{"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":"code","source":"import tensorflow as tf\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.TPUStrategy(tpu)\nelse:\n    strategy = tf.distribute.get_strategy()\n\nAUTO = tf.data.experimental.AUTOTUNE\nprint(\"REPLICAS: \", strategy.num_replicas_in_sync)","metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"from kaggle_datasets import KaggleDatasets","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"!pip install -q efficientnet\n!pip install tensorflow_addons\nimport re\nimport os\nimport numpy as np\nimport pandas as pd\nimport random\nimport math\nimport tensorflow as tf\nimport efficientnet.tfkeras as efn\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\nfrom tqdm.auto import tqdm\nimport matplotlib.pyplot as plt\nimport pickle\nimport json\nimport tensorflow_hub as tfhub\nfrom datetime import datetime","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"save_dir = '.'\nEXPERIMENT = 0\nrun_ts = datetime.now().strftime('%Y%m%d-%H%M%S')\nprint(run_ts)","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"class config:\n     \n    SEED = 56\n    FOLD_TO_RUN = 20\n    FOLDS = 30\n    DEBUG = False\n    EVALUATE = True\n    RESUME = False\n    RESUME_EPOCH = None\n       \n    BATCH_SIZE = 16 * strategy.num_replicas_in_sync\n    IMAGE_SIZE = 768\n    N_CLASSES = 15587\n    model_type = 'effnetv1'  \n    EFF_NET = 7\n    EFF_NETV2 = 's-21k-ft1k'\n    FREEZE_BATCH_NORM = False\n    head = 'arcface' \n    EPOCHS = 31\n    LR = 0.001\n    message='baseline'\n    CUTOUT = False\n    save_dir = save_dir\n    KNN = 1000\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\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    \nseed_everything(56)    ","metadata":{},"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_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_count":null,"outputs":[]},{"cell_type":"code","source":"GCS_PATH1 = KaggleDatasets().get_gcs_path('happywhale-tfrecords-fullbody-768')\n    \ntrain_files = np.sort(np.array(tf.io.gfile.glob(GCS_PATH1 + '/train*.tfrec')))\ntest_files = np.sort(np.array(tf.io.gfile.glob(GCS_PATH1 + '/test*.tfrec')))\nprint(GCS_PATH1)\nprint(len(train_files),len(test_files),count_data_items(train_files),count_data_items(test_files))","metadata":{},"execution_count":null,"outputs":[]},{"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, matches):\n    return image,posting_id\n\ndef arcface_eval_format(posting_id, image, label_group, matches):\n    return image,label_group\n\ndef data_augment(posting_id, image, label_group, matches):\n    image = tf.image.random_flip_left_right(image, 6)\n    image = tf.image.random_hue(image, 0.01, 3)\n    image = tf.image.random_saturation(image, 0.65, 1.15)\n    image = tf.image.random_contrast(image, 0.7, 1.10, 4)\n    image = tf.image.random_brightness(image, 0.1, 5)\n    return posting_id, image, label_group, matches\n\ndef test_augment(posting_id, image, matches):\n    image = tf.image.random_flip_left_right(image, 6)\n    image = tf.image.random_hue(image, 0.01, 3)\n    image = tf.image.random_saturation(image, 0.65, 1.15)\n    image = tf.image.random_contrast(image, 0.7, 1.10, 4)\n    image = tf.image.random_brightness(image, 0.1, 5)\n    return posting_id, image, matches\n    \ndef decode_image(image_data):\n    image = tf.image.decode_jpeg(image_data, channels = 3)\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\ndef read_labeled_tfrecord(example):\n    LABELED_TFREC_FORMAT = {'image_id': tf.io.FixedLenFeature([], tf.string),\n                            'image': tf.io.FixedLenFeature([], tf.string),\n                            'individual_id': tf.io.FixedLenFeature([], tf.int64),}\n    example = tf.io.parse_single_example(example, LABELED_TFREC_FORMAT)\n    posting_id = example['image_id']\n    image = decode_image(example['image'])\n    label_group = tf.cast(example['individual_id'], tf.int32)\n    matches = 1\n    return posting_id, image, label_group, matches\n\ndef read_test_tfrecord(example):\n    LABELED_TFREC_FORMAT = {'image_id': tf.io.FixedLenFeature([], tf.string),\n                            'image': tf.io.FixedLenFeature([], tf.string),\n                            }\n    example = tf.io.parse_single_example(example, LABELED_TFREC_FORMAT)\n    posting_id = example['image_id']\n    image = decode_image(example['image'])\n    matches = 1\n    return posting_id, image, matches\n\ndef load_dataset(filenames, ordered = False, test = False):\n    ignore_order = tf.data.Options()\n    if not ordered:\n        ignore_order.experimental_deterministic = False \n    dataset = tf.data.TFRecordDataset(filenames, num_parallel_reads = AUTO)\n    dataset = dataset.with_options(ignore_order)\n    if (test == True):\n        dataset = dataset.map(read_test_tfrecord, num_parallel_calls = AUTO) \n    else:\n        dataset = dataset.map(read_labeled_tfrecord, num_parallel_calls = AUTO) \n    return dataset\n\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\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\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\ndef get_test_dataset(filenames, get_names = True):\n    dataset = load_dataset(filenames, ordered = True, test = True)\n    dataset = dataset.map(test_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_count":null,"outputs":[]},{"cell_type":"code","source":"class ArcMarginProduct(tf.keras.layers.Layer):\n    \n    def __init__(self, n_classes, s=30, m=0.50, easy_margin=False, ls_eps=0.0, **kwargs):\n        super(ArcMarginProduct, self).