Quick Draw CNN.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# QuickDraw CNN with Tensorflow 2.0\n",
    "\n",
    "To start this notebook you need to download the following files. <br>\n",
    "\n",
    "    - full_numpy_bitmap_airplane.npy\n",
    "    - full_numpy_bitmap_book.npy\n",
    "    - full_numpy_bitmap_car.npy\n",
    "    - full_numpy_bitmap_dog.npy\n",
    "    - full_numpy_bitmap_face.npy\n",
    "    - full_numpy_bitmap_apple.npy\n",
    "    - full_numpy_bitmap_brain.npy\n",
    "    - full_numpy_bitmap_chair.npy\n",
    "    - full_numpy_bitmap_eye.npy\n",
    "    - full_numpy_bitmap_The Eiffel Tower.npy\n",
    "\n",
    "Then you need to store the file in a folder named: <b>quick_draw_dataset</b>\n",
    "\n",
    "[Download the files here](https://console.cloud.google.com/storage/browser/quickdraw_dataset/full/numpy_bitmap/?pli=1)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Dependencies"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Limited tf.compat.v2.summary API due to missing TensorBoard installation\n",
      "Limited tf.summary API due to missing TensorBoard installation\n"
     ]
    }
   ],
   "source": [
    "import tensorflow as tf\n",
    "from sklearn.model_selection import train_test_split #pip install sklearn\n",
    "from sklearn.preprocessing import StandardScaler\n",
    "import matplotlib.pyplot as plt\n",
    "import numpy as np\n",
    "import os"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Be sure to used Tensorflow 2.0"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "assert hasattr(tf, \"function\") # Be sure to use tensorflow 2.0"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Load the dataset"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "images.shape (10050, 28, 28)\n",
      "targets.shape (10050,)\n",
      "images_valid.shape (4950, 28, 28)\n",
      "targets_valid.shape (4950,)\n",
      "['face', 'brain', 'The Eiffel Tower', 'chair', 'eye', 'airplane', 'apple', 'car', 'book', 'dog']\n"
     ]
    }
   ],
   "source": [
    "dataset_dir = \"quick_draw_dataset\"\n",
    "files = [name for name in os.listdir(dataset_dir) if \".npy\" in name]\n",
    "max_size_per_cl = 1500\n",
    "draw_class = []\n",
    "\n",
    "# Evalueate the size of the dataset\n",
    "size = 0\n",
    "for name in files:\n",
    "    draws = np.load(os.path.join(dataset_dir, name))\n",
    "    draws = draws[:max_size_per_cl] # Take only 10 000 draw\n",
    "    size += draws.shape[0]\n",
    "\n",
    "images = np.zeros((size, 28, 28))\n",
    "targets = np.zeros((size,))\n",
    "\n",
    "it = 0\n",
    "t = 0\n",
    "for name in files:\n",
    "    # Open each dataset and add the new class\n",
    "    draw_class.append(name.replace(\"full_numpy_bitmap_\", \"\").replace(\".npy\", \"\"))\n",
    "    draws = np.load(os.path.join(dataset_dir, name))\n",
    "    draws = draws[:max_size_per_cl] # Take only 10 000 draw\n",
    "    # Add images to the buffer\n",
    "    images[it:it+draws.shape[0]] = np.invert(draws.reshape(-1, 28, 28))\n",
    "    targets[it:it+draws.shape[0]] = t\n",
    "    # Iter\n",
    "    it += draws.shape[0]\n",
    "    t += 1\n",
    "\n",
    "images = images.astype(np.float32)\n",
    "    \n",
    "# Shuffle dataset\n",
    "indexes = np.arange(size)\n",
    "np.random.shuffle(indexes)\n",
    "images = images[indexes]\n",
    "targets = targets[indexes]\n",
    "\n",
    "images, images_valid, targets, targets_valid = train_test_split(images, targets, test_size=0.33)\n",
    "\n",
    "print(\"images.shape\", images.shape)\n",
    "print(\"targets.shape\", targets.shape)\n",
    "\n",
    "print(\"images_valid.shape\", images_valid.shape)\n",
    "print(\"targets_valid.shape\", targets_valid.shape)\n",
    "\n",
    "print(draw_class)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Plot exemples of images"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 576x576 with 20 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "w=10\n",
    "h=10\n",
    "fig=plt.figure(figsize=(8, 8))\n",
    "columns = 4\n",
    "rows = 5\n",
    "for i in range(1, columns*rows +1):\n",
    "    index = np.random.randint(len(images))\n",
    "    img = images[index]\n",
    "    fig.add_subplot(rows, columns, i)\n",
    "    plt.title(draw_class[int(targets[index])]) \n",
    "    plt.imshow(img, cmap=\"gray\")\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Normalization"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Mean and std of images 209.48892 85.187294\n",
      "Mean and std of scaled images 2.3239094e-10 0.9285708\n"
     ]
    }
   ],
   "source": [
