{"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"pygments_lexer":"ipython3","nbconvert_exporter":"python","version":"3.6.4","file_extension":".py","codemirror_mode":{"name":"ipython","version":3},"name":"python","mimetype":"text/x-python"}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"markdown","source":"# Intro\n\nThe goal of this notebook is to extend upon the work of [MANOJ PRABHAKAR](https://www.kaggle.com/manojprabhaakr) in the [EFFNET B6 WHALE COMP](https://www.kaggle.com/manojprabhaakr/effnet-b6-whale-comp), utilising the [Detic bounding boxes](https://www.kaggle.com/c/happy-whale-and-dolphin/discussion/305503) to precrop the images.\n\nI have created a new set of TFRecords which includes the bounding boxes in [this](https://www.kaggle.com/lextoumbourou/happywhale-tfrecords-with-bounding-boxes) repo.\n\nAt the time of creation, this notebook achieves a top 100 score. Given there's still 2 months left in the competition and that this data is all publically available, I felt it to be reasonable to release.","metadata":{}},{"cell_type":"code","source":"!pip install -qq vit-keras\n!pip install -qq efficientnet\n!pip install -qq tensorflow_addons","metadata":{"execution":{"iopub.status.busy":"2022-02-26T14:45:55.004118Z","iopub.execute_input":"2022-02-26T14:45:55.004716Z","iopub.status.idle":"2022-02-26T14:46:22.399848Z","shell.execute_reply.started":"2022-02-26T14:45:55.004619Z","shell.execute_reply":"2022-02-26T14:46:22.398774Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"import tensorflow as tf\ntry:\n    # TPU detection. No parameters necessary if TPU_NAME environment variable is\n    # set: this is always the case on Kaggle.\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.TPUStrategy(tpu)\nelse:\n    # Default distribution strategy in Tensorflow. Works on CPU and single GPU.\n    strategy = tf.distribute.get_strategy()\n\nAUTO = tf.data.experimental.AUTOTUNE\nprint(\"REPLICAS: \", strategy.num_replicas_in_sync)","metadata":{"execution":{"iopub.status.busy":"2022-02-26T14:46:22.40204Z","iopub.execute_input":"2022-02-26T14:46:22.402475Z","iopub.status.idle":"2022-02-26T14:46:28.272065Z","shell.execute_reply.started":"2022-02-26T14:46:22.402437Z","shell.execute_reply":"2022-02-26T14:46:28.27111Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"from kaggle_datasets import KaggleDatasets","metadata":{"execution":{"iopub.status.busy":"2022-02-26T14:46:28.273569Z","iopub.execute_input":"2022-02-26T14:46:28.27478Z","iopub.status.idle":"2022-02-26T14:46:28.281746Z","shell.execute_reply.started":"2022-02-26T14:46:28.274732Z","shell.execute_reply":"2022-02-26T14:46:28.281013Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# basics\nimport re\nimport os\nimport numpy as np\nimport pandas as pd\nimport random\nimport math\n\n# tensorflow\nimport tensorflow as tf\nfrom sklearn import metrics\nfrom sklearn.model_selection import KFold, train_test_split\nfrom tensorflow.keras import backend as K\nimport tensorflow_addons as tfa\nimport tensorflow_hub as tfhub\n\n# tensorflow model\nimport efficientnet.tfkeras as efn\nfrom tensorflow.keras.models import Sequential\nfrom tensorflow.keras.layers import Dense, BatchNormalization, Flatten\nfrom tensorflow.keras import optimizers\nfrom vit_keras import vit, utils\n\n# utilities\nfrom tqdm.auto import tqdm\nimport matplotlib.pyplot as plt\nimport pickle\nimport json\nfrom datetime import datetime","metadata":{"execution":{"iopub.status.busy":"2022-02-26T14:46:28.28356Z","iopub.execute_input":"2022-02-26T14:46:28.284224Z","iopub.status.idle":"2022-02-26T14:46:30.242401Z","shell.execute_reply.started":"2022-02-26T14:46:28.284137Z","shell.execute_reply":"2022-02-26T14:46:30.241258Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"## Config","metadata":{}},{"cell_type":"code","source":"save_dir = '.'\n#EXPERIMENT = 0\n#EXPERIMENT = 1\n#EXPERIMENT = 2\nEXPERIMENT = 3\n#EXPERIMENT = 4\n\nrun_ts = datetime.now().strftime('%Y%m%d-%H%M%S')\nprint(run_ts)","metadata":{"execution":{"iopub.status.busy":"2022-02-26T14:46:30.245693Z","iopub.execute_input":"2022-02-26T14:46:30.246414Z","iopub.status.idle":"2022-02-26T14:46:30.252934Z","shell.execute_reply.started":"2022-02-26T14:46:30.246365Z","shell.execute_reply":"2022-02-26T14:46:30.251676Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"GCS_DS_PATH=KaggleDatasets().get_gcs_path('happywhale-tfrecords-bb')","metadata":{"execution":{"iopub.status.busy":"2022-02-26T14:46:30.254971Z","iopub.execute_input":"2022-02-26T14:46:30.255972Z","iopub.status.idle":"2022-02-26T14:46:30.837859Z","shell.execute_reply.started":"2022-02-26T14:46:30.255928Z","shell.execute_reply":"2022-02-26T14:46:30.836938Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"class config:\n    DEBUG = False\n    \n    SEED = 42\n    FOLD_TO_RUN = EXPERIMENT\n    FOLDS = 5\n    \n    EVALUATE = True\n    RESUME = False\n    RESUME_EPOCH = None\n    \n    \n    ### Dataset\n    BATCH_SIZE = 16 * strategy.num_replicas_in_sync\n    #IMAGE_SIZE = 512\n    IMAGE_SIZE = 384\n    N_CLASSES = 15587\n    MID_LAYER = 512 # output from ViT-model\n    \n    ### Model\n#     model_type = 'effnetv1'\n#     model_type = 'ViT_B_16'\n#     model_type = 'ViT_B_32'\n#     model_type = 'ViT_L_16'\n    model_type = 'ViT_L_32'\n    #EFF_NET = 6\n    EFF_NET = 7\n    EFF_NETV2 = 's-21k-ft1k'\n    FREEZE_BATCH_NORM = False\n    head = 'arcface' \n    #EPOCHS = 20\n    EPOCHS = 50\n#     LR = 0.001\n    LR = 5e-7\n    message='baseline'\n    \n    ### Augmentations\n    CUTOUT = False\n    \n    ### Save-Directory\n    save_dir = save_dir\n    \n    ### Inference\n    KNN = 100\n    \ndef count_data_items(filenames):\n    n = [int(re.compile(r\"-([0-9]*)\\.