{

  "nbformat": 4,

  "nbformat_minor": 0,

  "metadata": {

    "kernelspec": {

      "display_name": "Python [conda env:datasci] *",

      "language": "python",

      "name": "conda-env-datasci-py"

    },

    "language_info": {

      "codemirror_mode": {

        "name": "ipython",

        "version": 3

      },

      "file_extension": ".py",

      "mimetype": "text/x-python",

      "name": "python",

      "nbconvert_exporter": "python",

      "pygments_lexer": "ipython3",

      "version": "3.7.5"

    },

    "colab": {

      "name": "Training the CNN.ipynb",

      "provenance": [],

      "collapsed_sections": []

    }

  },

  "cells": [

    {

      "cell_type": "markdown",

      "metadata": {

        "id": "efCvb66QStgP",

        "colab_type": "text"

      },

      "source": [

        "*To run this notebook, please provide the following four file paths:*"

      ]

    },

    {

      "cell_type": "code",

      "metadata": {

        "id": "Iukqj1IzStyn",

        "colab_type": "code",

        "colab": {}

      },

      "source": [

        "path_to_train = '/path/to/train/images'\n",

        "path_to_test = '/path/to/test/images'\n",

        "path_to_labels = '/path/to/labels.csv'\n",

        "\n",

        "path_to_save_model = '/path/to/save/model/to/cnn-checkpoint-{epoch:02d}-{val_accuracy:.2f}.hdf5'"

      ],

      "execution_count": 0,

      "outputs": []

    },

    {

      "cell_type": "markdown",

      "metadata": {

        "id": "XqBOX3Y4MWs3",

        "colab_type": "text"

      },

      "source": [

        "## **Installing & Importing Dependencies**"

      ]

    },

    {

      "cell_type": "code",

      "metadata": {

        "id": "jnq7FwiHKZwY",

        "colab_type": "code",

        "colab": {}

      },

      "source": [

        "!pip install keras"

      ],

      "execution_count": 0,

      "outputs": []

    },

    {

      "cell_type": "code",

      "metadata": {

        "id": "OTnWqAwiKZxb",

        "colab_type": "code",

        "colab": {}

      },

      "source": [

        "import tensorflow as tf\n",

        "import pandas as pd\n",

        "\n",

        "from keras import applications\n",

        "from keras import optimizers\n",

        "from keras import backend as k \n",

        "\n",

        "from keras.models import Sequential, Model\n",

        "from keras.layers import Dropout, Flatten, Dense, GlobalAveragePooling2D\n",

        "from keras.callbacks import ModelCheckpoint, LearningRateScheduler, TensorBoard, EarlyStopping\n",

        "from keras.preprocessing.image import ImageDataGenerator"

      ],

      "execution_count": 0,

      "outputs": []

    },

    {

      "cell_type": "code",

      "metadata": {

        "id": "N60M6zljKZxs",

        "colab_type": "code",

        "colab": {}

      },

      "source": [

        "# Make sure that GPU is available on the machine\n",

        "assert tf.test.is_gpu_available()\n",

        "assert tf.test.is_built_with_cuda()"

      ],

      "execution_count": 0,

      "outputs": []

    },

    {

      "cell_type": "markdown",

      "metadata": {

        "id": "PIErVpE6NHx6",

        "colab_type": "text"

      },

      "source": [

        "## **Reading in Data Labels**"

      ]

    },

    {

      "cell_type": "code",

      "metadata": {

        "id": "SGUEeB-cKZx3",

        "colab_type": "code",

        "colab": {}

      },

      "source": [

        "labels_df = pd.read_csv(path_to_labels)\n",

        "\n",

        "# For flow_from_dataframe to function, string datatype is required\n",

        "labels_df = labels_df.astype(str)"

      ],

      "execution_count": 0,

      "outputs": []

    },

    {

      "cell_type": "code",

      "metadata": {

        "scrolled": true,

        "id": "qShGxNkPKZx7",

        "colab_type": "code",

        "colab": {},

        "outputId": "f6cf0aa8-a383-440a-e4ed-13499e3669af"

      },

      "source": [

        "# Inspect the DataFrame containing the labels\n",

        "labels_df.head(3)"

