xception_model_with_multi_output.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Train an Xception Model from Keras with 2 Output Layers\n",
    "**The output layers will classify an image into 2 different categories**\n",
    "\n",
    "This script was motivated by:\n",
    "- https://www.pyimagesearch.com/2018/06/04/keras-multiple-outputs-and-multiple-losses/ \n",
    "- https://www.kaggle.com/abnera/transfer-learning-keras-xception-cnn?scriptVersionId=818701\n",
    "\n",
    "I wanted to show that it is possible to use the Xception model with multiple output layers to classify one image into 2 separate categories."
   ]
  },
  {
   "attachments": {
    "diagram_xception_two_output_layer.png": {
     "image/png": "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"
    }
   },
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "![diagram_xception_two_output_layer.png](attachment:diagram_xception_two_output_layer.png)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2018-10-01T12:25:30.342787Z",
     "start_time": "2018-10-01T12:25:27.890549Z"
    }
   },
   "outputs": [],
   "source": [
    "from keras.layers import *\n",
    "from keras.optimizers import *\n",
    "from keras.applications import *\n",
    "from keras.models import Model\n",
    "from keras.preprocessing.image import ImageDataGenerator, img_to_array\n",
    "from keras.callbacks import ModelCheckpoint, EarlyStopping\n",
    "from keras import backend as k\n",
    "from keras.utils import to_categorical\n",
    "\n",
    "import matplotlib.pyplot as plt"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Settings"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Run the following lines, if you want to use CPU only or ignore this if you don't have a GPU."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2018-10-01T12:25:30.390504Z",
     "start_time": "2018-10-01T12:25:30.344935Z"
    }
   },
   "outputs": [],
   "source": [
    "config = tf.ConfigProto(allow_soft_placement=True, device_count = {'CPU' : 1, 'GPU' : 0})\n",
    "session = tf.Session(config=config)\n",
    "k.set_session(session)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "- Set the paths to the images and to the model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2018-10-01T12:25:35.165192Z",
     "start_time": "2018-10-01T12:25:35.160708Z"
    }
   },
   "outputs": [],
   "source": [
    "import os\n",
    "\n",
    "# set the paths to the images and the model\n",
    "data_dir = os.path.abspath(\"images/\")\n",
    "model_dir = os.path.abspath(\"model/model1/\")\n",
    "\n",
    "# create folders if they don't exist\n",
    "os.makedirs(os.path.abspath('preview/'), exist_ok=True)\n",
    "os.makedirs(model_dir, exist_ok=True)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Hyper Parameters"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2018-10-01T12:25:46.309389Z",
     "start_time": "2018-10-01T12:25:46.304631Z"
    }
   },
   "outputs": [],
   "source": [
    "# fix seed for reproducible results (only works on CPU, not GPU)\n",
    "seed = 9\n",
    "np.random.seed(seed=seed)\n",
    "tf.set_random_seed(seed=seed)\n",
    "\n",
    "img_width, img_height = 299, 299  # change based on the shape/structure of your images\n",
    "\n",
    "# hyper parameters for model\n",
    "based_model_last_block_layer_number = 126  # value is based on based model selected.\n",
    "batch_size = 32  # try 4, 8, 16, 32, 64, 128, 256 dependent on CPU/GPU memory capacity (powers of 2 values).\n",
    "nb_epoch = 50  # number of iteration the algorithm gets trained.\n",
    "np_classes = 4 # number of classes the algorithm uses\n",
    "learn_rate = 1e-4  # sgd learning rate\n",
    "momentum = .9  # sgd momentum to avoid local minimum\n",
    "transformation_ratio = .05  # how aggressive will be the data augmentation/transformation"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Gather Image Paths"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "**Important Note: the filename needs to contain cat1 and cat2 in order to be able to train the model**\n",
    "- Search for png images in the given directory and the subdirectories within and return a list of file paths\n",
    "- Set a random seed first to reproduce the results\n",
    "- Shuffle the list of file paths"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2018-09-28T14:40:02.140104Z",
     "start_time": "2018-09-28T14:40:02.016658Z"
    }
   },
   "outputs": [],
   "source": [
    "import glob # pip install glob\n",
    "import random\n",
    "\n",
    "print(\"[INFO] Searching Images...\")\n",
    "imagePaths = glob.glob(data_dir + '/**/*.png', recursive=True)\n",
    "random.seed(seed)\n",
    "random.shuffle(imagePaths)\n",
