{"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"name":"python","version":"3.10.13","mimetype":"text/x-python","codemirror_mode":{"name":"ipython","version":3},"pygments_lexer":"ipython3","nbconvert_exporter":"python","file_extension":".py"},"kaggle":{"accelerator":"nvidiaTeslaT4","dataSources":[{"sourceId":70203,"databundleVersionId":8068726,"sourceType":"competition"},{"sourceId":6127,"sourceType":"modelInstanceVersion","modelInstanceId":4598}],"dockerImageVersionId":30733,"isInternetEnabled":true,"language":"python","sourceType":"notebook","isGpuEnabled":true}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"code","source":"import os\nos.environ[\"KERAS_BACKEND\"] = \"tensorflow\"\n\nimport keras_cv\nimport keras\nimport keras.backend as K\nimport tensorflow as tf\nimport tensorflow_io as tfio\n\nimport numpy as np \nimport pandas as pd\n\nfrom glob import glob\nfrom tqdm import tqdm\n\nimport librosa\nimport IPython.display as ipd\nimport librosa.display as lid\n\nimport matplotlib.pyplot as plt\nimport matplotlib as mpl\n\ncmap = mpl.cm.get_cmap('coolwarm')","metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","execution":{"iopub.status.busy":"2024-06-09T15:50:23.546899Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"if tf.config.list_physical_devices('GPU'):\n    print(\"GPU is available\")\nelse:\n    print(\"GPU is not available, using CPU\")","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"print(\"TensorFlow:\", tf.__version__)\nprint(\"Keras:\", keras.__version__)\nprint(\"KerasCV:\", keras_cv.__version__)","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"class CFG:\n    seed = 42\n    \n    # Input image size and batch size\n    img_size = [128, 384]\n    batch_size = 64\n    \n    # Audio duration, sample rate, and length\n    duration = 15 # second\n    sample_rate = 32000\n    audio_len = duration*sample_rate\n    \n    # STFT parameters\n    nfft = 2028\n    window = 2048\n    hop_length = audio_len // (img_size[1] - 1)\n    fmin = 20\n    fmax = 16000\n    \n    # Number of epochs, model name\n    epochs = 10\n    preset = 'efficientnetv2_b2_imagenet'\n    \n    # Data augmentation parameters\n    augment=True\n\n    # Class Labels for BirdCLEF 24\n    class_names = sorted(os.listdir('/kaggle/input/birdclef-2024/train_audio/'))\n    num_classes = len(class_names)\n    class_labels = list(range(num_classes))\n    label2name = dict(zip(class_labels, class_names))\n    name2label = {v:k for k,v in label2name.items()}","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"tf.keras.utils.set_random_seed(CFG.seed)","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"BASE_PATH = '/kaggle/input/birdclef-2024'","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"df = pd.read_csv(f'{BASE_PATH}/train_metadata.csv')\ndf['filepath'] = BASE_PATH + '/train_audio/' + df.filename\ndf['target'] = df.primary_label.map(CFG.name2label)\ndf['filename'] = df.filepath.map(lambda x: x.split('/')[-1])\ndf['xc_id'] = df.filepath.map(lambda x: x.split('/')[-1].split('.')[0])\n\n# Display rwos\ndf.head(2)","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def load_audio(filepath):\n    audio, sr = librosa.load(filepath)\n    return audio, sr\n\ndef get_spectrogram(audio):\n    spec = librosa.feature.melspectrogram(y=audio, \n                                   sr=CFG.sample_rate, \n                                   n_mels=256,\n                                   n_fft=2048,\n                                   hop_length=512,\n                                   fmax=CFG.fmax,\n                                   fmin=CFG.fmin,\n                                   )\n    spec = librosa.power_to_db(spec, ref=1.0)\n    min_ = spec.min()\n    max_ = spec.max()\n    if max_ != min_:\n        spec = (spec - min_)/(max_ - min_)\n    return spec\n\ndef display_audio(row):\n    # Caption for viz\n    caption = f'Id: {row.filename} | Name: {row.common_name} | Sci.Name: {row.scientific_name} | Rating: {row.rating}'\n    # Read audio file\n    audio, sr = load_audio(row.filepath)\n    # Keep fixed length audio\n    audio = audio[:CFG.audio_len]\n    # Spectrogram from audio\n    spec = get_spectrogram(audio)\n    # Display audio\n    print(\"# Audio:\")\n    display(ipd.Audio(audio, rate=CFG.sample_rate))\n    print('# Visualization:')\n    fig, ax = plt.subplots(2, 1, figsize=(12, 2*3), sharex=True, tight_layout=True)\n    fig.suptitle(caption)\n    # Waveplot\n    lid.waveshow(audio,\n                 sr=CFG.sample_rate,\n                 ax=ax[0],\n                 color= cmap(0.1))\n    # Specplot\n    lid.specshow(spec, \n                 sr = CFG.sample_rate, \n                 hop_length=512,\n                 n_fft=2048,\n                 fmin=CFG.fmin,\n                 fmax=CFG.fmax,\n                 x_axis = 'time', \n                 y_axis = 'mel',\n                 cmap = 'coolwarm',\n                 ax=ax[1])\n    ax[0].set_xlabel('');\n    fig.show()","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"row = df.iloc[35]\n\n# Display audio\ndisplay_audio(row)","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"row = df.iloc[16]\n\n# Display audio\ndisplay_audio(row)","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# Import required packages\nfrom sklearn.model_selection import train_test_split\n\ntrain_df, valid_df = train_test_split(df, test_size=0.2)\n\nprint(f\"Num Train: {len(train_df)} | Num Valid: {len(valid_df)}\")","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# Decodes Audio\ndef build_decoder(with_labels=True, dim=1024):\n    def get_audio(filepath):\n        file_bytes = tf.io.read_file(filepath)\n        audio = tfio.audio.decode_vorbis(file_bytes)  # decode .ogg file\n        audio = tf.cast(audio, tf.float32)\n        if tf.shape(audio)[1] > 1:  # stereo -> mono\n            audio = audio[..., 0:1]\n        audio = tf.squeeze(audio, axis=-1)\n        return audio\n\n    def crop_or_pad(audio, target_len, pad_mode=\"constant\"):\n        audio_len = tf.shape(audio)[0]\n        diff_len = abs(\n            target_len - audio_len\n        )  # find difference between target and audio length\n        if audio_len < target_len:  # do padding if audio length is shorter\n            pad1 = tf.random.uniform([], maxval=diff_len, dtype=tf.int32)\n            pad2 = diff_len - pad1\n            audio = tf.pad(audio, paddings=[[pad1, pad2]], mode=pad_mode)\n        elif audio_len > target_len:  # do cropping if audio length is larger\n            idx = tf.random.uniform([], maxval=diff_len, dtype=tf.int32)\n            audio = audio[idx : (idx + target_len)]\n        return tf.reshape(audio, [target_len])\n\n    def apply_preproc(spec):\n        # Standardize\n        mean = tf.math.reduce_mean(spec)\n        std = tf.math.reduce_std(spec)\n        spec = tf.where(tf.math.equal(std, 0), spec - mean, (spec - mean) / std)\n\n        # Normalize using Min-Max\n        min_val = tf.math.reduce_min(spec)\n        max_val = tf.math.reduce_max(spec)\n        spec = tf.where(\n            tf.math.equal(max_val - min_val, 0),\n            spec - min_val,\n            (spec - min_val) / (max_val - min_val),\n        )\n        return spec\n\n    def get_target(target):\n        target = tf.reshape(target, [1])\n        target = tf.cast(tf.one_hot(target, CFG.num_classes), tf.float32)\n        target = tf.reshape(target, [CFG.num_classes])\n        return target\n\n    def decode(path):\n        # Load audio file\n        audio = get_audio(path)\n        # Crop or pad audio to keep a fixed length\n        audio = crop_or_pad(audio, dim)\n        # Audio to Spectrogram\n        spec = keras.layers.MelSpectrogram(\n            num_mel_bins=CFG.img_size[0],\n            fft_length=CFG.nfft,\n            sequence_stride=CFG.hop_length,\n            sampling_rate=CFG.sample_rate,\n        )(audio)\n        # Apply normalization and standardization\n        spec = apply_preproc(spec)\n        # Spectrogram