food_recognition_resnet.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {
    "execution": {
     "iopub.execute_input": "2021-12-19T22:14:27.835087Z",
     "iopub.status.busy": "2021-12-19T22:14:27.834600Z",
     "iopub.status.idle": "2021-12-19T22:14:32.368958Z",
     "shell.execute_reply": "2021-12-19T22:14:32.368231Z",
     "shell.execute_reply.started": "2021-12-19T22:14:27.834915Z"
    },
    "id": "ndWwn6igj8M6",
    "pycharm": {
     "name": "#%%\n"
    }
   },
   "outputs": [],
   "source": [
    "from keras.applications.resnet_v2 import ResNet50V2, ResNet101V2\n",
    "\n",
    "import datetime\n",
    "import keras\n",
    "import matplotlib.pyplot as plt\n",
    "import pandas as pd\n",
    "from pathlib import Path\n",
    "import numpy as np\n",
    "import tensorflow as tf\n",
    "\n",
    "from keras_preprocessing.image import ImageDataGenerator\n",
    "from keras.models import Sequential\n",
    "from keras.losses import CategoricalCrossentropy\n",
    "from keras.optimizers import adam_v2, rmsprop_v2\n",
    "from keras.layers import Input, Conv2D, MaxPool2D, Flatten, Dense, Dropout, BatchNormalization, Layer\n",
    "from typing import List, Set"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {
    "execution": {
     "iopub.execute_input": "2021-12-19T22:14:32.371044Z",
     "iopub.status.busy": "2021-12-19T22:14:32.370805Z",
     "iopub.status.idle": "2021-12-19T22:14:51.353289Z",
     "shell.execute_reply": "2021-12-19T22:14:51.352218Z",
     "shell.execute_reply.started": "2021-12-19T22:14:32.371012Z"
    },
    "id": "0mhytESUrSpZ",
    "outputId": "63a66722-3754-40c9-85aa-e30ad6804baf"
   },
   "outputs": [],
   "source": [
    "!pip install gdown"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {
    "execution": {
     "iopub.execute_input": "2021-12-19T22:14:51.354995Z",
     "iopub.status.busy": "2021-12-19T22:14:51.354737Z",
     "iopub.status.idle": "2021-12-19T22:15:00.246977Z",
     "shell.execute_reply": "2021-12-19T22:15:00.246178Z",
     "shell.execute_reply.started": "2021-12-19T22:14:51.354958Z"
    },
    "id": "jeZ8hE3BmbgN",
    "outputId": "553a1423-0b15-4a0a-d53f-3a8125f31293"
   },
   "outputs": [],
   "source": [
    "# Download dataset\n",
    "!gdown https://drive.google.com/uc?id=1MrjIi8zPwjoUIpwzn7UlomedT18i4gaD"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {
    "execution": {
     "iopub.execute_input": "2021-12-19T22:15:00.250455Z",
     "iopub.status.busy": "2021-12-19T22:15:00.249577Z",
     "iopub.status.idle": "2021-12-19T22:15:09.296341Z",
     "shell.execute_reply": "2021-12-19T22:15:09.295069Z",
     "shell.execute_reply.started": "2021-12-19T22:15:00.250395Z"
    },
    "id": "Xi_ZdVZspZVH",
    "outputId": "cd5813b7-a1ad-4fc8-a43d-0ca858f56b84"
   },
   "outputs": [],
   "source": [
    "!unzip -q \"food-recognition-challenge-2021.zip\" -d dataset"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {
    "execution": {
     "iopub.execute_input": "2021-12-19T22:15:09.300149Z",
     "iopub.status.busy": "2021-12-19T22:15:09.299808Z",
     "iopub.status.idle": "2021-12-19T22:15:13.583528Z",
     "shell.execute_reply": "2021-12-19T22:15:13.582597Z",
     "shell.execute_reply.started": "2021-12-19T22:15:09.300105Z"
    },
    "id": "01CaqPnktVlo"
   },
   "outputs": [],
   "source": [
    "!cp dataset/train_set/train_set/* dataset/train_set/\n",
    "!rm -rf dataset/train_set/train_set"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {
    "execution": {
     "iopub.execute_input": "2021-12-19T22:15:13.585882Z",
     "iopub.status.busy": "2021-12-19T22:15:13.585559Z",
     "iopub.status.idle": "2021-12-19T22:15:15.411975Z",
     "shell.execute_reply": "2021-12-19T22:15:15.411134Z",