    "print(len(imagePaths))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Populate Label- and Datalist"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "- Loop through the list of file paths\n",
    "- split the file name into cat1 and cat2\n",
    "- open the image, then resize it to the specified image size of the model\n",
    "- transfer the image to a numpy array and add it to the list of images"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "ExecuteTime": {
     "start_time": "2018-09-28T14:44:38.240Z"
    },
    "scrolled": true
   },
   "outputs": [],
   "source": [
    "import cv2 # pip install opencv-python\n",
    "\n",
    "print(\"[INFO] Gathering data and labels\")\n",
    "\n",
    "labelsCat1 = [] \n",
    "labelsCat2 = [] \n",
    "imageDataList = []\n",
    "\n",
    "counter = 0\n",
    "for files in imagePaths:\n",
    "    # progress counter\n",
    "    if (counter + 1) % 100 == 0:\n",
    "        print(counter + 1, \"/\", len(imagePaths))\n",
    "    \n",
    "    # split the filename into category 1 and 2, ignore the filenumber\n",
    "    filename = files.split(os.path.sep)[-1]\n",
    "    \n",
    "    categories = filename.split(\"_\")\n",
    "    (cat1, cat2, _) = categories\n",
    "    \n",
    "    # add the categories to the lists\n",
    "    labelsCat1.append(cat1)\n",
    "    labelsCat2.append(cat2)\n",
    "\n",
    "    # load the image, pre-process it, and store it in the data list\n",
    "    image = cv2.imread(files)\n",
    "    image = cv2.resize(image, (img_width, img_height))\n",
    "    image = img_to_array(image)\n",
    "    imageDataList.append(image)\n",
    "    \n",
    "    counter += 1"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Transform Image Data"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "- normalise the image by dividing every RGB pixel value with 255 to achieve a value between 0 and 1"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "ExecuteTime": {
     "start_time": "2018-09-28T14:45:34.967Z"
    }
   },
   "outputs": [],
   "source": [
    "import numpy as np\n",
    "\n",
    "print(\"[INFO] Transform the Data\")\n",
    "# scale the raw pixel intensities to the range [0, 1] and convert to a NumPy array\n",
    "imageData = np.array(imageDataList, dtype=\"float\") / 255.\n",
    "print(\"[INFO] data matrix: {} images ({:.2f}MB)\".format(\n",
    "    len(imagePaths), imageData.nbytes / (1024 * 1000.0)))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Preprocess Label Data"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "- transform the labels to a binary form so the model can process them\n",
    "- **Remember that the Binarizer is sorting the labels alphabetically**"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "ExecuteTime": {
     "start_time": "2018-09-28T14:58:51.363Z"
    }
   },
   "outputs": [],
   "source": [
    "from sklearn.preprocessing import LabelBinarizer\n",
    "\n",
    "# binarize both sets of labels\n",
    "print(\"[INFO] binarizing labels...\")\n",
    "\n",
    "# create numpy arrays from labels\n",
    "labelsCat1 = np.array(labelsCat1)\n",
    "labelsCat2 = np.array(labelsCat2)\n",
    "\n",
    "# initialize the binarizer to transform the labels from str to binary representation\n",
    "binarizer = LabelBinarizer()\n",
    "\n",
    "# labelsCat1 needs to be transformed into categorical representation because it only contains 1 values \n",
    "# which will be translated into 0 and 1\n",
    "# the fit_transform function translates them into an array of shape (,1)\n",
    "# one column creates troubles when the loss function is set to categorical\n",
    "# change the loss function or translate the labels to a categorical representation to solve this problem\n",
    "labelsCat1 = binarizer.fit_transform(labelsCat1)\n",
    "labelsCat1 = to_categorical(labelsCat1)\n",
    "\n",
    "# labelsCat2 does not need to be translated into categorical representation because it contains more than 2 values\n",
    "# calling the fit_transform function translates them into an array of shape (,4)\n",
    "# this does not create problems in combination with the categorical loss function\n",
    "labelsCat2 = binarizer.fit_transform(labelsCat2)\n",
    "\n",
    "# uncomment following line, if you need to sort the columns differently to read them better afterwards\n",
    "# the labelsCat2 columns need to be sorted differently\n",
    "# np.argsort([1,2,0,3]) means that the column gets moved to another position like:\n",
    "# column old - new\n",
    "#         0  -  1\n",
    "#         1  -  2\n",
    "#         2  -  0\n",
    "#         3  -  3\n",
    "# labelsCat2 = labelsCat2[:,np.argsort([1,2,0,3])]\n",
    "\n",
    "#print out labels and image\n",
    "index = 0\n",
    "\n",
