2.4 Transfer Learning & Fine-Tuning.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Transfer Learning and Fine Tuning"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "* Train a simple convnet on the MNIST dataset the first 5 digits [0..4].\n",
    "* Freeze convolutional layers and fine-tune dense layers for the classification of digits [5..9]."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Using GPU (highly recommended)\n",
    "\n",
    "-> If using `theano` backend:\n",
    "\n",
    "`THEANO_FLAGS=mode=FAST_RUN,device=gpu,floatX=float32`"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "import numpy as np\n",
    "import datetime\n",
    "\n",
    "np.random.seed(1337)  # for reproducibility"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/home/user/anaconda2/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 Theano backend.\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "2.1.6\n"
     ]
    }
   ],
   "source": [
    "from keras.datasets import mnist\n",
    "from keras.models import Sequential\n",
    "from keras.layers import Dense, Dropout, Activation, Flatten\n",
    "from keras.layers import Convolution2D, MaxPooling2D\n",
    "from keras.utils import np_utils\n",
    "from keras import backend as K\n",
    "from numpy import nan\n",
    "import keras\n",
    "\n",
    "print keras.__version__\n",
    "now = datetime.datetime.now"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Settings"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "now = datetime.datetime.now\n",
    "\n",
    "batch_size = 128\n",
    "nb_classes = 5\n",
    "nb_epoch = 5\n",
    "\n",
    "# input image dimensions\n",
    "img_rows, img_cols = 28, 28\n",
    "# number of convolutional filters to use\n",
    "nb_filters = 32\n",
    "# size of pooling area for max pooling\n",
    "pool_size = 2\n",
    "# convolution kernel size\n",
    "kernel_size = 3"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "if K.image_data_format() == 'channels_first':\n",
    "    input_shape = (1, img_rows, img_cols)\n",
    "else:\n",
    "    input_shape = (img_rows, img_cols, 1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "def train_model(model, train, test, nb_classes):\n",
    "    \n",
    "    X_train = train[0].reshape((train[0].shape[0],) + input_shape)\n",
    "    X_test = test[0].reshape((test[0].shape[0],) + input_shape)\n",
    "    X_train = X_train.astype('float32')\n",
    "    X_test = X_test.astype('float32')\n",
    "    X_train /= 255\n",
    "    X_test /= 255\n",
    "    \n",
    "    print('X_train shape:', X_train.shape)\n",
    "    print(X_train.shape[0], 'train samples')\n",
    "    print(X_test.shape[0], 'test samples')\n",
    "\n",
    "    # convert class vectors to binary class matrices\n",
    "    Y_train = np_utils.to_categorical(train[1], nb_classes)\n",
    "    Y_test = np_utils.to_categorical(test[1], nb_classes)\n",
    "\n",
    "    model.compile(loss='categorical_crossentropy',\n",
    "                  optimizer='adadelta',\n",
    "                  metrics=['accuracy'])\n",
    "\n",
    "    t = now()\n",
    "    model.fit(X_train, Y_train,\n",
    "              batch_size=batch_size, nb_epoch=nb_epoch,\n",
    "              verbose=1,\n",
    "              validation_data=(X_test, Y_test))\n",
    "    print('Training time: %s' % (now() - t))\n",
    "    score = model.evaluate(X_test, Y_test, verbose=0)\n",
    "    print('Test score:', score[0])\n",
    "    print('Test accuracy:', score[1])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Dataset Preparation"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "# the data, shuffled and split between train and test sets\n",
    "(X_train, y_train), (X_test, y_test) = mnist.load_data()\n",
    "\n",
    "# create two datasets one with digits below 5 and one with 5 and above\n",
    "X_train_lt5 = X_train[y_train < 5]\n",
    "y_train_lt5 = y_train[y_train < 5]\n",
    "X_test_lt5 = X_test[y_test < 5]\n",
    "y_test_lt5 = y_test[y_test < 5]\n",
    "\n",
    "X_train_gte5 = X_train[y_train >= 5]\n",
    "y_train_gte5 = y_train[y_train >= 5] - 5  # make classes start at 0 for\n",
    "X_test_gte5 = X_test[y_test >= 5]         # np_utils.to_categorical\n",
    "y_test_gte5 = y_test[y_test >= 5] - 5"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/home/user/anaconda2/lib/python2.7/site-packages/ipykernel/__main__.py:5: UserWarning: Update your `Conv2D` call to the Keras 2 API: `Conv2D(32, (3, 3), padding=\"valid\", input_shape=(28, 28, 1...)`\n",
      "/home/user/anaconda2/lib/python2.7/site-packages/ipykernel/__main__.py:7: UserWarning: Update your `Conv2D` call to the Keras 2 API: `Conv2D(32, (3, 3))`\n"
     ]
    }
   ],
   "source": [
    "# define two groups of layers: feature (convolutions) and classification (dense)\n",
    "feature_layers = [\n",
    "    Convolution2D(nb_filters, kernel_size, kernel_size,\n",
    "                  border_mode='valid',\n",
