{"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"name":"python","version":"3.10.15","mimetype":"text/x-python","codemirror_mode":{"name":"ipython","version":3},"pygments_lexer":"ipython3","nbconvert_exporter":"python","file_extension":".py"},"kaggle":{"accelerator":"tpu1vmV38","dataSources":[{"sourceId":21154,"databundleVersionId":1243559,"sourceType":"competition"}],"dockerImageVersionId":30806,"isInternetEnabled":true,"language":"python","sourceType":"notebook","isGpuEnabled":false}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"code","source":"import math, re, os\nimport numpy as np\nimport tensorflow as tf\nfrom tensorflow.keras.models import Model\nfrom tensorflow.keras import layers\nfrom tensorflow.keras.layers import Dense\nfrom tensorflow.keras.callbacks import ModelCheckpoint\nfrom kaggle_datasets import KaggleDatasets\nimport random\nimport matplotlib.pyplot as plt\n%matplotlib inline \nprint(\"Tensorflow version \" + tf.__version__)","metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","trusted":true,"execution":{"iopub.status.busy":"2024-12-19T17:43:00.248500Z","iopub.execute_input":"2024-12-19T17:43:00.249137Z","iopub.status.idle":"2024-12-19T17:43:18.041129Z","shell.execute_reply.started":"2024-12-19T17:43:00.249102Z","shell.execute_reply":"2024-12-19T17:43:18.040418Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"AUTO = tf.data.experimental.AUTOTUNE\n# Detect TPU, return appropriate distribution strategy\ntry:\n    tpu = tf.distribute.cluster_resolver.TPUClusterResolver() \n    print('Running on TPU ', tpu.master())\nexcept ValueError:\n    tpu = None\n\nif tpu:\n    tf.config.experimental_connect_to_cluster(tpu)\n    tf.tpu.experimental.initialize_tpu_system(tpu)\n    strategy = tf.distribute.experimental.TPUStrategy(tpu)\nelse:\n    strategy = tf.distribute.get_strategy() \n\nprint(\"REPLICAS: \", strategy.num_replicas_in_sync)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2024-12-19T17:43:18.042353Z","iopub.execute_input":"2024-12-19T17:43:18.042765Z","iopub.status.idle":"2024-12-19T17:43:26.182375Z","shell.execute_reply.started":"2024-12-19T17:43:18.042738Z","shell.execute_reply":"2024-12-19T17:43:26.181576Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"GCS_DS_PATH = KaggleDatasets().get_gcs_path()","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2024-12-19T17:43:26.183313Z","iopub.execute_input":"2024-12-19T17:43:26.183551Z","iopub.status.idle":"2024-12-19T17:43:26.187493Z","shell.execute_reply.started":"2024-12-19T17:43:26.183526Z","shell.execute_reply":"2024-12-19T17:43:26.186769Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"IMAGE_SIZE = [331, 331] # при таком размере графическому процессору не хватит памяти. Используйте TPU\nEPOCHS = 25\nBATCH_SIZE = 16 * strategy.num_replicas_in_sync\n\nGCS_PATH_SELECT = { # available image sizes\n    192: GCS_DS_PATH + '/tfrecords-jpeg-192x192',\n    224: GCS_DS_PATH + '/tfrecords-jpeg-224x224',\n    331: GCS_DS_PATH + '/tfrecords-jpeg-331x331',\n    512: GCS_DS_PATH + '/tfrecords-jpeg-512x512'\n}\nGCS_PATH = GCS_PATH_SELECT[IMAGE_SIZE[0]]\n\nTRAINING_FILENAMES = tf.io.gfile.glob(GCS_PATH + '/train/*.tfrec')\nVALIDATION_FILENAMES = tf.io.gfile.glob(GCS_PATH + '/val/*.tfrec')\nTEST_FILENAMES = tf.io.gfile.glob(GCS_PATH + '/test/*.tfrec')\n\nSEED = 101","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2024-12-19T17:43:26.189187Z","iopub.execute_input":"2024-12-19T17:43:26.189464Z","iopub.status.idle":"2024-12-19T17:43:26.225296Z","shell.execute_reply.started":"2024-12-19T17:43:26.189437Z","shell.execute_reply":"2024-12-19T17:43:26.224573Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"def random_erasing(img, sl=0.1, sh=0.2, rl=0.4, p=0.3):\n    h = tf.shape(img)[0]\n    w = tf.shape(img)[1]\n    c = tf.shape(img)[2]\n    origin_area = tf.cast(h*w, tf.float32)\n\n    e_size_l = tf.cast(tf.round(tf.sqrt(origin_area * sl * rl)), tf.int32)\n    e_size_h = tf.cast(tf.round(tf.sqrt(origin_area * sh / rl)), tf.int32)\n\n    e_height_h = tf.minimum(e_size_h, h)\n    e_width_h = tf.minimum(e_size_h, w)\n\n    erase_height = tf.random.uniform(shape=[], minval=e_size_l, maxval=e_height_h, dtype=tf.int32)\n    erase_width = tf.random.uniform(shape=[], minval=e_size_l, maxval=e_width_h, dtype=tf.int32)\n\n    erase_area = tf.zeros(shape=[erase_height, erase_width, c])\n    erase_area = tf.cast(erase_area, tf.uint8)\n\n    pad_h = h - erase_height\n    pad_top = tf.random.uniform(shape=[], minval=0, maxval=pad_h, dtype=tf.int32)\n    pad_bottom = pad_h - pad_top\n\n    pad_w = w - erase_width\n    pad_left = tf.random.uniform(shape=[], minval=0, maxval=pad_w, dtype=tf.int32)\n    pad_right = pad_w - pad_left\n\n    erase_mask = tf.pad([erase_area], [[0,0],[pad_top, pad_bottom], [pad_left, pad_right], [0,0]], constant_values=1)\n    erase_mask = tf.squeeze(erase_mask, axis=0)\n    erased_img = tf.multiply(tf.cast(img,tf.float32), tf.cast(erase_mask, tf.float32))\n\n    return tf.cond(tf.random.uniform([], 0, 1) > p, lambda: tf.cast(img, img.dtype), lambda:  tf.cast(erased_img, img.dtype))","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"def decode_image(image_data):\n    \"\"\"Декодирует изображение в vyjujvthye. vfnhbwe (тензор)\n    Нормализует данные и преобразовывает изображения к указанному размеру\"\"\"\n    image = tf.image.decode_jpeg(image_data, channels=3) # Декодирование изображения в формате JPEG в тензор uint8.\n    image = tf.cast(image, tf.float32) / 255.0  # преобразовать изображение в плавающее в диапазоне [0, 1]\n    image = tf.reshape(image, [*IMAGE_SIZE, 3]) # явный размер, необходимый для TPU\n#     image = tf.keras.applications.inception_resnet_v2.preprocess_input(image)\n    return image\n\ndef read_labeled_tfrecord(example):\n    LABELED_TFREC_FORMAT = {\n        \"image\": tf.io.FixedLenFeature([], tf.string), # tf.string означает байтовую строку\n        \"class\": tf.io.FixedLenFeature([], tf.int64),  # [] означает отдельный элемент\n    }\n    example = tf.io.parse_single_example(example, LABELED_TFREC_FORMAT) # парсим отдельный пример в указанном формате\n    image = decode_image(example['image']) # преобразуем изображение к нужному нам формату\n    label = tf.cast(example['class'], tf.int32)\n    return image, label # возвращает набор данных пар (изображение, метка)\n\ndef read_unlabeled_tfrecord(example):\n    UNLABELED_TFREC_FORMAT = {\n        \"image\": tf.io.FixedLenFeature([], tf.string), # tf.string означает байтовую строку\n        \"id\": tf.io.FixedLenFeature([], tf.string),  # [] означает отдельный элемент\n        # класс отсутствует, задача этого конкурса - предсказать классы цветов для тестового набора данных\n    }\n    example = tf.io.parse_single_example(example, UNLABELED_TFREC_FORMAT)\n    image = decode_image(example['image']) # преобразуем изображение к нужному нам формату\n    idnum = example['id']\n    return image, idnum # returns a dataset of image(s)\n\ndef load_dataset(filenames, labeled=True, ordered=False):\n    \"\"\"Читает из TFRecords. Для оптимальной производительности одновременное чтение из нескольких\n    файлов без учета порядка данных. Порядок не имеет значения, поскольку мы все равно будем перетасовывать данные\"\"\"\n\n    ignore_order = tf.data.Options() # Представляет параметры для tf.data.Dataset.