to 3 channel image (for imagenet)\n        spec = tf.tile(spec[..., None], [1, 1, 3])\n        spec = tf.reshape(spec, [*CFG.img_size, 3])\n        return spec\n\n    def decode_with_labels(path, label):\n        label = get_target(label)\n        return decode(path), label\n\n    return decode_with_labels if with_labels else decode\n","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def build_augmenter():\n    augmenters = [\n        keras_cv.layers.MixUp(alpha=0.4),\n        keras_cv.layers.RandomCutout(height_factor=(1.0, 1.0),\n                                     width_factor=(0.06, 0.12)), # time-masking\n        keras_cv.layers.RandomCutout(height_factor=(0.06, 0.1),\n                                     width_factor=(1.0, 1.0)), # freq-masking\n        keras_cv.layers.AugMix(severity=0.02, alpha=1.0, value_range=(0, 1)),\n        keras_cv.layers.ChannelShuffle(groups=3),\n        keras_cv.layers.GridMask(),\n        keras_cv.layers.RandomChannelShift(value_range=(0, 1), factor=0.5),\n        keras_cv.layers.RandomColorDegeneration(factor=0.2),\n        keras_cv.layers.RandomHue(value_range=(0, 1), factor=0.2),\n        keras_cv.layers.RandomSaturation(factor=0.2),\n        keras_cv.layers.RandomSharpness(value_range=(0, 1), factor=0.2)\n    ]\n    \n    def augment(img, label):\n        data = {\"images\":img, \"labels\":label}\n        for augmenter in augmenters:\n            if tf.random.uniform([]) < 0.35:\n                data = augmenter(data, training=True)\n        return data[\"images\"], data[\"labels\"]\n    \n    return augment","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def build_dataset(paths, labels=None, batch_size=32, \n                  decode_fn=None, augment_fn=None, cache=True,\n                  augment=False, shuffle=2048):\n\n    if decode_fn is None:\n        decode_fn = build_decoder(labels is not None, dim=CFG.audio_len)\n\n    if augment_fn is None:\n        augment_fn = build_augmenter()\n        \n    AUTO = tf.data.experimental.AUTOTUNE\n    slices = (paths,) if labels is None else (paths, labels)\n    ds = tf.data.Dataset.from_tensor_slices(slices)\n    ds = ds.map(decode_fn, num_parallel_calls=AUTO)\n    ds = ds.cache() if cache else ds\n    if shuffle:\n        opt = tf.data.Options()\n        ds = ds.shuffle(shuffle, seed=CFG.seed)\n        opt.experimental_deterministic = False\n        ds = ds.with_options(opt)\n    ds = ds.batch(batch_size, drop_remainder=True)\n    ds = ds.map(augment_fn, num_parallel_calls=AUTO) if augment else ds\n    ds = ds.prefetch(AUTO)\n    return ds","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# Train\ntrain_paths = train_df.filepath.values\ntrain_labels = train_df.target.values\ntrain_ds = build_dataset(train_paths, train_labels, batch_size=CFG.batch_size,\n                         shuffle=True, augment=CFG.augment)\n\n# Valid\nvalid_paths = valid_df.filepath.values\nvalid_labels = valid_df.target.values\nvalid_ds = build_dataset(valid_paths, valid_labels, batch_size=CFG.batch_size,\n                         shuffle=False, augment=False)","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"print(\"Devices available: \", tf.config.list_physical_devices())","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def plot_batch(batch, row=3, col=3, label2name=None,):\n    \"\"\"Plot one batch data\"\"\"\n    if isinstance(batch, tuple) or isinstance(batch, list):\n        specs, tars = batch\n    else:\n        specs = batch\n        tars = None\n    plt.figure(figsize=(col*5, row*3))\n    for idx in range(row*col):\n        ax = plt.subplot(row, col, idx+1)\n        lid.specshow(np.array(specs[idx, ..., 0]), \n                     n_fft=CFG.nfft, \n                     hop_length=CFG.hop_length, \n                     sr=CFG.sample_rate,\n                     x_axis='time',\n                     y_axis='mel',\n                     cmap='coolwarm')\n        if tars is not None:\n            label = tars[idx].numpy().argmax()\n            