    "print(\"Mean and std of images\", images.mean(), images.std())\n",
    "scaler = StandardScaler()\n",
    "scaled_images = scaler.fit_transform(images.reshape(-1, 28*28))\n",
    "scaled_images_valid = scaler.transform(images_valid.reshape(-1, 28*28))\n",
    "print(\"Mean and std of scaled images\", scaled_images.mean(), scaled_images.std())\n",
    "\n",
    "scaled_images = scaled_images.reshape(-1, 28, 28, 1)\n",
    "scaled_images_valid = scaled_images_valid.reshape(-1, 28, 28, 1)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Handle dataset\n",
    "### Create the dataset"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [],
   "source": [
    "train_dataset = tf.data.Dataset.from_tensor_slices(scaled_images)\n",
    "valid_dataset = tf.data.Dataset.from_tensor_slices(scaled_images_valid)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Iter in dataset"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(28, 28, 1)\n"
     ]
    }
   ],
   "source": [
    "for item in train_dataset:\n",
    "    print(item.shape)\n",
    "    break"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(32, 28, 28, 1)\n"
     ]
    }
   ],
   "source": [
    "# Iter in the dataset with a number of epoch and batch size\n",
    "\n",
    "epoch = 1\n",
    "batch_size = 32\n",
    "for batch_training in train_dataset.repeat(epoch).batch(32):\n",
    "    print(batch_training.shape)\n",
    "    break"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Create the training dataset"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [],
   "source": [
    "train_dataset = tf.data.Dataset.from_tensor_slices((scaled_images, targets))\n",
    "valid_dataset = tf.data.Dataset.from_tensor_slices((scaled_images_valid, targets_valid))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(32, 28, 28, 1) (32,)\n"
     ]
    }
   ],
   "source": [
    "# Iter in the dataset with a number of epoch and batch size\n",
    "epoch = 1\n",
    "batch_size = 32\n",
    "for images_batch, targets_batch in train_dataset.repeat(epoch).batch(batch_size):\n",
    "    print(images_batch.shape, targets_batch.shape)\n",
    "    break"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Build a model with the subclassing API"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "<img src=\"./images/cnn.png\"></img>"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([[0.11286189, 0.07945979, 0.11112614, 0.10460607, 0.08639237,\n",
       "        0.09611528, 0.10053033, 0.10822048, 0.09531808, 0.10536952]],\n",
       "      dtype=float32)"
      ]
     },
     "execution_count": 12,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "class ConvModel(tf.keras.Model):\n",
    "    \n",
    "    def __init__(self):\n",
    "        super(ConvModel, self).__init__()\n",
    "        # Convolutions\n",
    "        self.conv1 = tf.keras.layers.Conv2D(32, 4, activation='relu', name=\"conv1\")\n",
    "        self.conv2 = tf.keras.layers.Conv2D(64, 3, activation='relu', name=\"conv2\")\n",
    "        self.conv3 = tf.keras.layers.Conv2D(128, 3, activation='relu', name=\"conv3\")\n",
    "        # Flatten the convolution\n",
    "        self.flatten = tf.keras.layers.Flatten(name=\"flatten\")       \n",
    "        # Dense layers\n",
    "        self.d1 = tf.keras.layers.Dense(128, activation='relu', name=\"d1\")\n",
    "        self.out = tf.keras.layers.Dense(10, activation='softmax', name=\"output\")\n",
    "\n",
    "    def call(self, image):\n",
    "        conv1 = self.conv1(image)\n",
    "        conv2 = self.conv2(conv1)\n",
    "        conv3 = self.conv3(conv2)\n",
    "        flatten = self.flatten(conv3)\n",
    "        d1 = self.d1(flatten)\n",
    "        output = self.out(d1)\n",
    "        return output\n",
    "\n",
    "model = ConvModel()\n",
    "model.predict(scaled_images[0:1])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Set the loss and objectives"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [],
   "source": [
    "loss_object = tf.keras.losses.SparseCategoricalCrossentropy()\n",
    "optimizer = tf.keras.optimizers.Adam()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Set some metrics to track the progress of the training"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Loss\n",
    "train_loss = tf.keras.metrics.Mean(name='train_loss')\n",
    "valid_loss = tf.keras.metrics.Mean(name='valid_loss')\n",
    "# Accuracy\n",
    "train_accuracy = tf.keras.metrics.SparseCategoricalAccuracy(name='train_accuracy')\n",
    "valid_accuracy = tf.keras.metrics.SparseCategoricalAccuracy(name='valid_accuracy')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Training method\n",
