def create_eval_dags(
neural_factory, neural_modules, waveglow_params, eval_datasets, eval_batch_size, eval_freq, cpu_per_dl=1,
):
data_preprocessor, waveglow, _ = neural_modules
eval_dl_params = copy.deepcopy(waveglow_params["AudioDataLayer"])
eval_dl_params.update(waveglow_params["AudioDataLayer"]["eval"])
del eval_dl_params["train"]
del eval_dl_params["eval"]
callbacks = []
# assemble eval DAGs
for eval_dataset in eval_datasets:
data_layer_eval = nemo_tts.AudioDataLayer(
manifest_filepath=eval_dataset, batch_size=eval_batch_size, num_workers=cpu_per_dl, **eval_dl_params,
)
audio, audio_len, = data_layer_eval()
spec_target, spec_target_len = data_preprocessor(input_signal=audio, length=audio_len)
audio_pred, log_s_list, log_det_W_list = waveglow(mel_spectrogram=spec_target, audio=audio)
# create corresponding eval callback
tagname = os.path.basename(eval_dataset).split(".")[0]
eval_callback = nemo.core.EvaluatorCallback(
eval_tensors=[audio_pred, spec_target, spec_target_len],
user_iter_callback=waveglow_process_eval_batch,
user_epochs_done_callback=lambda x: x,
tb_writer_func=partial(waveglow_eval_log_to_tb_func, tag=tagname, mel_fb=data_preprocessor.filter_banks,),
eval_step=eval_freq,
tb_writer=neural_factory.tb_writer,
)
callbacks.append(eval_callback)
return callbacks
def test_waveglow_training(self):
data_layer = nemo_tts.AudioDataLayer(
manifest_filepath=self.manifest_filepath,
n_segments=4000,
batch_size=4,
)
preprocessing = nemo_asr.AudioToMelSpectrogramPreprocessor(
window_size=None,
window_stride=None,
n_window_size=512,
n_window_stride=128,
normalize=None,
preemph=None,
dither=0,
mag_power=1.0,
pad_value=-11.52,
)
waveglow = nemo_tts.WaveGlowNM(
n_mel_channels=64,
n_flows=6,
n_group=4,
n_early_every=4,
n_early_size=2,
n_wn_layers=4,
n_wn_channels=256,
wn_kernel_size=3,
)
waveglow_loss = nemo_tts.WaveGlowLoss()
# DAG
audio, audio_len, = data_layer()
spec_target, _ = preprocessing(input_signal=audio, length=audio_len)
z, log_s_list, log_det_W_list = waveglow(mel_spectrogram=spec_target,
audio=audio)
loss_t = waveglow_loss(z=z,
log_s_list=log_s_list,
log_det_W_list=log_det_W_list)
callback = nemo.core.SimpleLossLoggerCallback(
tensors=[loss_t],
print_func=lambda x: logging.info(f'Train Loss: {str(x[0].item())}'
),
)
# Instantiate an optimizer to perform `train` action
neural_factory = nemo.core.NeuralModuleFactory(
backend=nemo.core.Backend.PyTorch,
local_rank=None,
create_tb_writer=False,
)
optimizer = neural_factory.get_trainer()
optimizer.train(
[loss_t],
callbacks=[callback],
optimizer="sgd",
optimization_params={
"num_epochs": 10,
"lr": 0.0003
},
)
def test_waveglow_training(self):
"""Integtaion test that instantiates a smaller WaveGlow model and tests training with the sample asr data.
Training is run for 3 forward and backward steps and asserts that loss after 3 steps is smaller than the loss
at the first step.
"""
data_layer = nemo_tts.AudioDataLayer(
manifest_filepath=self.manifest_filepath, n_segments=4000, batch_size=4, sample_rate=16000
)
preprocessing = nemo_asr.AudioToMelSpectrogramPreprocessor(
window_size=None,
window_stride=None,
n_window_size=512,
n_window_stride=128,
normalize=None,
preemph=None,
dither=0,
mag_power=1.0,
pad_value=-11.52,
)
waveglow = nemo_tts.WaveGlowNM(
n_mel_channels=64,
n_flows=6,
n_group=4,
n_early_every=4,
n_early_size=2,
n_wn_layers=4,
n_wn_channels=256,
wn_kernel_size=3,
sample_rate=16000,
)
waveglow_loss = nemo_tts.WaveGlowLoss(sample_rate=16000)
# DAG
audio, audio_len, = data_layer()
spec_target, _ = preprocessing(input_signal=audio, length=audio_len)
z, log_s_list, log_det_W_list = waveglow(mel_spectrogram=spec_target, audio=audio)
loss_t = waveglow_loss(z=z, log_s_list=log_s_list, log_det_W_list=log_det_W_list)
loss_list = []
callback = SimpleLossLoggerCallback(
tensors=[loss_t], print_func=partial(self.print_and_log_loss, loss_log_list=loss_list), step_freq=1
)
# Instantiate an optimizer to perform `train` action
optimizer = PtActions()
optimizer.train(
[loss_t], callbacks=[callback], optimizer="sgd", optimization_params={"max_steps": 3, "lr": 0.01}
)
# Assert that training loss went down
assert loss_list[-1] < loss_list[0]
def create_train_dag(
neural_factory,
neural_modules,
waveglow_params,
train_dataset,
batch_size,
checkpoint_save_freq,
cpu_per_dl=1,
):
data_preprocessor, waveglow, waveglow_loss = neural_modules
train_dl_params = copy.deepcopy(waveglow_params["AudioDataLayer"])
train_dl_params.update(waveglow_params["AudioDataLayer"]["train"])
del train_dl_params["train"]
del train_dl_params["eval"]
data_layer = nemo_tts.AudioDataLayer(
manifest_filepath=train_dataset,
batch_size=batch_size,
num_workers=cpu_per_dl,
**train_dl_params,
)
N = len(data_layer)
steps_per_epoch = int(N / (batch_size * neural_factory.world_size))
logging.info('Have {0} examples to train on.'.format(N))
# Train DAG
audio, audio_len, = data_layer()
spec_target, spec_target_len = data_preprocessor(input_signal=audio,
length=audio_len)
z, log_s_list, log_det_W_list = waveglow(mel_spectrogram=spec_target,
audio=audio)
loss_t = waveglow_loss(z=z,
log_s_list=log_s_list,
log_det_W_list=log_det_W_list)
# Callbacks needed to print info to console and Tensorboard
train_callback = nemo.core.SimpleLossLoggerCallback(
tensors=[loss_t, z, spec_target, spec_target_len],
print_func=lambda x: logging.info(f"Loss: {x[0].data}"),
log_to_tb_func=partial(waveglow_log_to_tb_func, log_images=False),
tb_writer=neural_factory.tb_writer,
)
chpt_callback = nemo.core.CheckpointCallback(
folder=neural_factory.checkpoint_dir, step_freq=checkpoint_save_freq)
callbacks = [train_callback, chpt_callback]
return loss_t, callbacks, steps_per_epoch