name = label2name[label]\n            plt.title(name)\n    plt.tight_layout()\n    plt.show()","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"sample_ds = train_ds.take(100)\nbatch = next(iter(sample_ds))\nplot_batch(batch, label2name=CFG.label2name)","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# Create an input layer for the model\ninp = keras.layers.Input(shape=(None, None, 3))\n# Pretrained backbone\nbackbone = keras_cv.models.EfficientNetV2Backbone.from_preset(\n    CFG.preset,\n)\nout = keras_cv.models.ImageClassifier(\n    backbone=backbone,\n    num_classes=CFG.num_classes,\n    name=\"classifier\"\n)(inp)\n# Build model\nmodel = keras.models.Model(inputs=inp, outputs=out)\n# Compile model with optimizer, loss and metrics\nmodel.compile(optimizer=\"adam\",\n              loss=keras.losses.CategoricalCrossentropy(label_smoothing=0.02),\n              metrics=[keras.metrics.AUC(name='auc')],\n             )\nmodel.summary()","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"import math\n\ndef get_lr_callback(batch_size=8, mode='cos', epochs=10, plot=False):\n    lr_start, lr_max, lr_min = 5e-5, 8e-6 * batch_size, 1e-5\n    lr_ramp_ep, lr_sus_ep, lr_decay = 3, 0, 0.75\n\n    def lrfn(epoch):  # Learning rate update function\n        if epoch < lr_ramp_ep: lr = (lr_max - lr_start) / lr_ramp_ep * epoch + lr_start\n        elif epoch < lr_ramp_ep + lr_sus_ep: lr = lr_max\n        elif mode == 'exp': lr = (lr_max - lr_min) * lr_decay**(epoch - lr_ramp_ep - lr_sus_ep) + lr_min\n        elif mode == 'step': lr = lr_max * lr_decay**((epoch - lr_ramp_ep - lr_sus_ep) // 2)\n        elif mode == 'cos':\n            decay_total_epochs, decay_epoch_index = epochs - lr_ramp_ep - lr_sus_ep + 3, epoch - lr_ramp_ep - lr_sus_ep\n            phase = math.pi * decay_epoch_index / decay_total_epochs\n            lr = (lr_max - lr_min) * 0.5 * (1 + math.cos(phase)) + lr_min\n        return lr\n\n    if plot:  # Plot lr curve if plot is True\n        plt.figure(figsize=(10, 5))\n        plt.plot(np.arange(epochs), [lrfn(epoch) for epoch in np.arange(epochs)], marker='o')\n        plt.xlabel('epoch'); plt.ylabel('lr')\n        plt.title('LR Scheduler')\n        plt.show()\n\n    return keras.callbacks.LearningRateScheduler(lrfn, verbose=False)  # Create lr callback","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"lr_cb = get_lr_callback(CFG.batch_size, plot=True)","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"ckpt_cb = keras.callbacks.ModelCheckpoint(\"/kaggle/working/models/best_model_v2.weights.h5\",\n                                         monitor='val_auc',\n                                         save_best_only=True,\n                                         save_weights_only=True,\n                                         mode='max')","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# !rm /kaggle/working/models/best_model_v2.weights.h5\n# !ls /kaggle/working/models/","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"history = model.fit(\n    train_ds, \n    validation_data=valid_ds, \n    epochs=CFG.epochs,\n    callbacks=[lr_cb, ckpt_cb], \n    verbose=1\n)","metadata":{"execution":{"iopub.status.idle":"2024-06-09T20:14:37.300953Z","shell.execute_reply.started":"2024-06-09T15:51:35.326943Z","shell.execute_reply":"2024-06-09T20:14:37.299999Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"best_epoch = np.argmax(history.history[\"val_auc\"])\nbest_score = history.history[\"val_auc\"][best_epoch]\nprint('>>> Best AUC: ', best_score)\nprint('>>> Best Epoch: ', best_epoch+1)","metadata":{"execution":{"iopub.status.busy":"2024-06-09T20:14:37.308371Z","iopub.execute_input":"2024-06-09T20:14:37.308679Z","iopub.status.idle":"2024-06-09T20:14:37.328630Z","shell.execute_reply.started":"2024-06-09T20:14:37.308655Z","shell.execute_reply":"2024-06-09T20:14:37.327719Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"!ls /kaggle/working/models/","metadata":{"execution":{"iopub.status.busy":"2024-06-09T20:14:37.329855Z","iopub.execute_input":"2024-06-09T20:14:37.330137Z","iopub.status.idle":"2024-06-09T20:14:39.528869Z","shell.execute_reply.started":"2024-06-09T20:14:37.330111Z","shell.execute_reply":"2024-06-09T20:14:39.527701Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# /kaggle/working/models/best_model_v2.weights.h5","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Inference\ninference was done in the same notebook due to not having access to the tarained models outside the training notebook","metadata":{}},{"cell_type":"code","source":"class TestCFG:\n    seed = 42\n    \n    # Input image size and batch size\n    img_size = [128, 384]\n    \n    # Audio duration, sample rate, and length\n    duration = 15 # second\n    sample_rate = 32000\n    audio_len = duration*sample_rate\n    \n    # STFT parameters\n    nfft = 2028\n    window = 2048\n    hop_length = audio_len // (img_size[1] - 1)\n    fmin = 20\n    fmax = 16000\n    \n    # Number of epochs, model name\n    preset = 'efficientnetv2_b2_imagenet'\n\n    # Class Labels for BirdCLEF 24\n    class_names = sorted(os.listdir('/kaggle/input/birdclef-2024/train_audio/'))\n    num_classes = len(class_names)\n    class_labels = list(range(num_classes))\n    label2name = dict(zip(class_labels, class_names))\n    name2label = {v:k for k,v in label2name.items()}","metadata":{"execution":{"iopub.status.busy":"2024-06-09T20:14:39.533407Z","iopub.execute_input":"2024-06-09T20:14:39.533743Z","iopub.status.idle":"2024-06-09T20:14:39.560756Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"test_paths = glob(f'{BASE_PATH}/test_soundscapes/*ogg')\n# During commit use `unlabeled` data as there is no `test` data.\n# During submission `test` data will automatically be populated.\nif len(test_paths)==0:\n    test_paths = glob(f'{BASE_PATH}/unlabeled_soundscapes/*ogg')[:10]\ntest_df = pd.DataFrame(test_paths, columns=['filepath'])\ntest_df.head()","metadata":{"execution":{"iopub.status.busy":"2024-06-09T20:14:39.567347Z","iopub.execute_input":"2024-06-09T20:14:39.567602Z","iopub.status.idle":"2024-06-09T20:14:39.997058Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# Create an input layer for the model\ni_inp = keras.layers.Input(shape=(None, None, 3))\n# Pretrained backbone\ni_backbone = keras_cv.models.EfficientNetV2Backbone.from_preset(\n    TestCFG.preset,\n)\ni_out = keras_cv.models.ImageClassifier(\n    backbone=i_backbone,\n    num_classes=TestCFG.num_classes,\n    name=\"classifier\"\n)(i_inp)\n# Build model\ni_model = keras.models.Model(inputs=i_inp, outputs=i_out)\n# Load weights of trained model\ni_model.load_weights(\"/kaggle/working/models/best_model_v2.weights.h5\")","metadata":{"execution":{"iopub.status.busy":"2024-06-09T20:14:40.000171Z","iopub.execute_input":"2024-06-09T20:14:40.000547Z","iopub.status.idle":"2024-06-09T20:14:47.241243Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def build_test_decoder(with_labels=True, dim=1024):\n    def get_audio(filepath):\n        file_bytes = tf.io.read_file(filepath)\n        audio = tfio.audio.decode_vorbis(file_bytes) # decode .ogg file\n        audio = tf.cast(audio, tf.float32)\n        if tf.shape(audio)[1]>1: # stereo -> mono\n            audio = audio[...,0:1]\n        audio = tf.squeeze(audio, axis=-1)\n        return audio\n    \n    def create_frames(audio, duration=5, sr=32000):\n        frame_size = int(duration * sr)\n        audio = tf.pad(audio[..., None], [[0, tf.shape(audio)[0] % frame_size], [0, 0]]) # pad the end\n        audio = tf.squeeze(audio) # remove extra dimension added for padding\n        frames = tf.reshape(audio, [-1, frame_size]) # shape: [num_frames, frame_size]\n        return frames\n    \n    def apply_preproc(spec):\n        # Standardize\n        mean = tf.math.reduce_mean(spec)\n        std = tf.math.reduce_std(spec)\n        spec = tf.where(tf.math.equal(std, 0), spec - mean, (spec - mean) / std)\n\n        # Normalize using Min-Max\n        min_val = tf.math.reduce_min(spec)\n        max_val = tf.math.reduce_max(spec)\n        spec = tf.where(tf.math.equal(max_val - min_val, 0), spec - min_val,\n                              (spec - min_val) / (max_val - min_val))\n        return spec\n\n    def decode(path):\n        # Load audio file\n        audio = get_audio(path)\n        # Split audio file into frames with each having 5 seecond duration\n        audio = create_frames(audio)\n        # Convert audio to spectrogram\n        spec = keras.layers.MelSpectrogram(num_mel_bins=TestCFG.img_size[0],\n                                             fft_length=TestCFG.nfft, \n                                              sequence_stride=TestCFG.hop_length, \n                                              sampling_rate=TestCFG.sample_rate)(audio)\n        # Apply normalization and standardization\n        spec = apply_preproc(spec)\n        # Covnert spectrogram to 3 channel image (for imagenet)\n        spec = tf.tile(spec[..., None], [1, 1, 1, 3])\n        return spec\n    \n    return decode","metadata":{"execution":{"iopub.status.busy":"2024-06-09T20:14:47.243463Z","iopub.execute_input":"2024-06-09T20:14:47.243781Z","iopub.status.idle":"2024-06-09T20:14:47.256911Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def build_test_dataset(paths, batch_size=1, decode_fn=None, cache=False):\n    if decode_fn is None:\n        decode_fn = build_test_decoder(dim=TestCFG.audio_len) # decoder\n    AUTO = tf.data.experimental.AUTOTUNE\n    slices = (paths,)\n    ds = tf.data.Dataset.from_tensor_slices(slices)\n    ds = ds.map(decode_fn, num_parallel_calls=AUTO) # decode audio to spectrograms then create frames\n    ds = ds.cache() if cache else ds # cache files\n    ds = ds.batch(batch_size, drop_remainder=False) # create batches\n    ds = ds.prefetch(AUTO)\n    return ds","metadata":{"execution":{"iopub.status.busy":"2024-06-09T20:14:47.257957Z","iopub.execute_input":"2024-06-09T20:14:47.258282Z","iopub.status.idle":"2024-06-09T20:14:47.281435Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# Initialize empty list to store ids\nids = []\n\n# Initialize empty array to store predictions\npreds = np.empty(shape=(0, TestCFG.num_classes), dtype='float32')\n\n# Build test dataset\ntest_paths = test_df.filepath.tolist()\ntest_ds = build_test_dataset(paths=test_paths, batch_size=1)\n\n# Iterate over each audio file in the test dataset\nfor idx, specs in enumerate(tqdm(iter(test_ds), desc='test ', total=len(test_df))):\n    # Extract the filename without the extension\n    filename = test_paths[idx].split('/')[-1].replace('.ogg','')\n    \n    # Convert to backend-specific tensor while excluding extra dimension\n    specs = keras.ops.convert_to_tensor(specs[0])\n    \n    # Predict bird species for all frames in a recording using all trained models\n    frame_preds = i_model.predict(specs, verbose=0)\n    \n    # Create a ID for each frame in a recording using the filename and frame number\n    frame_ids = [f'{filename}_{(frame_id+1)*5}' for frame_id in range(len(frame_preds))]\n    \n    # Concatenate the ids\n    ids += frame_ids\n    # Concatenate the predictions\n    preds = np.concatenate([preds, frame_preds], axis=0)","metadata":{"execution":{"iopub.status.busy":"2024-06-09T20:14:47.282582Z","iopub.execute_input":"2024-06-09T20:14:47.282818Z","iopub.status.idle":"2024-06-09T20:15:21.056142Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# Submit prediction\npred_df = pd.DataFrame(ids, columns=['row_id'])\npred_df.loc[:, CFG.class_names] = preds\npred_df.to_csv('submission.csv',index=False)\npred_df.head()","metadata":{"execution":{"iopub.status.busy":"2024-06-09T20:15:21.058230Z","iopub.execute_input":"2024-06-09T20:15:21.058650Z","iopub.status.idle":"2024-06-09T20:15:21.380343Z"},"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":[]},{"cell_type":"code","source":"","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"","metadata":{},"execution_count":null,"outputs":[]}]}