\n    if not ordered:\n        ignore_order.experimental_deterministic = False # отключить порядок, увеличить скорость\n\n    dataset = tf.data.TFRecordDataset(filenames, num_parallel_reads=AUTO) # автоматически чередует чтение из нескольких файлов\n    dataset = dataset.with_options(ignore_order) # использует данные сразу после их поступления, а не в исходном порядке\n    dataset = dataset.map(read_labeled_tfrecord if labeled else read_unlabeled_tfrecord, num_parallel_calls=AUTO)\n    # возвращает набор данных пар (изображение, метка), если метка = Истина, или пар (изображение, идентификатор), если метка = Ложь\n    return dataset\n\ndef data_augment(image, label):\n    image = tf.image.random_flip_left_right(image)\n    image = random_erasing(image)\n    return image, label\n\n    # # data augmentation. Thanks to the dataset.prefetch(AUTO) statement in the next function (below),\n    # # this happens essentially for free on TPU. Data pipeline code is executed on the \"CPU\" part\n    # # of the TPU while the TPU itself is computing gradients.\n    # flag = random.randint(1,4)\n    # coef_1 = random.randint(70, 90) * 0.01\n    # coef_2 = random.randint(70, 90) * 0.01\n    # if flag == 1:\n    #     image = tf.image.random_flip_left_right(image, seed=SEED)\n    # elif flag == 2:\n    #     image = tf.image.random_flip_up_down(image, seed=SEED)\n    # #else:\n    #     #image = tf.image.random_crop(image, [int(IMAGE_SIZE[0]*coef_1), int(IMAGE_SIZE[0]*coef_2), 3],seed=SEED)\n    # return image, label   \n\ndef get_training_dataset():\n    dataset = load_dataset(TRAINING_FILENAMES, labeled=True)\n    dataset = dataset.map(data_augment, num_parallel_calls=AUTO)\n    dataset = dataset.repeat() # набор обучающих данных должен повторяться в течение нескольких эпох\n    dataset = dataset.shuffle(2048)\n    dataset = dataset.batch(BATCH_SIZE)\n    dataset = dataset.prefetch(AUTO) #готовим следующий набор, пока предыдущий обучается\n    return dataset\n\ndef get_validation_dataset():\n    dataset = load_dataset(VALIDATION_FILENAMES, labeled=True, ordered=False)\n    dataset = dataset.batch(BATCH_SIZE)\n    dataset = dataset.cache() # кешируем набор\n    dataset = dataset.prefetch(AUTO) #готовим следующий набор, пока предыдущий обучается\n    return dataset\n\ndef get_test_dataset(ordered=False):\n    dataset = load_dataset(TEST_FILENAMES, labeled=False, ordered=ordered)\n    dataset = dataset.batch(BATCH_SIZE)\n    dataset = dataset.prefetch(AUTO) #готовим следующий набор, пока предыдущий обучается\n    return dataset\n                               \ndef count_data_items(filenames):\n    # the number of data items is written in the name of the .tfrec files, i.e. flowers00-230.tfrec = 230 data items\n    n = [int(re.compile(r\"-([0-9]*)\\.\").search(filename).group(1)) for filename in filenames]\n    return np.sum(n)\n\n# training_dataset = get_training_dataset()\n# validation_dataset = get_validation_dataset()\n\nNUM_TRAINING_IMAGES = count_data_items(TRAINING_FILENAMES)\nNUM_VALIDATION_IMAGES = count_data_items(VALIDATION_FILENAMES)\nNUM_TEST_IMAGES = count_data_items(TEST_FILENAMES)\nSTEPS_PER_EPOCH = NUM_TRAINING_IMAGES // BATCH_SIZE\nprint('Dataset: {} training images, {} validation images, {} unlabeled test images'.format(NUM_TRAINING_IMAGES, NUM_VALIDATION_IMAGES, NUM_TEST_IMAGES))","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2024-12-19T17:43:26.226202Z","iopub.execute_input":"2024-12-19T17:43:26.226489Z","iopub.status.idle":"2024-12-19T17:43:26.244116Z","shell.execute_reply.started":"2024-12-19T17:43:26.226458Z","shell.execute_reply":"2024-12-19T17:43:26.243414Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# функция управляющая изменениями шага обучения в процессе тренировки нейронной сети\nLR_START = 0.00001\nLR_MAX = 0.00005 * strategy.num_replicas_in_sync#0.0001\nLR_MIN = 0.00001\nLR_RAMPUP_EPOCHS = 5\nLR_SUSTAIN_EPOCHS = 0\nLR_EXP_DECAY = .8\n\ndef lrfn(epoch):\n    if epoch < LR_RAMPUP_EPOCHS:\n        lr = (LR_MAX - LR_START) / LR_RAMPUP_EPOCHS * epoch + LR_START\n    elif epoch < LR_RAMPUP_EPOCHS + LR_SUSTAIN_EPOCHS:\n        lr = LR_MAX\n    else:\n        lr = (LR_MAX - LR_MIN) * LR_EXP_DECAY**(epoch - LR_RAMPUP_EPOCHS - LR_SUSTAIN_EPOCHS) + LR_MIN\n    return lr\n    \nlr_callback = tf.keras.callbacks.LearningRateScheduler(lrfn, verbose=True)\n\n# построим график изменения шага обучение в зависимости от эпох\nrng = [i for i in range(EPOCHS)]\ny = [lrfn(x) for x in rng]\nplt.plot(rng, y)\nprint(\"Learning rate schedule: {:.3g} to {:.3g} to {:.3g}\".format(y[0], max(y), y[-1]))","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2024-12-19T17:43:26.244901Z","iopub.execute_input":"2024-12-19T17:43:26.245140Z","iopub.status.idle":"2024-12-19T17:43:26.412961Z","shell.execute_reply.started":"2024-12-19T17:43:26.245115Z","shell.execute_reply":"2024-12-19T17:43:26.412348Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"### Функции для создания сетей","metadata":{}},{"cell_type":"code","source":"def get_model_ConvNeXtBase():\n    base_model = tf.keras.applications.ConvNeXtBase(weights='imagenet', \n                      include_top=False, pooling='avg',\n                      input_shape=(*IMAGE_SIZE, 3))\n    # base_model.trainable = False\n    x = base_model.output\n    predictions = Dense(104, activation='softmax')(x)\n    return Model(inputs=base_model.input, outputs=predictions)\n\n\ndef get_model_InceptionResNetV2():\n    base_model = tf.keras.applications.InceptionResNetV2(weights='imagenet', \n                          include_top=False, \n                          pooling='avg',\n                          input_shape=(*IMAGE_SIZE, 3))\n    # base_model.trainable = False\n    x = base_model.output\n    predictions = Dense(104, activation='softmax')(x)\n    return Model(inputs=base_model.input, outputs=predictions)\n\n\ndef get_model_EfficientNetB7():\n    base_model = tf.keras.applications.EfficientNetB7(weights='imagenet', \n                          include_top=False, \n                          pooling='avg',\n                          input_shape=(*IMAGE_SIZE, 3))\n    # base_model.trainable = False\n    x = base_model.output\n    predictions = Dense(104, activation='softmax')(x)\n    return Model(inputs=base_model.input, outputs=predictions)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2024-12-19T17:43:26.413776Z","iopub.execute_input":"2024-12-19T17:43:26.414328Z","iopub.status.idle":"2024-12-19T17:43:26.420120Z","shell.execute_reply.started":"2024-12-19T17:43:26.414303Z","shell.execute_reply":"2024-12-19T17:43:26.419439Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"### Обучение InceptionResNetV2","metadata":{}},{"cell_type":"code","source":"with strategy.scope():    \n    model = get_model_InceptionResNetV2()\n        \n    model.compile(\n        optimizer='adam',\n        loss = 'sparse_categorical_crossentropy',\n        metrics=['sparse_categorical_accuracy']\n    )\n\n    history = model.fit(get_training_dataset(), \n          steps_per_epoch=STEPS_PER_EPOCH, \n          epochs=EPOCHS, \n          callbacks=[lr_callback, ModelCheckpoint(filepath='my_InceptionResNetV2.keras', monitor='val_loss',\n                                  save_best_only=True)],\n          validation_data=get_validation_dataset())","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2024-12-19T17:43:26.420902Z","iopub.execute_input":"2024-12-19T17:43:26.421129Z","iopub.status.idle":"2024-12-19T17:52:11.820782Z","shell.execute_reply.started":"2024-12-19T17:43:26.421107Z","shell.execute_reply":"2024-12-19T17:52:11.819587Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"### Обучение ConvNeXtBase","metadata":{}},{"cell_type":"code","source":"with strategy.scope():    \n    model = get_model_ConvNeXtBase()\n        \n    model.compile(\n        optimizer='adam',\n        loss = 'sparse_categorical_crossentropy',\n        metrics=['sparse_categorical_accuracy']\n    )\n\n    history = model.fit(get_training_dataset(), \n          steps_per_epoch=STEPS_PER_EPOCH, \n          epochs=EPOCHS, \n          callbacks=[lr_callback, ModelCheckpoint(filepath='my_ConvNeXtBase.keras', monitor='val_loss',\n                                  