    "The train_step method used autograp, compute the gradient to optimize the model and gather the loss and accuracy on the training set"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [],
   "source": [
    "@tf.function\n",
    "def train_step(image, targets):\n",
    "    with tf.GradientTape() as tape:\n",
    "        # Make a prediction on all the batch\n",
    "        predictions = model(image)\n",
    "        # Get the error/loss on these predictions\n",
    "        loss = loss_object(targets, predictions)\n",
    "    # Compute the gradient which respect to the loss\n",
    "    gradients = tape.gradient(loss, model.trainable_variables)\n",
    "    # Change the weights of the model\n",
    "    optimizer.apply_gradients(zip(gradients, model.trainable_variables))\n",
    "    # The metrics are accumulate over time. You don't need to average it yourself.\n",
    "    train_loss(loss)\n",
    "    train_accuracy(targets, predictions)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Validation method\n",
    "The valid_step function is simply used to test the image on the validation set and to gather the loss and accuracy"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {},
   "outputs": [],
   "source": [
    "@tf.function\n",
    "def valid_step(image, targets):\n",
    "    predictions = model(image)\n",
    "    t_loss = loss_object(targets, predictions)\n",
    "    # Set the metrics for the test\n",
    "    valid_loss(t_loss)\n",
    "    valid_accuracy(targets, predictions)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Train and test the model on the validation set"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      " Batch 10048/10050, Loss: 0.6301611661911011, Accuracy: 80.57711791992188\n",
      "Epoch 1, Valid Loss: 0.41171038150787354, Valid Accuracy: 87.77777862548828\n",
      " Batch 20128/10050, Loss: 0.2828487753868103, Accuracy: 91.034820556640626\n",
      "Epoch 2, Valid Loss: 0.47590312361717224, Valid Accuracy: 86.62625885009766\n",
      " Batch 30208/10050, Loss: 0.14437039196491241, Accuracy: 95.44278717041016\n",
      "Epoch 3, Valid Loss: 0.6771371364593506, Valid Accuracy: 84.6868667602539\n",
      " Batch 40288/10050, Loss: 0.08336804807186127, Accuracy: 97.064674377441443\n",
      "Epoch 4, Valid Loss: 0.669844388961792, Valid Accuracy: 86.7272720336914\n",
      " Batch 50368/10050, Loss: 0.052768390625715256, Accuracy: 98.37811279296875\n",
      "Epoch 5, Valid Loss: 0.6764260530471802, Valid Accuracy: 86.94950103759766\n",
      " Batch 60448/10050, Loss: 0.031841933727264404, Accuracy: 98.76617431640625\n",
      "Epoch 6, Valid Loss: 0.75545734167099, Valid Accuracy: 88.28282928466797\n",
      " Batch 70528/10050, Loss: 0.03313273563981056, Accuracy: 98.965171813964841\n",
      "Epoch 7, Valid Loss: 0.8147743940353394, Valid Accuracy: 87.21212005615234\n",
      " Batch 80608/10050, Loss: 0.02087359130382538, Accuracy: 99.333335876464841\n",
      "Epoch 8, Valid Loss: 0.857755720615387, Valid Accuracy: 87.81818389892578\n",
      " Batch 90688/10050, Loss: 0.017239892855286598, Accuracy: 99.49253845214844\n",
      "Epoch 9, Valid Loss: 0.9015608429908752, Valid Accuracy: 87.37374114990234\n",
      " Batch 100768/10050, Loss: 0.017793819308280945, Accuracy: 99.39303588867188\n",
      "Epoch 10, Valid Loss: 0.8246178030967712, Valid Accuracy: 86.98989868164062\n"
     ]
    }
   ],
   "source": [
    "epoch = 10\n",
    "batch_size = 32\n",
    "b = 0\n",
    "for epoch in range(epoch):\n",
    "    # Training set\n",
    "    for images_batch, targets_batch in train_dataset.batch(batch_size):\n",
    "        train_step(images_batch, targets_batch)\n",
    "        template = '\\r Batch {}/{}, Loss: {}, Accuracy: {}'\n",
    "        print(template.format(\n",
    "            b, len(targets), train_loss.result(), \n",
    "            train_accuracy.result()*100\n",
    "        ), end=\"\")\n",
    "        b += batch_size\n",
    "    # Validation set\n",
    "    for images_batch, targets_batch in valid_dataset.batch(batch_size):\n",
    "        valid_step(images_batch, targets_batch)\n",
    "\n",
    "    template = '\\nEpoch {}, Valid Loss: {}, Valid Accuracy: {}'\n",
    "    print(template.format(\n",
    "        epoch+1,\n",
    "        valid_loss.result(), \n",
    "        valid_accuracy.result()*100)\n",
    "    )\n",
    "    valid_loss.reset_states()\n",
    "    valid_accuracy.reset_states()\n",
    "    train_accuracy.reset_states()\n",
    "    train_loss.reset_states()"
   ]
  },
  {
   "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": "3.6.7"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}