\").search(filename).group(1)) \n         for filename in filenames]\n    return np.sum(n)\n\n# Function to seed everything\ndef seed_everything(seed):\n    random.seed(seed)\n    np.random.seed(seed)\n    os.environ['PYTHONHASHSEED'] = str(seed)\n    tf.random.set_seed(seed)\n    \ndef is_interactive():\n    return 'runtime'    in get_ipython().config.IPKernelApp.connection_file\nIS_INTERACTIVE = is_interactive()\nprint(IS_INTERACTIVE)","metadata":{"execution":{"iopub.status.busy":"2022-02-26T14:46:30.839586Z","iopub.execute_input":"2022-02-26T14:46:30.840567Z","iopub.status.idle":"2022-02-26T14:46:30.854225Z","shell.execute_reply.started":"2022-02-26T14:46:30.840511Z","shell.execute_reply":"2022-02-26T14:46:30.853372Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"MODEL_NAME = None\nif config.model_type == 'effnetv1':\n    MODEL_NAME = f'effnetv1_b{config.EFF_NET}'\nelif config.model_type == 'effnetv2':\n    MODEL_NAME = f'effnetv2_{config.EFF_NETV2}'\n\nconfig.MODEL_NAME = MODEL_NAME\nprint(MODEL_NAME)","metadata":{"execution":{"iopub.status.busy":"2022-02-26T14:46:30.856074Z","iopub.execute_input":"2022-02-26T14:46:30.85642Z","iopub.status.idle":"2022-02-26T14:46:30.871727Z","shell.execute_reply.started":"2022-02-26T14:46:30.85638Z","shell.execute_reply":"2022-02-26T14:46:30.870674Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"with open(config.save_dir+'/config.json', 'w') as fp:\n    json.dump({x:dict(config.__dict__)[x] for x in dict(config.__dict__) if not x.startswith('_')}, fp)","metadata":{"execution":{"iopub.status.busy":"2022-02-26T14:46:30.873105Z","iopub.execute_input":"2022-02-26T14:46:30.873383Z","iopub.status.idle":"2022-02-26T14:46:30.883Z","shell.execute_reply.started":"2022-02-26T14:46:30.873354Z","shell.execute_reply":"2022-02-26T14:46:30.881959Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"train_files = np.sort(np.array(tf.io.gfile.glob(GCS_DS_PATH + '/happywhale-2022-train*.tfrec')))\ntest_files = np.sort(np.array(tf.io.gfile.glob(GCS_DS_PATH + '/happywhale-2022-test*.tfrec')))\nprint(GCS_DS_PATH)\nprint(len(train_files),len(test_files),count_data_items(train_files),count_data_items(test_files))","metadata":{"execution":{"iopub.status.busy":"2022-02-26T14:46:30.884047Z","iopub.execute_input":"2022-02-26T14:46:30.884619Z","iopub.status.idle":"2022-02-26T14:46:31.471015Z","shell.execute_reply.started":"2022-02-26T14:46:30.884586Z","shell.execute_reply":"2022-02-26T14:46:31.470086Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"## Data","metadata":{}},{"cell_type":"code","source":"def arcface_format(posting_id, image, label_group, matches):\n    return posting_id, {'inp1': image, 'inp2': label_group}, label_group, matches\n\ndef arcface_inference_format(posting_id, image, label_group, matches):\n    return image,posting_id\n\ndef arcface_eval_format(posting_id, image, label_group, matches):\n    return image,label_group\n\n# Data augmentation function\ndef data_augment(posting_id, image, label_group, matches):\n\n    ### CUTOUT\n    if tf.random.uniform([])>0.5 and config.CUTOUT:\n        N_CUTOUT = 6\n        for cutouts in range(N_CUTOUT):\n            if tf.random.uniform([])>0.5:\n                DIM = config.IMAGE_SIZE\n                CUTOUT_LENGTH = DIM//8\n                x1 = tf.cast( tf.random.uniform([],0,DIM-CUTOUT_LENGTH),tf.int32)\n                x2 = tf.cast( tf.random.uniform([],0,DIM-CUTOUT_LENGTH),tf.int32)\n                filter_ = tf.concat([tf.zeros((x1,CUTOUT_LENGTH)),tf.ones((CUTOUT_LENGTH,CUTOUT_LENGTH)),tf.zeros((DIM-x1-CUTOUT_LENGTH,CUTOUT_LENGTH))],axis=0)\n                filter_ = tf.concat([tf.zeros((DIM,x2)),filter_,tf.zeros((DIM,DIM-x2-CUTOUT_LENGTH))],axis=1)\n                cutout = tf.reshape(1-filter_,(DIM,DIM,1))\n                image = cutout*image\n\n    image = tf.image.random_flip_left_right(image)\n    # image = tf.image.random_flip_up_down(image)\n    image = tf.image.random_hue(image, 0.01)\n    image = tf.image.random_saturation(image, 0.70, 1.30)\n    image = tf.image.random_contrast(image, 0.80, 1.20)\n    image = tf.image.random_brightness(image, 0.10)\n    return posting_id, image, label_group, matches\n\n# Function to decode our images\n# Updated to include crops.