__init__(**kwargs)\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_classes': self.n_classes,'s': self.s, 'm': self.m,\n                       'ls_eps': self.ls_eps, 'easy_margin': self.easy_margin, })\n        return config\n\n    def build(self, input_shape):\n        super(ArcMarginProduct, self).build(input_shape[0])\n        self.W = self.add_weight(name='W', shape=(int(input_shape[0][-1]), self.n_classes),\n                                 initializer='glorot_uniform', dtype='float32',\n                                 trainable=True, regularizer=None)\n    def call(self, inputs):\n        X, y = inputs\n        y = tf.cast(y, dtype=tf.int32)\n        cosine = tf.matmul(tf.math.l2_normalize(X, axis=1), tf.math.l2_normalize(self.W, axis=0))\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(tf.one_hot(y, depth=self.n_classes), dtype=cosine.dtype)\n        if self.ls_eps > 0:\n            one_hot = (1 - self.ls_eps) * one_hot + self.ls_eps / self.n_classes\n        output = (one_hot * phi) + ((1.0 - one_hot) * cosine)\n        output *= self.s\n        return output\n\nEFNS = [efn.EfficientNetB0, efn.EfficientNetB1, efn.EfficientNetB2, efn.EfficientNetB3, \n        efn.EfficientNetB4, efn.EfficientNetB5, efn.EfficientNetB6, efn.EfficientNetB7]\n\ndef freeze_BN(model):\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\ndef get_model():\n    if config.head=='arcface':\n        head = ArcMarginProduct\n    else:\n        assert 1==2, \"INVALID HEAD\"\n    \n    with strategy.scope():\n        margin = head(n_classes = config.N_CLASSES, s = 30, m = 0.5,\n                      name=f'head/{config.head}', dtype='float32')\n\n        label = tf.keras.layers.Input(shape = (), name = 'inp2')\n        \n        if config.model_type == 'effnetv1':\n            inp = EFNS[config.EFF_NET](weights='noisy-student', include_top=False,\n                                       input_shape = [config.IMAGE_SIZE, config.IMAGE_SIZE, 3])\n            inp.layers[0]._name = 'inp1'\n            x1=tf.keras.layers.GlobalAveragePooling2D()(inp.layers[-1].output)\n            x2=tf.keras.layers.GlobalAveragePooling2D()(inp.layers[-5].output)\n            x3=tf.keras.layers.GlobalAveragePooling2D()(inp.layers[-7].output)\n            x4=tf.keras.layers.GlobalAveragePooling2D()(inp.layers[-13].output)\n            embed =  tf.concat([x1,x2,x3,x4],axis = 1)    \n            \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(0.3)(embed)\n        embed = tf.keras.layers.Dense(2048)(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.input, label], outputs = [output])\n        embed_model = tf.keras.models.Model(inputs = inp.input, 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(optimizer = opt, loss = [tf.keras.losses.SparseCategoricalCrossentropy()],\n                      metrics = [tf.keras.metrics.SparseCategoricalAccuracy(),\n                                 tf.keras.metrics.SparseTopKCategoricalAccuracy(k=5)]) \n        return model,embed_model","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def get_lr_callback(plot=False):\n    lr_start   = 0.000001\n    lr_max     = 0.000005 * config.BATCH_SIZE  \n    lr_min     = 0.000001\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        else:\n            lr = (lr_max - lr_min) * lr_decay**(epoch - lr_ramp_ep - lr_sus_ep) + lr_min\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_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    def on_epoch_end(self, epoch, logs=None):\n        if epoch in self.snapshot_epochs:     \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\")\n","metadata":{},"execution_count":null,"outputs":[]},{"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),\n      count_data_items(VALIDATION_FILENAMES))","metadata":{},"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)\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')\nK.clear_session()\nmodel,embed_model = get_model()\nsnap = Snapshot(fold=config.FOLD_TO_RUN,snapshot_epochs=[5,8])\n#model.summary()\n\nif config.RESUME:   \n    model.load_weights(config.resume_model_wts)","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"print('#### Image Size %i with EfficientNet B%i and batch_size %i'%\n      (config.IMAGE_SIZE,config.EFF_NET,config.BATCH_SIZE))\n\n#history = model.fit(train_dataset, validation_data = val_dataset, steps_per_epoch = STEPS_PER_EPOCH,\n #                   epochs = config.EPOCHS, callbacks = [snap,get_lr_callback(),train_logger,sv_loss], \n  #                  verbose = VERBOSE)","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"model.load_weights(\"../input/f30models/effnetv1_b7_lastF20.h5\")\n    #config.save_dir+f\"/{config.MODEL_NAME}_loss.h5\")","metadata":{},"execution_count":null,"outputs":[]},{"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    try:\n        return 1 / (predictions[:5].index(label) + 1)\n    