cRM2 = np.abs(F2) / np.abs(FM)
cRM2[~np.isfinite(cRM2)] = 0
cRM = np.zeros((1, 298, 257, 2, 2))
cRM[0, :, :, :, 0] = cRM1
cRM[0, :, :, :, 1] = cRM2
print('shape of cRM should be (1,298,257,2,2):', cRM.shape)
input_dim = (298, 257, 2)
output_dim = (298, 257, 2, 2)
filepath = "model-{epoch:02d}-{val_loss:.4f}.h5"
checkpoint = ModelCheckpoint(filepath,
monitor='val_loss',
verbose=1,
save_best_only=True,
mode='min')
model = model_AO.AO_model(2)
# Note: added compile to bypass error
adam = optimizers.Adam()
model.compile(optimizer=adam, loss='mse')
model.fit(FM,
cRM,
epochs=5,
batch_size=2,
validation_data=(FM, cRM),
shuffle=False,
callbacks=[TensorBoard(log_dir='./log'), checkpoint])
pred_cRM = model.predict(FM)
with open((database_dir_path + '/AV_log/AVdataset_train.txt'), 'r') as t:
trainfile = t.readlines()
with open((database_dir_path + '/AV_log/AVdataset_val.txt'), 'r') as v:
valfile = v.readlines()
# ///////////////////////////////////////////////////////// #
# the training steps
if RESTORE:
latest_file = latest_file(path + '/')
print("Restoring model from:", latest_file)
loss = audio_loss(gamma=gamma_loss, beta=beta_loss, num_speaker=people_num)
AO_model = load_model(latest_file, custom_objects={"loss_func": loss})
info = latest_file.strip().split('-')
initial_epoch = int(info[-2])
else:
AO_model = AO.AO_model(people_num)
train_generator = AudioGenerator(trainfile,
database_dir_path=database_dir_path,
batch_size=batch_size,
shuffle=True)
val_generator = AudioGenerator(valfile,
database_dir_path=database_dir_path,
batch_size=batch_size,
shuffle=True)
if NUM_GPU > 1:
parallel_model = ModelMGPU(AO_model, NUM_GPU)
adam = optimizers.Adam()
loss = audio_loss(gamma=gamma_loss, beta=beta_loss, num_speaker=people_num)
parallel_model.compile(loss=loss, optimizer=adam, metrics=['accuracy'])