      ],

      "execution_count": 0,

      "outputs": [

        {

          "output_type": "execute_result",

          "data": {

            "text/html": [

              "<div>\n",

              "<style scoped>\n",

              "    .dataframe tbody tr th:only-of-type {\n",

              "        vertical-align: middle;\n",

              "    }\n",

              "\n",

              "    .dataframe tbody tr th {\n",

              "        vertical-align: top;\n",

              "    }\n",

              "\n",

              "    .dataframe thead th {\n",

              "        text-align: right;\n",

              "    }\n",

              "</style>\n",

              "<table border=\"1\" class=\"dataframe\">\n",

              "  <thead>\n",

              "    <tr style=\"text-align: right;\">\n",

              "      <th></th>\n",

              "      <th>Unnamed: 0</th>\n",

              "      <th>ID</th>\n",

              "      <th>any</th>\n",

              "    </tr>\n",

              "  </thead>\n",

              "  <tbody>\n",

              "    <tr>\n",

              "      <th>0</th>\n",

              "      <td>0</td>\n",

              "      <td>ID_000039fa0.png</td>\n",

              "      <td>0</td>\n",

              "    </tr>\n",

              "    <tr>\n",

              "      <th>1</th>\n",

              "      <td>1</td>\n",

              "      <td>ID_00005679d.png</td>\n",

              "      <td>0</td>\n",

              "    </tr>\n",

              "    <tr>\n",

              "      <th>2</th>\n",

              "      <td>2</td>\n",

              "      <td>ID_00008ce3c.png</td>\n",

              "      <td>0</td>\n",

              "    </tr>\n",

              "  </tbody>\n",

              "</table>\n",

              "</div>"

            ],

            "text/plain": [

              "  Unnamed: 0                ID any\n",

              "0          0  ID_000039fa0.png   0\n",

              "1          1  ID_00005679d.png   0\n",

              "2          2  ID_00008ce3c.png   0"

            ]

          },

          "metadata": {

            "tags": []

          },

          "execution_count": 19

        }

      ]

    },

    {

      "cell_type": "markdown",

      "metadata": {

        "id": "ElzPsVvaODOk",

        "colab_type": "text"

      },

      "source": [

        "## **Building Transfer-Learning Model**"

      ]

    },

    {

      "cell_type": "markdown",

      "metadata": {

        "id": "qFQ7sFZjx_ba",

        "colab_type": "text"

      },

      "source": [

        "*As our pretrained model, we choose VGG19 with ImageNet weights. Note that include_top = False (otherwise, we would be including VGG19's final 1000-node dense softmax prediction layer)*"

      ]

    },

    {

      "cell_type": "code",

      "metadata": {

        "id": "vh0b4EQtKZyN",

        "colab_type": "code",

        "colab": {}

      },

      "source": [

        "model = applications.VGG19(weights = \"imagenet\", include_top=False, input_shape = (128, 128, 3))"

      ],

      "execution_count": 0,

      "outputs": []

    },

    {

      "cell_type": "markdown",

      "metadata": {

        "id": "tszxYuOwyURR",

        "colab_type": "text"

      },

      "source": [

        "*Freeze the first three convolutional blocks, leave the last two unfrozen. Thereby, we transfer the model's knowledge of low-level features (like edges and angles) while allowing for it to learn new high-level features (like hemorrhages).*"

      ]

    },

    {

      "cell_type": "code",

      "metadata": {

        "id": "BruLVtwJKZ0g",

        "colab_type": "code",

        "colab": {}

      },

      "source": [

        "for layer in model.layers[0:12]:\n",

        "    layer.trainable = False"

      ],

      "execution_count": 0,

      "outputs": []

    },

    {

      "cell_type": "markdown",

      "metadata": {

        "id": "g_Kb-J_Qylgz",

        "colab_type": "text"

      },

      "source": [

        "*We append our own custom layers to the end of VGG19. Note that a single sigmoid final prediction node is equivalent to two softmax final prediction nodes.*"

      ]

    },

    {

      "cell_type": "code",

      "metadata": {

        "id": "Z7uPdX-8c5eU",

        "colab_type": "code",

        "colab": {}

      },

      "source": [

        "x = model.output\n",

        "x = Flatten()(x)\n",

        "\n",

        "x = Dense(1000, activation = 'relu')(x)\n",

        "x = Dropout(0.5)(x)\n",

        "x = Dense(1000, activation = 'relu')(x)\n",

        "output = Dense(1, activation = 'sigmoid')(x)\n",

        "\n",

        "custom_model = Model(inputs = model.input, outputs = output)"