\ndef decode_image(image_data, box):\n    if box is not None and box[0] != -1:\n        left, top, right, bottom = box[0], box[1], box[2], box[3]\n        bbs = tf.convert_to_tensor([top, left, bottom - top, right - left])\n        image = tf.io.decode_and_crop_jpeg(image_data, bbs, channels=3)\n    else:\n        image = tf.image.decode_jpeg(image_data, channels = 3)\n\n    image = tf.image.resize(image, [config.IMAGE_SIZE,config.IMAGE_SIZE])\n    image = tf.cast(image, tf.float32) / 255.0\n    return image\n\n# This function parse our images and also get the target variable\ndef read_labeled_tfrecord(example):\n    LABELED_TFREC_FORMAT = {\n        \"image_name\": tf.io.FixedLenFeature([], tf.string),\n        \"image\": tf.io.FixedLenFeature([], tf.string),\n        \"target\": tf.io.FixedLenFeature([], tf.int64),\n        'detic_box': tf.io.FixedLenFeature([4], tf.int64),\n         # 'yolov5_box': tf.io.FixedLenFeature([4], tf.int64),\n    }\n\n    example = tf.io.parse_single_example(example, LABELED_TFREC_FORMAT)\n    posting_id = example['image_name']\n    bb = tf.cast(example['detic_box'], tf.int32)\n    image = decode_image(example['image'], bb)\n#     label_group = tf.one_hot(tf.cast(example['label_group'], tf.int32), depth = N_CLASSES)\n    label_group = tf.cast(example['target'], tf.int32)\n#     matches = example['matches']\n    matches = 1\n    return posting_id, image, label_group, matches\n\n# This function loads TF Records and parse them into tensors\ndef load_dataset(filenames, ordered = False):\n    \n    ignore_order = tf.data.Options()\n    if not ordered:\n        ignore_order.experimental_deterministic = False \n        \n    dataset = tf.data.TFRecordDataset(filenames, num_parallel_reads = AUTO)\n#     dataset = dataset.cache()\n    dataset = dataset.with_options(ignore_order)\n    dataset = dataset.map(read_labeled_tfrecord, num_parallel_calls = AUTO) \n    return dataset\n\n# This function is to get our training tensors\ndef get_training_dataset(filenames):\n    dataset = load_dataset(filenames, ordered = False)\n    dataset = dataset.map(data_augment, num_parallel_calls = AUTO)\n    dataset = dataset.map(arcface_format, num_parallel_calls = AUTO)\n    dataset = dataset.map(lambda posting_id, image, label_group, matches: (image, label_group))\n    dataset = dataset.repeat()\n    dataset = dataset.shuffle(2048)\n    dataset = dataset.batch(config.BATCH_SIZE)\n    dataset = dataset.prefetch(AUTO)\n    return dataset\n\n# This function is to get our training tensors\ndef get_val_dataset(filenames):\n    dataset = load_dataset(filenames, ordered = True)\n    dataset = dataset.map(data_augment, num_parallel_calls = AUTO)\n    dataset = dataset.map(arcface_format, num_parallel_calls = AUTO)\n    dataset = dataset.map(lambda posting_id, image, label_group, matches: (image, label_group))\n    dataset = dataset.batch(config.BATCH_SIZE)\n    dataset = dataset.prefetch(AUTO)\n    return dataset\n\n# This function is to get our training tensors\ndef get_eval_dataset(filenames, get_targets = True):\n    dataset = load_dataset(filenames, ordered = True)\n    dataset = dataset.map(data_augment, num_parallel_calls = AUTO)\n    dataset = dataset.map(arcface_eval_format, num_parallel_calls = AUTO)\n    if not get_targets:\n        dataset = dataset.map(lambda image, target: image)\n    dataset = dataset.batch(config.BATCH_SIZE)\n    dataset = dataset.prefetch(AUTO)\n    return dataset\n\n# This function is to get our training tensors\ndef get_test_dataset(filenames, get_names = True):\n    dataset = load_dataset(filenames, ordered = True)\n    dataset = dataset.map(data_augment, num_parallel_calls = AUTO)\n    dataset = dataset.map(arcface_inference_format, num_parallel_calls = AUTO)\n    if not get_names:\n        dataset = dataset.map(lambda image, posting_id: image)\n    dataset = dataset.batch(config.BATCH_SIZE)\n    dataset = dataset.prefetch(AUTO)\n    return dataset","metadata":{"execution":{"iopub.status.busy":"2022-02-26T14:46:31.473021Z","iopub.execute_input":"2022-02-26T14:46:31.473791Z","iopub.status.idle":"2022-02-26T14:46:31.508231Z","shell.execute_reply.started":"2022-02-26T14:46:31.473747Z","shell.execute_reply":"2022-02-26T14:46:31.507449Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"row = 10; col = 8;\nrow = min(row,config.BATCH_SIZE//col)\nN_TRAIN = count_data_items(train_files)\nprint(N_TRAIN)\nds = get_training_dataset(train_files)\n\nfor (sample,label) in ds:\n    img = sample['inp1']\n    plt.figure(figsize=(25,int(25*row/col)))\n    for j in range(row*col):\n        plt.subplot(row,col,j+1)\n        plt.title(label[j].numpy())\n        plt.axis('off')\n        plt.imshow(img[j,])\n    plt.show()\n    break\nprint(img.shape)","metadata":{"execution":{"iopub.status.busy":"2022-02-26T14:46:31.509488Z","iopub.execute_input":"2022-02-26T14:46:31.510649Z","iopub.status.idle":"2022-02-26T14:47:00.958759Z","shell.execute_reply.started":"2022-02-26T14:46:31.510601Z","shell.execute_reply":"2022-02-26T14:47:00.957687Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"row = 10; col = 8;\nrow = min(row,config.BATCH_SIZE//col)\nN_TEST = count_data_items(test_files)\nprint(N_TEST)\nds = get_test_dataset(test_files)\n\nfor (img,label) in ds:\n    plt.figure(figsize=(25,int(25*row/col)))\n    for j in range(row*col):\n        plt.subplot(row,col,j+1)\n        plt.title(label[j].numpy())\n        plt.axis('off')\n        plt.imshow(img[j,])\n    plt.show()\n    break\nprint(img.shape)","metadata":{"execution":{"iopub.status.busy":"2022-02-26T14:47:00.960036Z","iopub.execute_input":"2022-02-26T14:47:00.960264Z","iopub.status.idle":"2022-02-26T14:47:04.532017Z","shell.execute_reply.started":"2022-02-26T14:47:00.960237Z","shell.execute_reply":"2022-02-26T14:47:04.530962Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"## Model","metadata":{}},{"cell_type":"code","source":"# Arcmarginproduct class keras layer\nclass ArcMarginProduct(tf.keras.layers.Layer):\n    '''\n    Implements large margin arc distance.