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,
n_input_hops,
batch_size,
mode="train",
aug_kind_str=aug_kind_str,
tfr_str=tfr_str,
partial_labels=False,
structured=False,
single_output="coarse",
active_streamers=active_streamers,
streamer_rate=streamer_rate,
num_cpus=num_cpus,
align_perturb=align_perturb)
history = model.fit_generator(training_streamer,
steps_per_epoch=steps_per_epoch,
epochs=4,
verbose=False,
callbacks=[history_callback],
workers=0,
validation_data=validation_data,
use_multiprocessing=True,
dense2 = keras.layers.Dense(
1,
kernel_initializer="normal",
activation="sigmoid",
kernel_regularizer=keras.regularizers.l2(0.00002))(dense1)
# Compile model, print model summary.
model = keras.models.Model(inputs=inputs, outputs=dense2)
model.compile(loss="binary_crossentropy",
optimizer="adam",
metrics=["accuracy"])
model.summary()
# Build Pescador streamers corresponding to log-mel-spectrograms in augmented
# training and validation sets.
training_streamer = localmodule.multiplex_tfr(aug_kind_str, training_units,
n_input_hops, batch_size)
validation_streamer = localmodule.multiplex_tfr(aug_kind_str, validation_units,
n_input_hops, batch_size)
# Create directory for model, unit, and trial.
model_name = "icassp-convnet"
if not aug_kind_str == "none":
model_name = "_".join([model_name, "aug-" + aug_kind_str])
model_dir = os.path.join(models_dir, model_name)
os.makedirs(model_dir, exist_ok=True)
unit_dir = os.path.join(model_dir, unit_str)
os.makedirs(unit_dir, exist_ok=True)
trial_dir = os.path.join(unit_dir, trial_str)
os.makedirs(trial_dir, exist_ok=True)
# Define Keras callback for checkpointing model.