save_best_only=True)],\n          validation_data=get_validation_dataset())","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2024-12-19T17:52:11.822956Z","iopub.execute_input":"2024-12-19T17:52:11.823271Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"### Обучение EfficientNetB7","metadata":{}},{"cell_type":"code","source":"with strategy.scope():    \n    model = get_model_EfficientNetB7()\n        \n    model.compile(\n        optimizer='adam',\n        loss = 'sparse_categorical_crossentropy',\n        metrics=['sparse_categorical_accuracy']\n    )\n\n    history = model.fit(get_training_dataset(), \n          steps_per_epoch=STEPS_PER_EPOCH, \n          epochs=EPOCHS, \n          callbacks=[lr_callback, ModelCheckpoint(filepath='my_EfficientNetB7.keras', monitor='val_loss',\n                                  save_best_only=True)],\n          validation_data=get_validation_dataset())","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"### Загрузка весов обученных моделей","metadata":{}},{"cell_type":"code","source":"with strategy.scope():    \n    model1 = get_model_ConvNeXtBase()\n    model1.load_weights(\"/kaggle/working/my_ConvNeXtBase.keras\")\n    model2 = get_model_InceptionResNetV2()\n    model2.load_weights(\"/kaggle/working/my_InceptionResNetV2.keras\")\n    model3 = get_model_EfficientNetB7()\n    model3.load_weights(\"/kaggle/working/my_EfficientNetB7.keras\")","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"### Подбор коэффициентов","metadata":{}},{"cell_type":"code","source":"from sklearn.metrics import f1_score\nval_dataset = get_validation_dataset()\nimages_ds = val_dataset.map(lambda image, label: image)\nlabels_ds = val_dataset.map(lambda image, label: label).unbatch()\nval_labels = next(iter(labels_ds.batch(NUM_VALIDATION_IMAGES))).numpy() # get everything as one batch\nm1 = model1.predict(images_ds)\nm2 = model2.predict(images_ds)\nm3 = model3.predict(images_ds)\n\nscores = []\nalphas = np.linspace(0, 1, 50)  # Вариация для alpha\nbetas = np.linspace(0, 1, 50)   # Вариация для beta\n\n# Подбираем оптимальные alpha и beta\nfor alpha in alphas:\n    for beta in betas:\n        if alpha + beta > 1:  # Обеспечиваем, что alpha + beta <= 1\n            continue\n        gamma = 1 - alpha - beta  # Вес для m3\n        val_probabilities = alpha * m1 + beta * m2 + gamma * m3\n        val_predictions = np.argmax(val_probabilities, axis=-1)\n        scores.append((f1_score(val_labels, val_predictions, labels=range(104), average='macro'), alpha, beta))\n\n# Находим лучшие alpha и beta\nbest_score, best_alpha, best_beta = max(scores, key=lambda x: x[0])\nbest_gamma = 1 - best_alpha - best_beta\n\nprint(f'Best alpha: {best_alpha}')\nprint(f'Best beta: {best_beta}')\nprint(f'Best gamma: {best_gamma}')\nprint(f'Best F1 score: {best_score}')","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"### Инференс","metadata":{}},{"cell_type":"code","source":"# Поскольку мы разделяем набор данных и выполняем итерацию отдельно для изображений и идентификаторов, порядок имеет значение.\ntest_ds = get_test_dataset(ordered=True) \n\nprint('Вычисляем предсказания...')\ntest_images_ds = test_ds.map(lambda image, idnum: image)\n\nprobabilities1 = model1.predict(test_images_ds)\nprobabilities2 = model2.predict(test_images_ds)\nprobabilities = best_alpha * probabilities1 + (1 - best_alpha) * probabilities2\npredictions = np.argmax(probabilities, axis=-1)\nprint(predictions)\n\nprint('Создание файла submission.csv...')\ntest_ids_ds = test_ds.map(lambda image, idnum: idnum).unbatch()\ntest_ids = next(iter(test_ids_ds.batch(NUM_TEST_IMAGES))).numpy().astype('U') # все в одной партии\nnp.savetxt('submission.csv', np.rec.fromarrays([test_ids, predictions]), fmt=['%s', '%d'], delimiter=',', header='id,label', comments='')","metadata":{"trusted":true},"outputs":[],"execution_count":null}]}