history_callback = keras.callbacks.LambdaCallback(
on_epoch_end=lambda epoch, logs: write_row(history_path, epoch, logs))
callbacks = [checkpoint, history_callback]
if lr_annealing:
reduce_lr = keras.callbacks.ReduceLROnPlateau()
callbacks.append(reduce_lr)
print("Loading validation data.")
# Load validation data once to save on IO costs
validation_data = localmodule.get_validation_data(valid_hdf5_dir,
n_input_hops,
valid_batch_size,
validation_steps,
tfr_str=tfr_str,
structured=False,
label_inputs=False,
single_output="coarse")
print("Performing rejection sampling for initialization.")
# Rejection sampling for best initialization.
n_inits = 10
for init_id in range(n_inits):
model = keras.models.Model(inputs=inputs, outputs=outputs)
model.compile(loss='binary_crossentropy',
optimizer=Adam(lr),
metrics=["accuracy"])
training_streamer = localmodule.multiplex_tfr(
train_data_dir,
"y_medium": [categorical_accuracy],
"y_fine": [categorical_accuracy]
}
loss_weights = {
"y_coarse": loss_weight_coarse,
"y_medium": loss_weight_medium,
"y_fine": 1.0
}
print("Loading validation data.")
# Load validation data once to save on IO costs
validation_data = localmodule.get_validation_data(valid_hdf5_dir,
n_input_hops,
valid_batch_size,
validation_steps,
tfr_str=tfr_str,
structured=True,
label_inputs=False)
print("Performing rejection sampling for initialization.")
# Rejection sampling for best initialization.
n_inits = 10
for init_id in range(n_inits):
model = keras.models.Model(inputs=inputs, outputs=outputs)
model.compile(loss=losses,
loss_weights=loss_weights,
optimizer=Adam(lr),
metrics=metrics)
training_streamer = localmodule.multiplex_tfr(