\n\n    Reference:\n        https://arxiv.org/pdf/1801.07698.pdf\n        https://github.com/lyakaap/Landmark2019-1st-and-3rd-Place-Solution/\n            blob/master/src/modeling/metric_learning.py\n    '''\n#     def __init__(self, n_classes, s=30, m=0.50, easy_margin=False,\n#                  ls_eps=0.0, **kwargs):\n    def __init__(self, n_classes, s=30, m=0.10, easy_margin=False,\n                 ls_eps=0.0, **kwargs):\n        super(ArcMarginProduct, self).__init__(**kwargs)\n\n        self.n_classes = n_classes\n        self.s = s\n        self.m = m\n        self.ls_eps = ls_eps\n        self.easy_margin = easy_margin\n        self.cos_m = tf.math.cos(m)\n        self.sin_m = tf.math.sin(m)\n        self.th = tf.math.cos(math.pi - m)\n        self.mm = tf.math.sin(math.pi - m) * m\n\n    def get_config(self):\n\n        config = super().get_config().copy()\n        config.update({\n            'n_classes': self.n_classes,\n            's': self.s,\n            'm': self.m,\n            'ls_eps': self.ls_eps,\n            'easy_margin': self.easy_margin,\n        })\n        return config\n\n    def build(self, input_shape):\n        super(ArcMarginProduct, self).build(input_shape[0])\n\n        self.W = self.add_weight(\n            name='W',\n            shape=(int(input_shape[0][-1]), self.n_classes),\n            initializer='glorot_uniform',\n            dtype='float32',\n            trainable=True,\n            regularizer=None)\n\n    def call(self, inputs):\n        X, y = inputs\n        y = tf.cast(y, dtype=tf.int32)\n        cosine = tf.matmul(\n            tf.math.l2_normalize(X, axis=1),\n            tf.math.l2_normalize(self.W, axis=0)\n        )\n        sine = tf.math.sqrt(1.0 - tf.math.pow(cosine, 2))\n        phi = cosine * self.cos_m - sine * self.sin_m\n        if self.easy_margin:\n            phi = tf.where(cosine > 0, phi, cosine)\n        else:\n            phi = tf.where(cosine > self.th, phi, cosine - self.mm)\n        one_hot = tf.cast(\n            tf.one_hot(y, depth=self.n_classes),\n            dtype=cosine.dtype\n        )\n        if self.ls_eps > 0:\n            one_hot = (1 - self.ls_eps) * one_hot + self.ls_eps / self.n_classes\n\n        output = (one_hot * phi) + ((1.0 - one_hot) * cosine)\n        output *= self.s\n        return output","metadata":{"execution":{"iopub.status.busy":"2022-02-26T14:47:04.537488Z","iopub.execute_input":"2022-02-26T14:47:04.537746Z","iopub.status.idle":"2022-02-26T14:47:04.556203Z","shell.execute_reply.started":"2022-02-26T14:47:04.537719Z","shell.execute_reply":"2022-02-26T14:47:04.555011Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# EFNS = [efn.EfficientNetB0, efn.EfficientNetB1, efn.EfficientNetB2, efn.EfficientNetB3, \n#         efn.EfficientNetB4, efn.EfficientNetB5, efn.EfficientNetB6, efn.EfficientNetB7]\n\n# def freeze_BN(model):\n#     # Unfreeze layers while leaving BatchNorm layers frozen\n#     for layer in model.layers:\n#         if not isinstance(layer, tf.keras.layers.BatchNormalization):\n#             layer.trainable = True\n#         else:\n#             layer.trainable = False\n\n# # Function to create our EfficientNetB3 model\n# def get_model():\n\n#     if config.head=='arcface':\n#         head = ArcMarginProduct\n#     else:\n#         assert 1==2, \"INVALID HEAD\"\n    \n#     with strategy.scope():\n        \n#         margin = head(\n#             n_classes = config.N_CLASSES, \n#             s = 30, \n#             m = 0.3, \n#             name=f'head/{config.head}', \n#             dtype='float32'\n#             )\n\n#         inp = tf.keras.layers.Input(shape = [config.IMAGE_SIZE, config.IMAGE_SIZE, 3], name = 'inp1')\n#         label = tf.keras.layers.Input(shape = (), name = 'inp2')\n        \n#         if config.model_type == 'effnetv1':\n#             x = EFNS[config.EFF_NET](weights = 'noisy-student', include_top = False)(inp)\n#             embed = tf.keras.layers.GlobalAveragePooling2D()(x)\n#         elif config.model_type == 'effnetv2':\n#             FEATURE_VECTOR = f'{EFFNETV2_ROOT}/tfhub_models/efficientnetv2-{config.EFF_NETV2}/feature_vector'\n#             embed = tfhub.KerasLayer(FEATURE_VECTOR, trainable=True)(inp)\n            \n#         embed = tf.keras.layers.Dropout(0.2)(embed)\n#         embed = tf.keras.layers.Dense(512)(embed)\n#         x = margin([embed, label])\n        \n#         output = tf.keras.layers.Softmax(dtype='float32')(x)\n        \n#         model = tf.keras.models.Model(inputs = [inp, label], outputs = [output])\n#         embed_model = tf.keras.models.Model(inputs = inp, outputs = embed)  \n        \n#         opt = tf.keras.optimizers.Adam(learning_rate = config.LR)\n#         if config.FREEZE_BATCH_NORM:\n#             freeze_BN(model)\n\n#         model.compile(\n#             optimizer = opt,\n#             loss = [tf.keras.losses.SparseCategoricalCrossentropy()],\n#             metrics = [tf.keras.metrics.SparseCategoricalAccuracy(),tf.keras.metrics.SparseTopKCategoricalAccuracy(k=5)]\n#             ) \n        \n#         return model,embed_model","metadata":{"execution":{"iopub.status.busy":"2022-02-26T14:47:04.558183Z","iopub.execute_input":"2022-02-26T14:47:04.5585Z","iopub.status.idle":"2022-02-26T14:47:04.574942Z","shell.execute_reply.started":"2022-02-26T14:47:04.558471Z","shell.execute_reply":"2022-02-26T14:47:04.574117Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# available ViT models:ViT-B_16,ViT-B_32,ViT-L_16,ViT-L_32,\n\ndef freeze_BN(model):\n    # Unfreeze layers while leaving BatchNorm layers frozen\n    for layer in model.layers:\n        if not isinstance(layer, tf.keras.layers.BatchNormalization):\n            layer.trainable = True\n        else:\n            layer.trainable = False\n\n# Function to create our EfficientNetB3 model\ndef get_model():\n\n    if config.head=='arcface':\n        head = ArcMarginProduct\n    else:\n        assert 1==2, \"INVALID HEAD\"\n    \n    with strategy.scope():\n        \n        margin = head(\n            n_classes = config.N_CLASSES, \n            s = 30, \n            m = 0.3, \n            name=f'head/{config.head}', \n            dtype='float32'\n            )\n        if config.model_type == 'ViT_B_16':\n            vit_model = vit.vit_b16(image_size = config.IMAGE_SIZE,\n                                    activation = 'softmax',\n                                    pretrained = True,\n                                    include_top = False,\n                                    pretrained_top = False,\n                                    classes = config.MID_LAYER)\n        elif config.model_type == 'ViT_B_32':\n            vit_model = vit.vit_b32(image_size = config.IMAGE_SIZE,\n                                    activation = 'softmax',\n                                    pretrained = True,\n                                    include_top = False,\n                                    pretrained_top = False,\n                                    classes = config.MID_LAYER)\n        elif config.model_type == 'ViT_L_16':\n            vit_model = vit.vit_l16(image_size = config.IMAGE_SIZE,\n                                    activation = 'softmax',\n                                    pretrained = True,\n                                    include_top = False,\n                                    pretrained_top = False,\n                                    classes = config.MID_LAYER)\n        elif config.model_type == 'ViT_L_32':\n            vit_model = vit.vit_l32(image_size = config.IMAGE_SIZE,\n                                    activation = 'softmax',\n                                    pretrained = True,\n                                    include_top = False,\n                                    pretrained_top = False,\n                                    classes = config.MID_LAYER)\n        else:\n            vit_model = vit.vit_b16(image_size = config.IMAGE_SIZE,\n                                    activation = 'softmax',\n                                    pretrained = True,\n                                    include_top = False,\n                                    pretrained_top = False,\n                                    classes = config.MID_LAYER)\n        \n        inp = tf.keras.layers.Input(shape = [config.IMAGE_SIZE, config.IMAGE_SIZE, 3], name = 'inp1')\n        label = tf.keras.layers.Input(shape = (), name = 'inp2')\n        \n#         if config.model_type == 'effnetv1':\n#             x = EFNS[config.EFF_NET](weights = 'noisy-student', include_top = False)(inp)\n#             embed = tf.keras.layers.GlobalAveragePooling2D()(x)\n#         elif config.model_type == 'effnetv2':\n#             FEATURE_VECTOR = f'{EFFNETV2_ROOT}/tfhub_models/efficientnetv2-{config.EFF_NETV2}/feature_vector'\n#             embed = tfhub.KerasLayer(FEATURE_VECTOR, trainable=True)(inp)\n\n        embed = vit_model(inp)\n#         embed = tf.keras.layers.GlobalAveragePooling2D()(x)\n            \n        embed = tf.keras.layers.Dropout(0.2)(embed)\n        embed = tf.keras.layers.Dense(512)(embed)\n        x = margin([embed, label])\n        \n        output = tf.keras.layers.Softmax(dtype='float32')(x)\n        \n        model = tf.keras.models.Model(inputs = [inp, label], outputs = [output])\n        embed_model = tf.keras.models.Model(inputs = inp, outputs = embed)  \n        \n        opt = tf.keras.optimizers.Adam(learning_rate = config.LR)\n        if config.FREEZE_BATCH_NORM:\n            freeze_BN(model)\n\n        model.compile(\n            optimizer = opt,\n            loss = [tf.keras.losses.SparseCategoricalCrossentropy()],\n            metrics = [tf.keras.metrics.SparseCategoricalAccuracy(),tf.keras.metrics.SparseTopKCategoricalAccuracy(k=5)]\n            ) \n        \n        return