     "shell.execute_reply.started": "2021-12-19T22:15:13.585834Z"
    },
    "id": "cFTA5T_Mrwob"
   },
   "outputs": [],
   "source": [
    "INPUT_DATASET = './dataset/train_set'             # Path to a folder that contains all training images\n",
    "PATH_LABELS = './dataset/train_labels.csv'        # Path to CSV file containing training labels\n",
    "PATH_DESTINATION = './dataset/train_set_labelled' # Where new folder will be created with dataset organised in folders by label\n",
    "\n",
    "# Get a dict mapping file name to its label\n",
    "# Example: {'train_1.jpg': 21}\n",
    "file_label_dict = pd.read_csv(PATH_LABELS).set_index('img_name').to_dict()['label']\n",
    "\n",
    "for file in Path(INPUT_DATASET).iterdir():\n",
    "    label = file_label_dict[file.name]\n",
    "    dest_path = Path(PATH_DESTINATION) / str(label)\n",
    "    # Create missing directories if necessary\n",
    "    dest_path.mkdir(parents=True, exist_ok=True)\n",
    "    # Copy the file to the folder with a correct label\n",
    "    file.rename(dest_path / file.name)\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {
    "execution": {
     "iopub.execute_input": "2021-12-19T22:15:15.413907Z",
     "iopub.status.busy": "2021-12-19T22:15:15.413601Z",
     "iopub.status.idle": "2021-12-19T22:15:15.423959Z",
     "shell.execute_reply": "2021-12-19T22:15:15.423170Z",
     "shell.execute_reply.started": "2021-12-19T22:15:15.413866Z"
    }
   },
   "outputs": [],
   "source": [
    "for folder in Path(PATH_DESTINATION).iterdir():\n",
    "    new_filename = ''\n",
    "    for char in folder.name:\n",
    "        new_filename += chr(int(char) + 65)\n",
    "    new_filename = new_filename.rjust(3, 'A')\n",
    "    folder.rename(PATH_DESTINATION + '/' + new_filename)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {
    "execution": {
     "iopub.execute_input": "2021-12-19T22:15:15.714712Z",
     "iopub.status.busy": "2021-12-19T22:15:15.714183Z",
     "iopub.status.idle": "2021-12-19T22:15:17.820663Z",
     "shell.execute_reply": "2021-12-19T22:15:17.819670Z",
     "shell.execute_reply.started": "2021-12-19T22:15:15.714661Z"
    },
    "id": "_038uxn6j8M8",
    "outputId": "645f1869-02a2-4d8b-8cde-f4f219bf1f13",
    "pycharm": {
     "name": "#%%\n"
    }
   },
   "outputs": [],
   "source": [
    "# TODO Find a way to turn off that red debugging spam from tensorflow, this does not work\n",
    "tf.get_logger().setLevel('WARN')\n",
    "\n",
    "print(f'Using GPU {tf.test.gpu_device_name()}')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {
    "execution": {
     "iopub.execute_input": "2021-12-19T22:15:17.822650Z",
     "iopub.status.busy": "2021-12-19T22:15:17.822190Z",
     "iopub.status.idle": "2021-12-19T22:15:18.146736Z",
     "shell.execute_reply": "2021-12-19T22:15:18.145760Z",
     "shell.execute_reply.started": "2021-12-19T22:15:17.822609Z"
    },
    "id": "KAxpYOo-j8M9",
    "outputId": "94976dd9-d3dd-4208-eb67-5a08716a5622",
    "pycharm": {
     "name": "#%%\n"
    }
   },
   "outputs": [],
   "source": [
    "# Global params and constants\n",
    "WIDTH = 224\n",
    "HEIGHT = 224\n",
    "BATCH_SIZE = 32\n",
    "EPOCHS = 50\n",
    "TRAIN_IMAGES_PATH = r'./dataset/train_set_labelled'\n",
    "TEST_IMAGES_PATH = r'./dataset/test_set'\n",
    "TRAIN_LABELS_PATH = r'./dataset/train_labels.csv'\n",
    "PREDICTIONS_PATH = r'predictions.csv'\n",
    "NUM_EXAMPLES = len(list(Path(TRAIN_IMAGES_PATH).rglob('*.jpg')))\n",
    "NUM_CLASSES = len(list(Path(TRAIN_IMAGES_PATH).iterdir()))\n",
    "print(f'Num classes: {NUM_CLASSES}  num samples: {NUM_EXAMPLES}')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {
    "execution": {
     "iopub.execute_input": "2021-12-19T22:15:18.151432Z",