except ValueError:\n        return 0.0\n    \ntrain_encoded = pd.read_csv('../input/happywhale-tfrecords-fullbody-768/train_encoded.csv')\n\ntarget_encodings = dict(zip(list(train_encoded['individual_id_encode'].values),\n                            list(train_encoded['individual_id'].values)))\n\nsample_list = ['938b7e931166', '5bf17305f073', '7593d2aee842', '7362d7a01d00','956562ff2888']","metadata":{},"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":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"train_targets = []\ntrain_embeddings = []\ntrain_data_list=[] \n\nfor filename in tqdm(TRAINING_FILENAMES):\n    embeddings = get_embeddings(filename)\n    with open(config.save_dir+f'train_{filename.split(\"/\")[-1]}_{config.FOLD_TO_RUN}.npy', 'wb') as f:\n        np.save(f, embeddings)    \n    targets = get_targets(filename)\n    train_embeddings.append(embeddings)\n    train_targets.append(targets)\n    train_data_list.append([filename,embeddings])\ntrain_embeddings_df = pd.DataFrame(train_data_list, columns=['filename', 'embeddings'])\ntrain_embeddings_df['FOLD_TO_RUN']=config.FOLD_TO_RUN\ntrain_embeddings_df.to_csv(config.save_dir+f\"/train_embeddings_{config.FOLD_TO_RUN}.csv\",index=False)\ntrain_embeddings = np.concatenate(train_embeddings)\ntrain_targets = np.concatenate(train_targets)","metadata":{},"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":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"test_ids = []\ntest_nn_distances = []\ntest_nn_idxs = []\nval_targets = []\nval_embeddings = []\nval_data_list=[] \nfor filename in tqdm(VALIDATION_FILENAMES[:]):\n    embeddings = get_embeddings(filename)\n    with open(config.save_dir+f'val_{filename.split(\"/\")[-1]}_{config.FOLD_TO_RUN}.npy', 'wb') as f:\n        np.save(f, embeddings) \n    val_data_list.append([filename,embeddings])\n    targets = get_targets(filename)\n    ids = get_ids(filename)\n    targets = targets[~np.isnan(embeddings).any(axis = 1)]\n    ids = ids[~np.isnan(embeddings).any(axis = 1)]\n    embeddings = embeddings[~np.isnan(embeddings).any(axis = 1)]\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)\n    \nval_embeddings_df = pd.DataFrame(val_data_list, columns=['filename', 'embeddings'])\nval_embeddings_df['FOLD_TO_RUN']=config.FOLD_TO_RUN\nval_embeddings_df.to_csv(config.save_dir+f\"/val_embeddings_{config.FOLD_TO_RUN}.csv\",index=False)    \n    \ntest_nn_distances = np.concatenate(test_nn_distances)\ntest_nn_idxs = np.concatenate(test_nn_idxs)\ntest_ids = np.concatenate(test_ids)\nval_embeddings = np.concatenate(val_embeddings)\nval_targets = np.concatenate(val_targets)","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":{},"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":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"best_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":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"print(\"Best threshold\",best_th)\nprint(\"Best cv\",best_cv)\nval_targets_df.describe()","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"val_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":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"train_embeddings = np.concatenate([train_embeddings,val_embeddings])\ntrain_targets = np.concatenate([train_targets,val_targets])\nprint(train_embeddings.shape,train_targets.shape)","metadata":{},"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":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"test_ids = []\ntest_nn_distances = []\ntest_nn_idxs = []\ntest_data_list=[] \nfor filename in tqdm(test_files):\n    embeddingsN = get_embeddings(filename)\n    embeddings = np.nan_to_num(embeddingsN)\n    with open(config.save_dir+f'test_{filename.split(\"/\")[-1]}_{config.FOLD_TO_RUN}.npy', 'wb') as f:\n        np.save(f, embeddings) \n    test_data_list.append([filename,embeddings])\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)\ntest_nn_distances = np.concatenate(test_nn_distances)\ntest_nn_idxs = np.concatenate(test_nn_idxs)\ntest_ids = np.concatenate(test_ids)","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"sample_submission = pd.read_csv('../input/happy-whale-and-dolphin/sample_submission.csv',index_col='image')\nprint(len(test_ids),len(sample_submission))\ntest_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('test_neighbors.csv')\ntest_df.image.value_counts().value_counts()","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"sample_list = ['938b7e931166', '5bf17305f073', '7593d2aee842', '7362d7a01d00','956562ff2888']","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"predictions = {}\nfor 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>best_threshold_adjusted:\n        predictions[row.image] = [row.target,'new_individual']\n    else:\n        predictions[row.image] = ['new_individual',row.target]\n        \nfor 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    predictions[x] = ' '.join(predictions[x])\n    \npredictions = pd.Series(predictions).reset_index()\npredictions.columns = ['image','predictions']\npredictions.to_csv('submission.csv',index=False)\npredictions.head()","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"","metadata":{},"execution_count":null,"outputs":[]}]}