model,embed_model","metadata":{"execution":{"iopub.status.busy":"2022-02-26T14:47:04.576267Z","iopub.execute_input":"2022-02-26T14:47:04.577159Z","iopub.status.idle":"2022-02-26T14:47:04.599205Z","shell.execute_reply.started":"2022-02-26T14:47:04.577103Z","shell.execute_reply":"2022-02-26T14:47:04.598451Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def get_lr_callback(plot=False):\n#     lr_start   = 0.000001\n    lr_start   = config.LR\n    lr_max     = 5 * config.LR * config.BATCH_SIZE  \n    lr_min     = config.LR\n    lr_ramp_ep = 4\n    lr_sus_ep  = 0\n    lr_decay   = 0.9\n   \n    def lrfn(epoch):\n        if config.RESUME:\n            epoch = epoch + config.RESUME_EPOCH\n        if epoch < lr_ramp_ep:\n            lr = (lr_max - lr_start) / lr_ramp_ep * epoch + lr_start\n            \n        elif epoch < lr_ramp_ep + lr_sus_ep:\n            lr = lr_max\n            \n        else:\n            lr = (lr_max - lr_min) * lr_decay**(epoch - lr_ramp_ep - lr_sus_ep) + lr_min\n            \n        return lr\n        \n    if plot:\n        epochs = list(range(config.EPOCHS))\n        learning_rates = [lrfn(x) for x in epochs]\n        plt.scatter(epochs,learning_rates)\n        plt.show()\n\n    lr_callback = tf.keras.callbacks.LearningRateScheduler(lrfn, verbose=False)\n    return lr_callback\n\nget_lr_callback(plot=True)","metadata":{"execution":{"iopub.status.busy":"2022-02-26T14:47:04.601714Z","iopub.execute_input":"2022-02-26T14:47:04.602052Z","iopub.status.idle":"2022-02-26T14:47:04.835971Z","shell.execute_reply.started":"2022-02-26T14:47:04.60201Z","shell.execute_reply":"2022-02-26T14:47:04.834965Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"class Snapshot(tf.keras.callbacks.Callback):\n    \n    def __init__(self,fold,snapshot_epochs=[]):\n        super(Snapshot, self).__init__()\n        self.snapshot_epochs = snapshot_epochs\n        self.fold = fold\n        \n        \n    def on_epoch_end(self, epoch, logs=None):\n        # logs is a dictionary\n#         print(f\"epoch: {epoch}, train_acc: {logs['acc']}, valid_acc: {logs['val_acc']}\")\n        if epoch in self.snapshot_epochs: # your custom condition         \n            self.model.save_weights(config.save_dir+f\"/EF{config.MODEL_NAME}_epoch{epoch}.h5\")\n        self.model.save_weights(config.save_dir+f\"/{config.MODEL_NAME}_last.h5\")","metadata":{"execution":{"iopub.status.busy":"2022-02-26T14:47:04.837332Z","iopub.execute_input":"2022-02-26T14:47:04.837583Z","iopub.status.idle":"2022-02-26T14:47:04.845536Z","shell.execute_reply.started":"2022-02-26T14:47:04.837554Z","shell.execute_reply":"2022-02-26T14:47:04.844461Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"## Train","metadata":{}},{"cell_type":"code","source":"TRAINING_FILENAMES = [x for i,x in enumerate(train_files) if i%config.FOLDS!=config.FOLD_TO_RUN]\nVALIDATION_FILENAMES = [x for i,x in enumerate(train_files) if i%config.FOLDS==config.FOLD_TO_RUN]\nprint(len(TRAINING_FILENAMES),len(VALIDATION_FILENAMES),count_data_items(TRAINING_FILENAMES),count_data_items(VALIDATION_FILENAMES))","metadata":{"execution":{"iopub.status.busy":"2022-02-26T14:47:04.847085Z","iopub.execute_input":"2022-02-26T14:47:04.847379Z","iopub.status.idle":"2022-02-26T14:47:04.858121Z","shell.execute_reply.started":"2022-02-26T14:47:04.847343Z","shell.execute_reply":"2022-02-26T14:47:04.857495Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"if config.DEBUG:\n    TRAINING_FILENAMES = [TRAINING_FILENAMES[0]]\n    VALIDATION_FILENAMES = [VALIDATION_FILENAMES[0]]\n    print(len(TRAINING_FILENAMES),len(VALIDATION_FILENAMES),count_data_items(TRAINING_FILENAMES),count_data_items(VALIDATION_FILENAMES))\n    test_files = [test_files[0]]","metadata":{"execution":{"iopub.status.busy":"2022-02-26T14:47:04.859615Z","iopub.execute_input":"2022-02-26T14:47:04.859834Z","iopub.status.idle":"2022-02-26T14:47:04.871093Z","shell.execute_reply.started":"2022-02-26T14:47:04.859808Z","shell.execute_reply":"2022-02-26T14:47:04.86992Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"seed_everything(config.SEED)\nVERBOSE = 1\ntrain_dataset = get_training_dataset(TRAINING_FILENAMES)\nval_dataset = get_val_dataset(VALIDATION_FILENAMES)\nSTEPS_PER_EPOCH = count_data_items(TRAINING_FILENAMES) // config.BATCH_SIZE\ntrain_logger = tf.keras.callbacks.CSVLogger(config.save_dir+'/training-log-fold-%i.h5.csv'%config.FOLD_TO_RUN)\n# SAVE BEST MODEL EACH FOLD        \nsv_loss = tf.keras.callbacks.ModelCheckpoint(\n    config.save_dir+f\"/{config.MODEL_NAME}_loss.h5\", monitor='val_loss', verbose=0, save_best_only=True,\n    save_weights_only=True, mode='min', save_freq='epoch')\n# BUILD MODEL\nK.clear_session()\nmodel,embed_model = get_model()\nsnap = Snapshot(fold=config.FOLD_TO_RUN,snapshot_epochs=[5,8])\nmodel.summary()\n\nif config.RESUME:   \n    model.load_weights(config.resume_model_wts)","metadata":{"execution":{"iopub.status.busy":"2022-02-26T14:47:04.872739Z","iopub.execute_input":"2022-02-26T14:47:04.873698Z","iopub.status.idle":"2022-02-26T14:48:54.088207Z","shell.execute_reply.started":"2022-02-26T14:47:04.873663Z","shell.execute_reply":"2022-02-26T14:48:54.087449Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"print(f\"Image Size {config.IMAGE_SIZE} with {config.model_type} and batch_size {config.BATCH_SIZE}\")\n\nhistory = model.fit(train_dataset,\n                    validation_data = val_dataset,\n                    steps_per_epoch = STEPS_PER_EPOCH,\n                    epochs = config.EPOCHS,\n                    callbacks = [snap,get_lr_callback(),train_logger,sv_loss], \n                    verbose = VERBOSE)","metadata":{"execution":{"iopub.status.busy":"2022-02-26T14:48:54.089909Z","iopub.execute_input":"2022-02-26T14:48:54.091002Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"model.load_weights(config.save_dir+f\"/{config.MODEL_NAME}_loss.h5\")","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"## Evaluation","metadata":{}},{"cell_type":"code","source":"def get_ids(filename):\n    ds = get_test_dataset([filename],get_names=True).map(lambda image, image_name: image_name).unbatch()\n    NUM_IMAGES = count_data_items([filename])\n    ids = next(iter(ds.batch(NUM_IMAGES))).numpy().astype('U')\n    return ids\n\ndef get_targets(filename):\n    ds = get_eval_dataset([filename],get_targets=True).map(lambda image, target: target).unbatch()\n    NUM_IMAGES = count_data_items([filename])\n    ids = next(iter(ds.batch(NUM_IMAGES))).numpy()\n    return ids\n\ndef get_embeddings(filename):\n    ds = get_test_dataset([filename],get_names=False)\n    embeddings = embed_model.predict(ds,verbose=0)\n    return embeddings\n\ndef get_predictions(test_df,threshold=0.2):\n    predictions = {}\n    for i,row in tqdm(test_df.iterrows()):\n        if row.image in predictions:\n            if len(predictions[row.image])==5:\n                continue\n            predictions[row.image].append(row.target)\n        elif row.confidence>threshold:\n            predictions[row.image] = [row.target,'new_individual']\n        else:\n            predictions[row.image] = ['new_individual',row.target]\n\n    for x in tqdm(predictions):\n        if len(predictions[x])<5:\n            remaining = [y for y in sample_list if y not in predictions]\n            predictions[x] = predictions[x]+remaining\n            predictions[x] = predictions[x][:5]\n        \n    return predictions\n\ndef map_per_image(label, predictions):\n    \"\"\"Computes the precision score of one image.\n\n    Parameters\n    ----------\n    label : string\n            The true label of the image\n    predictions : list\n            A list of predicted elements (order does matter, 5 predictions allowed per image)\n\n    Returns\n    -------\n    score : double\n    \"\"\"    \n    try:\n        return 1 / (predictions[:5].index(label) + 1)\n    except ValueError:\n        return 0.0\n    \nf = open ('../input/happywhale-splits/individual_ids.json', \"r\")\ntarget_encodings = json.loads(f.read())\ntarget_encodings = {target_encodings[x]:x for x in target_encodings}\nsample_list = ['938b7e931166', '5bf17305f073', '7593d2aee842', '7362d7a01d00','956562ff2888']","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"train_targets = []\ntrain_embeddings = []\nfor filename in tqdm(TRAINING_FILENAMES):\n    embeddings = get_embeddings(filename)\n    targets = get_targets(filename)\n    train_embeddings.append(embeddings)\n    train_targets.append(targets)\ntrain_embeddings = np.concatenate(train_embeddings)\ntrain_targets = np.concatenate(train_targets)","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"from sklearn.neighbors import NearestNeighbors\nneigh = NearestNeighbors(n_neighbors=config.KNN,metric='cosine')\nneigh.fit(train_embeddings)","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"test_ids = []\ntest_nn_distances = []\ntest_nn_idxs = []\nval_targets = []\nval_embeddings = []\n\nfor filename in tqdm(VALIDATION_FILENAMES):\n    embeddings = get_embeddings(filename)\n    targets = get_targets(filename)\n    ids = get_ids(filename)\n    distances,idxs = neigh.kneighbors(embeddings, config.KNN, return_distance=True)\n    test_ids.append(ids)\n    test_nn_idxs.append(idxs)\n    test_nn_distances.append(distances)\n#     val_embeddings.append(embeddings)\n    val_targets.append(targets)\ntest_nn_distances = np.concatenate(test_nn_distances)\ntest_nn_idxs = np.concatenate(test_nn_idxs)\ntest_ids = np.concatenate(test_ids)\n# val_embeddings = np.concatenate(val_embeddings)\nval_targets = np.concatenate(val_targets)","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"try:\n    np.save(\"test_nn_distances\",test_nn_distances)\n    np.save(\"test_nn_idxs\",test_nn_idxs)\n    np.save(\"test_ids\",test_ids)\n    np.save(\"val_targets\",val_targets)\nexcept Exception as e:\n    print('Error:'+ str(e))","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"try:\n    test_nn_distances_df = pd.Series(test_nn_distances).reset_index()\n    test_nn_distances_df.to_csv(\"pred_distances.csv\", index=False)\n    \n    test_nn_idxs_df = pd.Series(test_nn_idxs).reset_index()\n    test_nn_idxs_df.to_csv(\"pred_IDindexs.csv\", index=False)\n    \n    test_ids_df = pd.Series(test_ids).reset_index()\n    test_ids_df.to_csv(\"pred_IDstr.csv\", index=False)\n    \n    val_embeddings_df = pd.Series(val_embeddings).reset_index()\n    