      ],

      "execution_count": 0,

      "outputs": []

    },

    {

      "cell_type": "markdown",

      "metadata": {

        "id": "X49CwEDJy7tN",

        "colab_type": "text"

      },

      "source": [

        "*Compiling the model. The primary metric we care about is recall (that is, the CNN's ability to correctly detect intracranial hemorrhages).*"

      ]

    },

    {

      "cell_type": "code",

      "metadata": {

        "id": "xECXyzILnWTa",

        "colab_type": "code",

        "colab": {}

      },

      "source": [

        "custom_model.compile(loss = 'binary_crossentropy',\n",

        "                     optimizer = optimizers.Adam(lr=0.0001),\n",

        "                     metrics=['accuracy',\n",

        "                              tf.keras.metrics.Recall(),\n",

        "                              tf.keras.metrics.AUC(),\n",

        "                              tf.keras.metrics.Precision()])"

      ],

      "execution_count": 0,

      "outputs": []

    },

    {

      "cell_type": "markdown",

      "metadata": {

        "id": "u7BWA55koyNp",

        "colab_type": "text"

      },

      "source": [

        "# **Creating Train & Test Generators**"

      ]

    },

    {

      "cell_type": "code",

      "metadata": {

        "id": "fgTeFSx5olWu",

        "colab_type": "code",

        "colab": {}

      },

      "source": [

        "# Initializing train & test generators to flow train and test images straight from the folders\n",

        "# that they are stored in\n",

        "train_datagen = ImageDataGenerator(rescale = 1./255,\n",

        "                                   horizontal_flip = True,\n",

        "                                   fill_mode = \"nearest\",\n",

        "                                   zoom_range = 0.3,\n",

        "                                   width_shift_range = 0.3,\n",

        "                                   height_shift_range=0.3,\n",

        "                                   rotation_range=30)\n",

        "\n",

        "test_datagen = ImageDataGenerator(rescale = 1./255)"

      ],

      "execution_count": 0,

      "outputs": []

    },

    {

      "cell_type": "code",

      "metadata": {

        "id": "7-zf69uioyiE",

        "colab_type": "code",

        "colab": {}

      },

      "source": [

        "# We flow from DataFrames, that is, our images are not stored in class-specific folders---instead,\n",

        "# their labels are stored in separate files (specifically, in DataFrames)\n",

        "train_generator = train_datagen.flow_from_dataframe(dataframe=labels_df,\n",

        "                                                    directory= path_to_train,\n",

        "                                                    x_col='ID',\n",

        "                                                    y_col='any',\n",

        "                                                    target_size=(128, 128),\n",

        "                                                    class_mode='binary')\n",

        "\n",

        "test_generator = test_datagen.flow_from_dataframe(dataframe=labels_df,\n",

        "                                                  directory=path_to_test,\n",

        "                                                  x_col='ID',\n",

        "                                                  y_col='any',\n",

        "                                                  target_size=(128, 128),\n",

        "                                                  class_mode='binary')"

      ],

      "execution_count": 0,

      "outputs": []

    },

    {

      "cell_type": "markdown",

      "metadata": {

        "id": "_Xh6lg1tszMo",

        "colab_type": "text"

      },

      "source": [

        "# **Fitting the Model**"

      ]

    },

    {

      "cell_type": "code",

      "metadata": {

        "id": "HpcKttO0qdEe",

        "colab_type": "code",

        "colab": {}

      },

      "source": [

        "# Callbacks\n",

        "checkpoint = ModelCheckpoint(path_to_save_model, monitor='val_acc', verbose=1, save_best_only=False, save_weights_only=False, mode='auto', period=1)\n",

        "early_stopper = EarlyStopping(monitor='val_acc', min_delta=0, patience=3, verbose=1, mode='auto')"

      ],

      "execution_count": 0,

      "outputs": []

    },

    {

      "cell_type": "code",

      "metadata": {

        "id": "5e5xPMF2KZ1h",

        "colab_type": "code",

        "colab": {}

      },

      "source": [

        "# Fitting the model \n",

        "custom_model.fit_generator(train_generator,\n",

        "                           epochs = 50,\n",

        "                           validation_data = test_generator,\n",

        "                           callbacks = [checkpoint, early_stopper])"

      ],

      "execution_count": 0,

      "outputs": []

    }

  ]

}