     "iopub.status.busy": "2021-12-19T22:15:18.151027Z",
     "iopub.status.idle": "2021-12-19T22:15:20.776976Z",
     "shell.execute_reply": "2021-12-19T22:15:20.774440Z",
     "shell.execute_reply.started": "2021-12-19T22:15:18.151350Z"
    },
    "id": "BdWM7K4Xj8M-",
    "outputId": "4b90dd61-a0c0-4c31-8811-4894fc61a168",
    "pycharm": {
     "name": "#%%\n"
    }
   },
   "outputs": [],
   "source": [
    "# Generators allow to get the data in batches without having to worry about the memory\n",
    "generator = ImageDataGenerator(\n",
    "    validation_split=0.1,\n",
    "    featurewise_center=True,\n",
    "    featurewise_std_normalization=True,\n",
    "    rotation_range=90,\n",
    "    width_shift_range=0.2,\n",
    "    height_shift_range=0.2,\n",
    "    rescale=1. // 255,\n",
    "    shear_range=0.2,\n",
    "    zoom_range=0.2,\n",
    "    fill_mode='nearest',\n",
    "    brightness_range=[0.8, 1.2]\n",
    ")\n",
    "train_gen = generator.flow_from_directory(\n",
    "    directory=TRAIN_IMAGES_PATH,\n",
    "    class_mode='categorical',\n",
    "    batch_size=BATCH_SIZE,\n",
    "    target_size=(WIDTH, HEIGHT),\n",
    "    shuffle=True,\n",
    "    subset='training'\n",
    ")\n",
    "validation_gen = generator.flow_from_directory(\n",
    "    directory=TRAIN_IMAGES_PATH,\n",
    "    class_mode='categorical',\n",
    "    batch_size=BATCH_SIZE,\n",
    "    target_size=(WIDTH, HEIGHT),\n",
    "    shuffle=True,\n",
    "    subset='validation'\n",
    ")\n",
    "test_gen = generator.flow_from_directory(\n",
    "    directory=TEST_IMAGES_PATH,\n",
    "    class_mode=None,\n",
    "    batch_size=BATCH_SIZE,\n",
    "    target_size=(WIDTH, HEIGHT),\n",
    "    shuffle=False\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {
    "execution": {
     "iopub.execute_input": "2021-12-19T22:15:20.786449Z",
     "iopub.status.busy": "2021-12-19T22:15:20.786018Z",
     "iopub.status.idle": "2021-12-19T22:15:23.792935Z",
     "shell.execute_reply": "2021-12-19T22:15:23.792255Z",
     "shell.execute_reply.started": "2021-12-19T22:15:20.786368Z"
    },
    "id": "LrKPobvGj8M_",
    "outputId": "f56bd335-f8d8-45ba-92b4-6a5ac2caab17",
    "pycharm": {
     "name": "#%%\n"
    }
   },
   "outputs": [],
   "source": [
    "from matplotlib import pyplot as plt\n",
    "\n",
    "for _ in range(5):\n",
    "    img, label = train_gen.next()\n",
    "    print(np.max(img[0][0]))\n",
    "    plt.imshow(img[0].astype(np.uint8))\n",
    "    plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {
    "execution": {
     "iopub.execute_input": "2021-12-19T22:15:23.796679Z",
     "iopub.status.busy": "2021-12-19T22:15:23.796146Z",
     "iopub.status.idle": "2021-12-19T22:15:23.804787Z",
     "shell.execute_reply": "2021-12-19T22:15:23.803832Z",
     "shell.execute_reply.started": "2021-12-19T22:15:23.796639Z"
    },
    "id": "YcBELpUxj8M_",
    "pycharm": {
     "name": "#%%\n"
    }
   },
   "outputs": [],
   "source": [
    "def make_predictions(model: keras.Model, test_gen: ImageDataGenerator):\n",
    "    \"\"\"\n",
    "    Output a CSV with model's predictions on test set that is ready to be submitted to Kaggle.\n",
    "    The file will be created in the main directory of the project, named 'predictions <current_time>'\n",
    "    \"\"\"\n",
    "    predictions = model.predict(test_gen, verbose=True, batch_size=BATCH_SIZE)\n",
    "    # Get names of test files in the same order they were used for predictions\n",
    "    file_names = list(map(lambda x: x.split('/')[1], test_gen.filenames))\n",
    "    # Obtain final labels for predictions, add one since classes start from one\n",
    "    predictions = predictions.argmax(axis=1) + 1\n",
