train_vgg_model.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Import libraries"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Using TensorFlow backend.\n"
     ]
    }
   ],
   "source": [
    "from keras.models import Model\n",
    "from keras.layers import Flatten, Dense, Lambda\n",
    "from keras.applications.vgg16 import VGG16\n",
    "from keras.optimizers import Adam\n",
    "from keras.preprocessing import image\n",
    "from keras import backend as K"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Set variables and import data"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "# set variables\n",
    "gen = image.ImageDataGenerator()\n",
    "batch_size = 64"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Found 8721 images belonging to 3 classes.\n"
     ]
    }
   ],
   "source": [
    "# import training data\n",
    "batches = gen.flow_from_directory('data/train',\n",
    "                                  target_size=(224,224),\n",
    "                                  class_mode='categorical',\n",
    "                                  shuffle=True,\n",
    "                                  batch_size=batch_size)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Found 2100 images belonging to 3 classes.\n"
     ]
    }
   ],
   "source": [
    "# import validation data\n",
    "val_batches = gen.flow_from_directory('data/valid',\n",
    "                                      target_size=(224,224),\n",
    "                                      class_mode='categorical',\n",
    "                                      shuffle=True,\n",
    "                                      batch_size=batch_size)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Import and configure model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [],
   "source": [
    "# retrieve the full Keras VGG model including imagenet weights\n",
    "vgg = VGG16(include_top=True, weights='imagenet',\n",
    "                               input_tensor=None, input_shape=(224,224,3), pooling=None)\n",
    "\n",
    "# define a new output layer to connect with the last fc layer in vgg\n",
    "x = vgg.layers[-2].output\n",
    "output_layer = Dense(3, activation='softmax', name='predictions')(x) # 3 ouputs - forward, left, right\n",
    "\n",
    "# combine the original VGG model with the new output layer\n",
    "vgg2 = Model(inputs=vgg.input, outputs=output_layer)\n",
    "\n",
    "# compile the new model\n",
    "vgg2.compile(optimizer=Adam(lr=0.0001), #was 0.001\n",
    "                loss='categorical_crossentropy', metrics=['accuracy'])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Train it!"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 1/1\n",
      "136/136 [==============================] - 111s - loss: 1.1160 - acc: 0.5068 - val_loss: 0.7706 - val_acc: 0.6724\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<keras.callbacks.History at 0x7f574145bd68>"
      ]
     },
     "execution_count": 9,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "vgg2.fit_generator(batches,\n",
    "                   steps_per_epoch = batches.samples // batch_size,\n",
    "                   validation_data = val_batches, \n",
    "                   validation_steps = val_batches.samples // batch_size,\n",
    "                   epochs = 1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 1/1\n",
      "136/136 [==============================] - 107s - loss: 0.7390 - acc: 0.6740 - val_loss: 0.7764 - val_acc: 0.6375\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<keras.callbacks.History at 0x7f57397bcda0>"
      ]
     },
     "execution_count": 10,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "vgg2.fit_generator(batches,\n",
    "                   steps_per_epoch = batches.samples // batch_size,\n",
    "                   validation_data = val_batches, \n",
    "                   validation_steps = val_batches.samples // batch_size,\n",
    "                   epochs = 1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 1/1\n",
      "136/136 [==============================] - 103s - loss: 0.6726 - acc: 0.7077 - val_loss: 0.6749 - val_acc: 0.7230\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<keras.callbacks.History at 0x7f5738f5ab70>"
      ]
     },
     "execution_count": 11,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "vgg2.fit_generator(batches,\n",
    "                   steps_per_epoch = batches.samples // batch_size,\n",
    "                   validation_data = val_batches, \n",
    "                   validation_steps = val_batches.samples // batch_size,\n",
    "                   epochs = 1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 1/3\n",
      "136/136 [==============================] - 105s - loss: 0.5994 - acc: 0.7454 - val_loss: 0.6667 - val_acc: 0.7245\n",
      "Epoch 2/3\n",
      "136/136 [==============================] - 105s - loss: 0.5376 - acc: 0.7702 - val_loss: 0.7630 - val_acc: 0.7038\n",
      "Epoch 3/3\n",
      "136/136 [==============================] - 104s - loss: 0.4818 - acc: 0.7973 - val_loss: 0.6560 - val_acc: 0.7451\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<keras.callbacks.History at 0x7f5738f5aef0>"
      ]
     },
     "execution_count": 12,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "vgg2.fit_generator(batches,\n",
    "                   steps_per_epoch = batches.samples // batch_size,\n",
    "                   validation_data = val_batches, \n",
    "                   validation_steps = val_batches.samples // batch_size,\n",
    "                   epochs = 3)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [],
   "source": [
    "vgg2.save('keras_models/gtavc1.h5')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [],
   "source": [
    "from keras.models import load_model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [],
   "source": [
    "vgg2 = load_model('keras_models/gtavc1.h5')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Overtraining - not used"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 1/3\n",
      "136/136 [==============================] - 110s - loss: 0.4211 - acc: 0.8313 - val_loss: 0.6827 - val_acc: 0.7471\n",
      "Epoch 2/3\n",
      "136/136 [==============================] - 110s - loss: 0.3776 - acc: 0.8422 - val_loss: 0.6867 - val_acc: 0.7579\n",
      "Epoch 3/3\n",
      "136/136 [==============================] - 108s - loss: 0.3069 - acc: 0.8739 - val_loss: 0.8031 - val_acc: 0.7362\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<keras.callbacks.History at 0x7f608a59e128>"
      ]
     },
     "execution_count": 14,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "vgg2.fit_generator(batches,\n",
    "                   steps_per_epoch = batches.samples // batch_size,\n",
    "                   validation_data = val_batches, \n",
    "                   validation_steps = val_batches.samples // batch_size,\n",
    "                   epochs = 3)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 1/3\n",
      "136/136 [==============================] - 107s - loss: 0.2498 - acc: 0.9011 - val_loss: 0.8753 - val_acc: 0.7367\n",
      "Epoch 2/3\n",
      "136/136 [==============================] - 107s - loss: 0.2133 - acc: 0.9150 - val_loss: 0.9853 - val_acc: 0.7397\n",
      "Epoch 3/3\n",
      "136/136 [==============================] - 109s - loss: 0.1787 - acc: 0.9313 - val_loss: 0.9651 - val_acc: 0.7279\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<keras.callbacks.History at 0x7f608a59e2e8>"
      ]
     },
     "execution_count": 15,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "vgg2.fit_generator(batches,\n",
    "                   steps_per_epoch = batches.samples // batch_size,\n",
    "                   validation_data = val_batches, \n",
    "                   validation_steps = val_batches.samples // batch_size,\n",
    "                   epochs = 3)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# print learning rate\n",
    "print(K.eval(vgg2.optimizer.lr))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# change learning rate\n",
    "K.set_value(vgg2.optimizer.lr, 0.1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "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": "3.6.3"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}