    "                  input_shape=input_shape),\n",
    "    Activation('relu'),\n",
    "    Convolution2D(nb_filters, kernel_size, kernel_size),\n",
    "    Activation('relu'),\n",
    "    MaxPooling2D(pool_size=(pool_size, pool_size)),\n",
    "    Dropout(0.25),\n",
    "    Flatten(),\n",
    "]\n",
    "classification_layers = [\n",
    "    Dense(128),\n",
    "    Activation('relu'),\n",
    "    Dropout(0.5),\n",
    "    Dense(nb_classes),\n",
    "    Activation('softmax')\n",
    "]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "('X_train shape:', (30596, 28, 28, 1))\n",
      "(30596, 'train samples')\n",
      "(5139, 'test samples')\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/home/user/anaconda2/lib/python2.7/site-packages/keras/models.py:981: UserWarning: The `nb_epoch` argument in `fit` has been renamed `epochs`.\n",
      "  warnings.warn('The `nb_epoch` argument in `fit` '\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Train on 30596 samples, validate on 5139 samples\n",
      "Epoch 1/5\n",
      "30596/30596 [==============================] - 91s 3ms/step - loss: 0.1576 - acc: 0.9512 - val_loss: 0.0289 - val_acc: 0.9885\n",
      "Epoch 2/5\n",
      "30596/30596 [==============================] - 124s 4ms/step - loss: 0.0410 - acc: 0.9869 - val_loss: 0.0109 - val_acc: 0.9969\n",
      "Epoch 3/5\n",
      "30596/30596 [==============================] - 107s 4ms/step - loss: 0.0292 - acc: 0.9912 - val_loss: 0.0099 - val_acc: 0.9967\n",
      "Epoch 4/5\n",
      "30596/30596 [==============================] - 79s 3ms/step - loss: 0.0242 - acc: 0.9925 - val_loss: 0.0087 - val_acc: 0.9969\n",
      "Epoch 5/5\n",
      "30596/30596 [==============================] - 73s 2ms/step - loss: 0.0195 - acc: 0.9944 - val_loss: 0.0096 - val_acc: 0.9971\n",
      "Training time: 0:08:25.213091\n",
      "('Test score:', 0.009554179413411535)\n",
      "('Test accuracy:', 0.997081144191477)\n"
     ]
    }
   ],
   "source": [
    "# create complete model\n",
    "model = Sequential(feature_layers + classification_layers)\n",
    "\n",
    "# train model for 5-digit classification [0..4]\n",
    "train_model(model,\n",
    "            (X_train_lt5, y_train_lt5),\n",
    "            (X_test_lt5, y_test_lt5), nb_classes)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "('X_train shape:', (29404, 28, 28, 1))\n",
      "(29404, 'train samples')\n",
      "(4861, 'test samples')\n",
      "Train on 29404 samples, validate on 4861 samples\n",
      "Epoch 1/5\n",
      "29404/29404 [==============================] - 29s 996us/step - loss: 0.2349 - acc: 0.9334 - val_loss: 0.0450 - val_acc: 0.9866\n",
      "Epoch 2/5\n",
      "29404/29404 [==============================] - 33s 1ms/step - loss: 0.0791 - acc: 0.9747 - val_loss: 0.0347 - val_acc: 0.9887\n",
      "Epoch 3/5\n",
      "29404/29404 [==============================] - 30s 1ms/step - loss: 0.0569 - acc: 0.9828 - val_loss: 0.0295 - val_acc: 0.9905\n",
      "Epoch 4/5\n",
      "29404/29404 [==============================] - 30s 1ms/step - loss: 0.0480 - acc: 0.9846 - val_loss: 0.0255 - val_acc: 0.9914\n",
      "Epoch 5/5\n",
      "29404/29404 [==============================] - 30s 1ms/step - loss: 0.0427 - acc: 0.9866 - val_loss: 0.0244 - val_acc: 0.9916\n",
      "Training time: 0:02:33.375177\n",
      "('Test score:', 0.02438091574341)\n",
      "('Test accuracy:', 0.9915655214976342)\n"
     ]
    }
   ],
   "source": [
    "# freeze feature layers and rebuild model\n",
    "for l in feature_layers:\n",
    "    l.trainable = False\n",
    "\n",
    "# transfer: train dense layers for new classification task [5..9]\n",
    "train_model(model,\n",
    "            (X_train_gte5, y_train_gte5),\n",
    "            (X_test_gte5, y_test_gte5), nb_classes)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Your Turn"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Try to Fine Tune a VGG16 Network"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "input_1 (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"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/home/user/anaconda2/lib/python2.7/site-packages/ipykernel/__main__.py:26: UserWarning: Update your `Model` call to the Keras 2 API: `Model(outputs=Softmax.0, inputs=/image_inp...)`\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "image_input (InputLayer)     (None, 48, 48, 3)         0         \n",
      "_________________________________________________________________\n",
      "vgg16 (Model)                multiple                  14714688  \n",
      "_________________________________________________________________\n",
      "flatten (Flatten)            (None, 512)               0         \n",
      "_________________________________________________________________\n",
      "fc1 (Dense)                  (None, 4096)              2101248   \n",
      "_________________________________________________________________\n",