    "    result = pd.DataFrame({'img_name': file_names, 'label': predictions})\n",
    "    result = result.set_index('img_name') \n",
    "    # Save the CSV file to main project directory\n",
    "    result.to_csv(f'predictions {datetime.datetime.now().strftime(\"%d-%m-%Y %Hh %Mm %Ss\")}')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "metadata": {
    "execution": {
     "iopub.execute_input": "2021-12-19T22:16:36.133304Z",
     "iopub.status.busy": "2021-12-19T22:16:36.133052Z",
     "iopub.status.idle": "2021-12-19T22:16:37.871140Z",
     "shell.execute_reply": "2021-12-19T22:16:37.870429Z",
     "shell.execute_reply.started": "2021-12-19T22:16:36.133276Z"
    },
    "id": "ZjxVRsYXj8NB",
    "outputId": "a1b74be8-f8e0-47b0-c738-52f4d2e92ac3",
    "pycharm": {
     "name": "#%%\n"
    }
   },
   "outputs": [],
   "source": [
    "from tensorflow.keras.applications import ResNet50\n",
    "\n",
    "feature_extractor = ResNet50(weights='imagenet', \n",
    "                             input_shape=(WIDTH, HEIGHT, 3),\n",
    "                             include_top=False)\n",
    "\n",
    "num_layers = len(feature_extractor.layers)\n",
    "for layer in feature_extractor.layers[:num_layers // 2]:\n",
    "    layer.trainable = False\n",
    "\n",
    "model = Sequential()\n",
    "model.add(tf.keras.Input(shape=(WIDTH, HEIGHT, 3)))\n",
    "model.add(feature_extractor)\n",
    "model.add(tf.keras.layers.GlobalAveragePooling2D())\n",
    "model.add(Dropout(0.5))\n",
    "model.add(BatchNormalization())\n",
    "model.add(tf.keras.layers.Dense(NUM_CLASSES, activation='softmax'))\n",
    "\n",
    "model.compile(optimizer='adam',\n",
    "             loss='categorical_crossentropy',\n",
    "             metrics=['accuracy'])\n",
    "\n",
    "model.summary()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {
    "execution": {
     "iopub.execute_input": "2021-12-19T22:16:48.570126Z",
     "iopub.status.busy": "2021-12-19T22:16:48.569868Z",
     "iopub.status.idle": "2021-12-19T22:16:49.274248Z",
     "shell.execute_reply": "2021-12-19T22:16:49.273189Z",
     "shell.execute_reply.started": "2021-12-19T22:16:48.570096Z"
    },
    "id": "0svaRww1F70L"
   },
   "outputs": [],
   "source": [
    "!mkdir training\n",
    "checkpoint_filepath = 'training/weights.{epoch:02d}-{val_loss:.2f}.hdf5'\n",
    "model_checkpoint_callback = tf.keras.callbacks.ModelCheckpoint(\n",
    "    filepath=checkpoint_filepath,\n",
    "    save_weights_only=True,\n",
    "    monitor='val_accuracy',\n",
    "    mode='max',\n",
    "    save_best_only=True)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "metadata": {
    "execution": {
     "iopub.execute_input": "2021-12-19T22:16:50.716642Z",
     "iopub.status.busy": "2021-12-19T22:16:50.715989Z",
     "iopub.status.idle": "2021-12-19T22:17:01.522573Z",
     "shell.execute_reply": "2021-12-19T22:17:01.519921Z",
     "shell.execute_reply.started": "2021-12-19T22:16:50.716601Z"
    },
    "id": "xrLzhBC19dvx",
    "outputId": "1d206005-2b4c-4105-9df9-b8ba45d00777"
   },
   "outputs": [],
   "source": [
    "model.fit(train_gen, epochs=EPOCHS, validation_data=validation_gen, callbacks=[model_checkpoint_callback])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "execution": {
     "iopub.status.busy": "2021-12-19T22:16:24.902628Z",
     "iopub.status.idle": "2021-12-19T22:16:24.902937Z",
     "shell.execute_reply": "2021-12-19T22:16:24.902797Z",
     "shell.execute_reply.started": "2021-12-19T22:16:24.902775Z"
    }
   },
   "outputs": [],
   "source": [
    "make_predictions(model, test_gen)"
   ]
  }
 ],
 "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": "3.8.8"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 4
}