    "# in my case it looked like this:\n",
    "# [1. 0.] represents cat1 a\n",
    "# [0. 1.] represents cat1 b\n",
    "print(labelsCat1[index])\n",
    "\n",
    "# in my case it looked like this:\n",
    "# [1 0 0 0] represents cat2 a\n",
    "# [0 1 0 0] represents cat2 b\n",
    "# [0 0 1 0] represents cat2 c\n",
    "# [0 0 0 1] represents cat2 d\n",
    "print(labelsCat2[index])\n",
    "\n",
    "# plot image\n",
    "plt.imshow(imageData[index])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Split Labels and Data"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "ExecuteTime": {
     "start_time": "2018-09-28T14:58:54.090Z"
    }
   },
   "outputs": [],
   "source": [
    "from sklearn.model_selection import train_test_split\n",
    "\n",
    "print(\"[INFO] Split dataset into train and validation...\")\n",
    "# partition the data into training and testing splits using 80% of\n",
    "# the data for training and the remaining 20% for testing\n",
    "split = train_test_split(imageData, labelsCat1, labelsCat2,\n",
    "    test_size=0.2, random_state=seed)\n",
    "(trainX, testX, trainCat1Y, testCat1Y, trainCat2Y, testCat2Y) = split\n",
    "\n",
    "# print length of training and test set\n",
    "print(len(trainX), len(testX))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Create Model"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "- Create the base model\n",
    "- load the base model with pretrained weights of imagenet\n",
    "- don't include the top layer because we want to create our own top layer\n",
    "- create two output layers with labelsCat1.shape[-1] and labelsCat2.shape[-1] output nodes respectively\n",
    "- _in my case it was 2 + 4 nodes for cat1 and cat2 respectively_\n",
    "- **With 2 output layers, val_acc can not be monitored in the callbacks it has to be changed to val_loss**"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "ExecuteTime": {
     "start_time": "2018-09-28T14:59:01.796Z"
    }
   },
   "outputs": [],
   "source": [
    "# Pre-Trained CNN Model using imagenet dataset for pre-trained weights\n",
    "base_model = Xception(input_shape=(img_width, img_height, 3), weights='imagenet', include_top=False)\n",
    "\n",
    "# Top Model Block\n",
    "x = base_model.output\n",
    "x = GlobalAveragePooling2D()(x)\n",
    "\n",
    "# add the first output layer (2 nodes)\n",
    "output_cat_1 = Dense(labelsCat1.shape[-1], activation='softmax', name='output_cat_1')(x)\n",
    "\n",
    "# add the second output layer (4 nodes)\n",
    "output_cat_2 = Dense(labelsCat2.shape[-1], activation='softmax', name='output_cat_2')(x)\n",
    "\n",
    "\n",
    "# add your top layer(s) block to your base model\n",
    "model = Model(inputs=base_model.input, outputs=[output_cat_1, output_cat_2])\n",
    "print(model.summary())\n",
    "\n",
    "# # let's visualize layer names and layer indices to see how many layers/blocks to re-train\n",
    "# # uncomment when choosing based_model_last_block_layer\n",
    "# for i, layer in enumerate(model.layers):\n",
    "#     print(i, layer.name)\n",
    "\n",
    "# first: train only the top layers (which were randomly initialized)\n",
    "# i.e. freeze all layers of the based model that is already pre-trained.\n",
    "for layer in base_model.layers:\n",
    "    layer.trainable = False\n",
    "\n",
    "# add two loss functions for each output layer separate\n",
    "model.compile(optimizer='nadam',\n",
    "              loss={\"output_cat_1\": \"categorical_crossentropy\", \"output_cat_2\": \"categorical_crossentropy\"},\n",
    "              metrics=['accuracy'])\n",
    "\n",
    "# save weights of best training epoch: monitor either val_loss or val_acc\n",
    "\n",
    "top_weights_path = os.path.join(os.path.abspath(model_dir), 'top_model_weights.h5')\n",
    "callbacks_list = [\n",
    "    ModelCheckpoint(top_weights_path, monitor=\"val_loss\", verbose=1, save_best_only=True),\n",
    "    EarlyStopping(monitor=\"val_loss\", patience=10, verbose=0),\n",
    "]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Training"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "ExecuteTime": {
     "start_time": "2018-09-28T14:59:20.124Z"
    },
    "scrolled": false
   },
   "outputs": [],
   "source": [
    "print(\"\\nStart Training Model\\n\")\n",
    "# Train Simple CNN\n",
    "# Sequential Mode\n",
    "model.fit(trainX,\n",
    "        {\"output_cat_1\": trainCat1Y, \"output_cat_2\": trainCat2Y},\n",
    "        validation_data=(testX,\n",
    "                        {\"output_cat_1\": testCat1Y, \"output_cat_2\": testCat2Y}),\n",
    "        epochs=nb_epoch,\n",
    "        callbacks=callbacks_list)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Fine Tuning"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "- Load the weights for the model with two output layers to perform the fine tuning of the model\n",
