train.py

import librosa.display
from keras.optimizers import Adam
from keras.optimizers.schedules import ExponentialDecay
from keras.losses import SparseCategoricalCrossentropy
from keras.metrics import SparseCategoricalAccuracy
from keras.callbacks import TensorBoard, ModelCheckpoint
from model import create_model
from data import make_dataset,split_data

model = create_model()

lr_schedule = ExponentialDecay(initial_learning_rate=1e-3,
                               decay_steps= 4000, decay_rate = 0.5, staircase=False)

model.compile(optimizer = Adam(lr_schedule),
              loss = SparseCategoricalCrossentropy(from_logits = True),
              metrics = [ SparseCategoricalAccuracy()])

# Creating callbacks
logdir = 'tb_logs/model'
tensorboard_callback = TensorBoard(log_dir=logdir)

checkpoint_filepath = 'model_checkpoint/model'
model_checkpoint_callback = ModelCheckpoint(
    filepath=checkpoint_filepath,
    save_weights_only=False,
    monitor='val_sparse_categorical_accuracy',
    mode='max',
    save_best_only=True)

bird_filepaths_train, bird_filepaths_val, bird_labels_train, bird_labels_val = split_data()
dataset_train = make_dataset(bird_labels_train, bird_filepaths_train, shuffle=True)
dataset_val = make_dataset(bird_labels_val, bird_filepaths_val, shuffle=False)

training = model.fit(dataset_train, validation_data = dataset_val, epochs=120,
                     callbacks=[tensorboard_callback,model_checkpoint_callback])