test_ids_df.to_csv(\"pred_emdeddings.csv\", index=False)\n    \n    val_targets_df = pd.Series(val_targets).reset_index()\n    test_ids_df.to_csv(\"val_targets.csv\", index=False)\n    \nexcept Exception as e:\n    print('Error:'+ str(e))","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"try:\n    idxs_dict = {}\n\n    for idx, id_idx in tqdm(enumerate(test_nn_idxs)):\n        idxs_dict[idx] = id_idx[:10]\n\n    predict_df = pd.Series(idxs_dict).reset_index()\n    predict_df.columns = [\"INDEX\", \"pred_id_index\"]\n    predict_df.to_csv(\"pred_id_index.csv\",index=False)\nexcept Exception as e:\n    print('Error:'+ str(e))","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"try:\n    distance_dict = {}\n\n    for idx, distance in tqdm(enumerate(distances)):\n        distance_dict[idx] = 1 - distance[:10]\n\n    predict_conf_df = pd.Series(distance_dict).reset_index()\n    predict_conf_df.columns = [\"INDEX\", \"confidence\"]\n    predict_conf_df.to_csv(\"pred_id_conf.csv\",index=False)\nexcept Exception as e:\n    print('Error:'+ str(e))","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"try:\n    target_dict = {}\n\n    for idx, target in tqdm(enumerate(val_targets)):\n        target_dict[idx] = target\n\n    target_df = pd.Series(target_dict).reset_index()\n    target_df.columns = [\"INDEX\", \"target\"]\n    target_df.to_csv(\"targets.csv\",index=False)\nexcept Exception as e:\n    print('Error:'+ str(e))","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"try:\n    id_dict = {}\n\n    for i in tqdm(range(len(test_ids))):\n        id_dict[i] = test_ids[i]\n\n    id_dict_df = pd.Series(id_dict).reset_index()\n    id_dict_df.columns = [\"INDEX\", \"id\"]\n    id_dict_df.to_csv(\"testids.csv\",index=False)\nexcept Exception as e:\n    print('Error:'+ str(e))","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"try:\n    predict_idconf_df = pd.merge(predict_df, predict_conf_df, on=\"INDEX\")\n    predict_idconf_df.to_csv(\"pred.csv\",index=False)\nexcept Exception as e:\n    print('Error:'+ str(e))","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"try:\n    predict_idconf_df = pd.merge(predict_idconf_df, target_df, on=\"INDEX\")\n    predict_idconf_df.to_csv(\"pred_and_target.csv\",index=False)\nexcept Exception as e:\n    print('Error:'+ str(e))","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"try:\n    id_dict_df = pd.merge(id_dict_df, predict_idconf_df, on=\"INDEX\")\n    id_dict_df.to_csv(\"pred_and_target_and_id.csv\",index=False)\nexcept Exception as e:\n    print('Error:'+ str(e))","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"allowed_targets = set([target_encodings[x] for x in np.unique(train_targets)])\nval_targets_df = pd.DataFrame(np.stack([test_ids,val_targets],axis=1),columns=['image','target'])\nval_targets_df['target'] = val_targets_df['target'].astype(int).map(target_encodings)\nval_targets_df.loc[~val_targets_df.target.isin(allowed_targets),'target'] = 'new_individual'\nval_targets_df.target.value_counts()","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"test_df = []\nfor i in tqdm(range(len(test_ids))):\n    id_ = test_ids[i]\n    targets = train_targets[test_nn_idxs[i]]\n    distances = test_nn_distances[i]\n    subset_preds = pd.DataFrame(np.stack([targets,distances],axis=1),columns=['target','distances'])\n    subset_preds['image'] = id_\n    test_df.append(subset_preds)\ntest_df = pd.concat(test_df).reset_index(drop=True)\ntest_df['confidence'] = 1-test_df['distances']\ntest_df = test_df.groupby(['image','target']).confidence.max().reset_index()\ntest_df = test_df.sort_values('confidence',ascending=False).reset_index(drop=True)\ntest_df['target'] = test_df['target'].map(target_encodings)\ntest_df.to_csv('val_neighbors.csv')\ntest_df.image.value_counts().value_counts()","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"## Compute CV\nbest_th = 0\nbest_cv = 0\nfor th in [0.1*x for x in range(11)]:\n    all_preds = get_predictions(test_df,threshold=th)\n    cv = 0\n    for i,row in val_targets_df.iterrows():\n        target = row.target\n        preds = all_preds[row.image]\n        val_targets_df.loc[i,th] = map_per_image(target,preds)\n    cv = val_targets_df[th].mean()\n    print(f\"CV at threshold {th}: {cv}\")\n    if cv > best_cv:\n        best_th = th\n        best_cv = cv","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"print(\"Best threshold\",best_th)\nprint(\"Best cv\",best_cv)\nval_targets_df.to_csv(\"val_targets_df.csv\", index=False)\nval_targets_df.describe()","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"## Adjustment: Since Public lb has nearly 10% 'new_individual' (Be Careful for private LB)\nval_targets_df['is_new_individual'] = val_targets_df.target=='new_individual'\nprint(val_targets_df.is_new_individual.value_counts().to_dict())\nval_scores = val_targets_df.groupby('is_new_individual').mean().T\nval_scores['adjusted_cv'] = val_scores[True]*0.1 + val_scores[False]*0.9\nbest_threshold_adjusted = val_scores['adjusted_cv'].idxmax()\nprint(\"best_threshold\",best_threshold_adjusted)\nval_scores","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"","metadata":{},"execution_count":null,"outputs":[]}]}