    "- Freeze all layers except the last 2 layers of the model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "ExecuteTime": {
     "start_time": "2018-09-28T14:59:32.645Z"
    }
   },
   "outputs": [],
   "source": [
    "# add the best weights from the train top model\n",
    "# at this point we have the pre-train weights of the base model and the trained weight of the new/added top model\n",
    "# we re-load model weights to ensure the best epoch is selected and not the last one.\n",
    "model.load_weights(top_weights_path)\n",
    "\n",
    "# based_model_last_block_layer_number points to the layer in your model you want to train.\n",
    "# For example if you want to train the last block of a 19 layer VGG16 model this should be 15\n",
    "# If you want to train the last Two blocks of an Inception model it should be 172\n",
    "# layers before this number will used the pre-trained weights, layers above and including this number\n",
    "# will be re-trained based on the new data.\n",
    "for layer in model.layers[:based_model_last_block_layer_number]:\n",
    "    layer.trainable = False\n",
    "for layer in model.layers[based_model_last_block_layer_number:]:\n",
    "    layer.trainable = True\n",
    "\n",
    "# compile the model with a SGD/momentum optimizer\n",
    "# and a very slow learning rate.\n",
    "\n",
    "\n",
    "model.compile(optimizer='nadam',\n",
    "              loss={\"output_cat_1\": \"categorical_crossentropy\", \"output_cat_2\": \"categorical_crossentropy\"},\n",
    "              metrics=['accuracy'])\n",
    "\n",
    "# save weights of best training epoch: monitor either val_loss or val_acc\n",
    "final_weights_path = os.path.join(os.path.abspath(model_dir), 'model_weights.h5')\n",
    "callbacks_list = [\n",
    "    ModelCheckpoint(final_weights_path, monitor='val_loss', verbose=1, save_best_only=True),\n",
    "    EarlyStopping(monitor='val_loss', patience=10, verbose=0)\n",
    "]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "- Start fine tune for two output layers"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "ExecuteTime": {
     "start_time": "2018-09-28T14:59:37.111Z"
    },
    "scrolled": false
   },
   "outputs": [],
   "source": [
    "print(\"\\nStart Fine Tuning Model\\n\")\n",
    "# Fine Tune Model\n",
    "# Sequential Mode\n",
    "model.fit(trainX,\n",
    "        {\"output_cat_1\": trainCat1Y, \"output_cat_2\": trainCat2Y},\n",
    "        validation_data=(testX,\n",
    "                        {\"output_cat_1\": testCat1Y, \"output_cat_2\": testCat2Y}),\n",
    "        epochs=nb_epoch,\n",
    "        callbacks=callbacks_list)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "heading_collapsed": true
   },
   "source": [
    "## Prediction"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2018-09-27T15:48:04.065067Z",
     "start_time": "2018-09-27T15:48:00.628414Z"
    },
    "hidden": true
   },
   "outputs": [],
   "source": [
    "probabilities = model.predict(testX, verbose=1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "ExecuteTime": {
     "end_time": "2018-09-27T15:48:31.640394Z",
     "start_time": "2018-09-27T15:48:31.636438Z"
    },
    "hidden": true
   },
   "outputs": [],
   "source": [
    "print(probabilities)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Save Model To JSON"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "ExecuteTime": {
     "start_time": "2018-09-28T14:59:58.479Z"
    }
   },
   "outputs": [],
   "source": [
    "# save the model\n",
    "model_json = model.to_json()\n",
    "with open(os.path.join(os.path.abspath(model_dir), 'model.json'), 'w') as json_file:\n",
    "    json_file.write(model_json)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Release Memory"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "ExecuteTime": {
     "start_time": "2018-09-28T15:00:01.513Z"
    },
    "scrolled": true
   },
   "outputs": [],
   "source": [
    "# release memory\n",
    "k.clear_session()"
   ]
  }
 ],
 "metadata": {
  "celltoolbar": "Initialization Cell",
  "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": "3.6.5"
  },
  "toc": {
   "base_numbering": 1,
   "nav_menu": {
    "height": "494px",
    "width": "628px"
   },
   "number_sections": true,
   "sideBar": true,
   "skip_h1_title": false,
   "title_cell": "Table of Contents",
   "title_sidebar": "Table of Contents",
   "toc_cell": false,
   "toc_position": {
    "height": "calc(100% - 180px)",
    "left": "10px",
    "top": "150px",
    "width": "174px"
   },
   "toc_section_display": true,
   "toc_window_display": true
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}