      "fc2 (Dense)                  (None, 4096)              16781312  \n",
      "_________________________________________________________________\n",
      "predictions (Dense)          (None, 5)                 20485     \n",
      "=================================================================\n",
      "Total params: 33,617,733\n",
      "Trainable params: 33,617,733\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "from keras.applications import VGG16\n",
    "from keras.applications.vgg16 import VGG16\n",
    "from keras.preprocessing import image\n",
    "from keras.applications.vgg16 import preprocess_input\n",
    "from keras.layers import Input, Flatten, Dense\n",
    "from keras.models import Model\n",
    "import numpy as np\n",
    "\n",
    "#Get back the convolutional part of a VGG network trained on ImageNet\n",
    "model_vgg16_conv = VGG16(weights='imagenet', include_top=False)\n",
    "model_vgg16_conv.summary()\n",
    "\n",
    "#Create your own input format (here 3x200x200)\n",
    "inp = Input(shape=(48,48,3),name = 'image_input')\n",
    "\n",
    "#Use the generated model \n",
    "output_vgg16_conv = model_vgg16_conv(inp)\n",
    "\n",
    "#Add the fully-connected layers \n",
    "x = Flatten(name='flatten')(output_vgg16_conv)\n",
    "x = Dense(4096, activation='relu', name='fc1')(x)\n",
    "x = Dense(4096, activation='relu', name='fc2')(x)\n",
    "x = Dense(5, activation='softmax', name='predictions')(x)\n",
    "\n",
    "#Create your own model \n",
    "my_model = Model(input=inp, output=x)\n",
    "\n",
    "#In the summary, weights and layers from VGG part will be hidden, but they will be fit during the training\n",
    "my_model.summary()\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "```python\n",
    "...\n",
    "...\n",
    "# Plugging new Layers\n",
    "    model.add(Dense(768, activation='sigmoid'))\n",
    "    model.add(Dropout(0.0))\n",
    "    model.add(Dense(768, activation='sigmoid'))\n",
    "    model.add(Dropout(0.0))\n",
    "    model.add(Dense(n_labels, activation='softmax'))\n",
    "```"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/home/user/anaconda2/lib/python2.7/site-packages/ipykernel/__main__.py:6: DeprecationWarning: `imresize` is deprecated!\n",
      "`imresize` is deprecated in SciPy 1.0.0, and will be removed in 1.2.0.\n",
      "Use ``skimage.transform.resize`` instead.\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(29404, 48, 48, 3)\n",
      "('X_train shape:', (29404, 48, 48, 3))\n",
      "(29404, 'train samples')\n",
      "(29404, 'test samples')\n",
      "(60000,)\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/home/user/anaconda2/lib/python2.7/site-packages/ipykernel/__main__.py:33: UserWarning: The `nb_epoch` argument in `fit` has been renamed `epochs`.\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 1/5\n"
     ]
    }
   ],
   "source": [
    "import scipy\n",
    "new_shape = (48,48)\n",
    "X_train_new = np.empty(shape=(X_train_gte5.shape[0],)+(48,48,3))\n",
    "\n",
    "for idx in xrange(X_train_gte5.shape[0]):\n",
    "    X_train_new[idx] = np.resize(scipy.misc.imresize(X_train_gte5[idx], (new_shape)), (48, 48, 3))\n",
    "    X_train_new[idx] = np.resize(X_train_new[idx], (48, 48, 3))\n",
    "    \n",
    "#X_train_new = np.expand_dims(X_train_new, axis=-1)\n",
    "\n",
    "print X_train_new.shape\n",
    "\n",
    "X_train_new = X_train_new.astype('float32')\n",
    "X_train_new /= 255\n",
    "    \n",
    "print('X_train shape:', X_train_new.shape)\n",
    "print(X_train_new.shape[0], 'train samples')\n",
    "print(X_train_new.shape[0], 'test samples')\n",
    "\n",
    "\n",
    "# convert class vectors to binary class matrices\n",
    "Y_train = np_utils.to_categorical(y_train_gte5, nb_classes)\n",
    "Y_test = np_utils.to_categorical(y_test_gte5, nb_classes)\n",
    "\n",
    "print y_train.shape\n",
    "\n",
    "my_model.compile(loss='categorical_crossentropy',\n",
    "                  optimizer='adadelta',\n",
    "                  metrics=['accuracy'])\n",
    "\n",
    "my_model.fit(X_train_new, Y_train,\n",
    "              batch_size=batch_size, nb_epoch=nb_epoch,\n",
    "              verbose=1)\n",
    "\n",
    "#print('Training time: %s' % (now() - t))\n",
    "#score = my_model.evaluate(X_test, Y_test, verbose=0)\n",
    "#print('Test score:', score[0])\n",
    "#print('Test accuracy:', score[1])\n",
    "    \n",
    "    \n",
    "    \n",
    "#train_model(my_model,\n",
    "#            (X_train_new, y_train_gte5),\n",
    "#            (X_test_gte5, y_test_gte5), nb_classes)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "anaconda-cloud": {},
  "kernelspec": {
   "display_name": "Python [conda root]",
   "language": "python",
   "name": "conda-root-py"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 2
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython2",
   "version": "2.7.12"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}