============================] - 20s 164ms/step - loss: 0.0263 - acc: 0.9925 - val_loss: 0.3818 - val_acc: 0.9163\n",
      "Epoch 16/50\n",
      "125/125 [==============================] - 20s 164ms/step - loss: 0.0275 - acc: 0.9905 - val_loss: 0.4401 - val_acc: 0.9250\n",
      "Epoch 17/50\n",
      "125/125 [==============================] - 20s 164ms/step - loss: 0.0284 - acc: 0.9895 - val_loss: 0.3713 - val_acc: 0.9150\n",
      "Epoch 18/50\n",
      "125/125 [==============================] - 20s 164ms/step - loss: 0.0285 - acc: 0.9900 - val_loss: 0.3477 - val_acc: 0.9263\n",
      "Epoch 19/50\n",
      "125/125 [==============================] - 20s 164ms/step - loss: 0.0206 - acc: 0.9940 - val_loss: 0.3831 - val_acc: 0.9200\n",
      "Epoch 20/50\n",
      "125/125 [==============================] - 20s 164ms/step - loss: 0.0157 - acc: 0.9910 - val_loss: 0.5340 - val_acc: 0.9125\n",
      "Epoch 21/50\n",
      "125/125 [==============================] - 20s 164ms/step - loss: 0.0202 - acc: 0.9925 - val_loss: 0.4502 - val_acc: 0.9125\n",
      "Epoch 22/50\n",
      "125/125 [==============================] - 21s 164ms/step - loss: 0.0206 - acc: 0.9940 - val_loss: 0.5521 - val_acc: 0.9200\n",
      "Epoch 23/50\n",
      "125/125 [==============================] - 20s 164ms/step - loss: 0.0190 - acc: 0.9935 - val_loss: 0.3896 - val_acc: 0.9225\n",
      "Epoch 24/50\n",
      "125/125 [==============================] - 20s 164ms/step - loss: 0.0122 - acc: 0.9950 - val_loss: 0.4255 - val_acc: 0.9313\n",
      "Epoch 25/50\n",
      "125/125 [==============================] - 20s 163ms/step - loss: 0.0084 - acc: 0.9970 - val_loss: 0.4214 - val_acc: 0.9200\n",
      "Epoch 26/50\n",
      "125/125 [==============================] - 20s 164ms/step - loss: 0.0058 - acc: 0.9985 - val_loss: 0.3923 - val_acc: 0.9250\n",
      "Epoch 27/50\n",
      "125/125 [==============================] - 20s 164ms/step - loss: 0.0224 - acc: 0.9925 - val_loss: 0.4568 - val_acc: 0.9287\n",
      "Epoch 28/50\n",
      "125/125 [==============================] - 20s 164ms/step - loss: 0.0379 - acc: 0.9920 - val_loss: 0.3002 - val_acc: 0.9250\n",
      "Epoch 29/50\n",
      "125/125 [==============================] - 20s 163ms/step - loss: 0.0285 - acc: 0.9945 - val_loss: 0.5268 - val_acc: 0.9137\n",
      "Epoch 30/50\n",
      "125/125 [==============================] - 20s 164ms/step - loss: 0.0400 - acc: 0.9875 - val_loss: 0.3688 - val_acc: 0.9225\n",
      "Epoch 31/50\n",
      "125/125 [==============================] - 20s 164ms/step - loss: 0.0102 - acc: 0.9955 - val_loss: 0.4523 - val_acc: 0.9313\n",
      "Epoch 32/50\n",
      "125/125 [==============================] - 20s 164ms/step - loss: 0.0054 - acc: 0.9985 - val_loss: 0.5035 - val_acc: 0.9263\n",
      "Epoch 33/50\n",
      "125/125 [==============================] - 20s 164ms/step - loss: 0.0195 - acc: 0.9940 - val_loss: 0.4182 - val_acc: 0.9225\n",
      "Epoch 34/50\n",
      "125/125 [==============================] - 20s 164ms/step - loss: 0.0136 - acc: 0.9950 - val_loss: 0.4399 - val_acc: 0.9300\n",
      "Epoch 35/50\n",
      "125/125 [==============================] - 20s 164ms/step - loss: 0.0168 - acc: 0.9945 - val_loss: 0.4108 - val_acc: 0.9237\n",
      "Epoch 36/50\n",
      "125/125 [==============================] - 20s 164ms/step - loss: 0.0047 - acc: 0.9995 - val_loss: 0.4880 - val_acc: 0.9250\n",
      "Epoch 37/50\n",
      "125/125 [==============================] - 20s 164ms/step - loss: 0.0202 - acc: 0.9970 - val_loss: 0.4422 - val_acc: 0.9237\n",
      "Epoch 38/50\n",
      "125/125 [==============================] - 20s 164ms/step - loss: 0.0142 - acc: 0.9965 - val_loss: 0.4092 - val_acc: 0.9213\n",
      "Epoch 39/50\n",
      "125/125 [==============================] - 20s 164ms/step - loss: 0.0097 - acc: 0.9955 - val_loss: 0.3900 - val_acc: 0.9275\n",
      "Epoch 40/50\n",
      "125/125 [==============================] - 20s 164ms/step - loss: 0.0113 - acc: 0.9960 - val_loss: 0.4725 - val_acc: 0.9137\n",
      "Epoch 41/50\n",
      "125/125 [==============================] - 20s 163ms/step - loss: 0.0058 - acc: 0.9985 - val_loss: 0.4197 - val_acc: 0.9237\n",
      "Epoch 42/50\n",
      "125/125 [==============================] - 20s 164ms/step - loss: 0.0084 - acc: 0.9970 - val_loss: 0.4394 - val_acc: 0.9237\n",
      "Epoch 43/50\n",
      "125/125 [==============================] - 20s 164ms/step - loss: 0.0109 - acc: 0.9970 - val_loss: 0.4607 - val_acc: 0.9225\n",
      "Epoch 44/50\n",
      "125/125 [==============================] - 20s 164ms/step - loss: 0.0046 - acc: 0.9980 - val_loss: 0.4826 - val_acc: 0.9250\n",
      "Epoch 45/50\n",
      "125/125 [==============================] - 20s 163ms/step - loss: 0.0026 - acc: 0.9990 - val_loss: 0.4939 - val_acc: 0.9313\n",
      "Epoch 46/50\n",
      "125/125 [==============================] - 20s 163ms/step - loss: 0.0014 - acc: 0.9990 - val_loss: 0.5016 - val_acc: 0.9237\n",
      "Epoch 47/50\n",
      "125/125 [==============================] - 20s 164ms/step - loss: 0.0361 - acc: 0.9915 - val_loss: 0.4228 - val_acc: 0.9137\n",
      "Epoch 48/50\n",
      "125/125 [==============================] - 20s 164ms/step - loss: 0.0213 - acc: 0.9935 - val_loss: 0.3931 - val_acc: 0.9250\n",
      "Epoch 49/50\n",
      "125/125 [==============================] - 20s 164ms/step - loss: 0.0086 - acc: 0.9965 - val_loss: 0.4426 - val_acc: 0.9200\n",
      "Epoch 50/50\n",
      "125/125 [==============================] - 20s 164ms/step - loss: 0.0045 - acc: 0.9985 - val_loss: 0.4834 - val_acc: 0.9300\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<keras.callbacks.History at 0x7f19b8e6d150>"
      ]
     },
     "execution_count": 8,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# fine-tune the model\n",
    "model.fit_generator(\n",
    "    train_generator,\n",
    "    steps_per_epoch=nb_train_samples // batch_size,\n",
    "    epochs=epochs,\n",
    "    validation_data=validation_generator,\n",
    "    validation_steps=nb_validation_samples // batch_size,\n",
    "    verbose=1)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "**Fine tuning, after 50 epochs, we have an accuracy of 93%**"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "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
}