from tiramisu_asr.utils.utils import bytes_to_string, merge_two_last_dims
decoder_config = {
"vocabulary":
"/mnt/Projects/asrk16/TiramisuASR/examples/deepspeech2/vocabularies/vietnamese.txt",
"beam_width": 100,
"blank_at_zero": True,
"lm_config": {
"model_path": "/mnt/Data/ML/NLP/vntc_asrtrain_5gram_trie.binary",
"alpha": 2.0,
"beta": 2.0
}
}
text_featurizer = TextFeaturizer(decoder_config)
text_featurizer.add_scorer(
Scorer(**decoder_config["lm_config"],
vocabulary=text_featurizer.vocab_array))
speech_featurizer = TFSpeechFeaturizer({
"sample_rate": 16000,
"frame_ms": 25,
"stride_ms": 10,
"num_feature_bins": 80,
"feature_type": "spectrogram",
"preemphasis": 0.97,
# "delta": True,
# "delta_delta": True,
"normalize_signal": True,
"normalize_feature": True,
"normalize_per_feature": False,
# "pitch": False,
})
def run(args):
assert args.mode in modes, f"Mode must in {modes}"
config = UserConfig(DEFAULT_YAML, args.config, learning=True)
speech_featurizer = SpeechFeaturizer(config["speech_config"])
text_featurizer = TextFeaturizer(config["decoder_config"])
if args.mode == "train":
tf.random.set_seed(2020)
if args.mixed_precision:
policy = tf.keras.mixed_precision.experimental.Policy("mixed_float16")
tf.keras.mixed_precision.experimental.set_policy(policy)
print("Enabled mixed precision training")
ctc_trainer = CTCTrainer(speech_featurizer, text_featurizer,
config["learning_config"]["running_config"],
args.mixed_precision)
if args.tfrecords:
train_dataset = ASRTFRecordDataset(
config["learning_config"]["dataset_config"]["train_paths"],
config["learning_config"]["dataset_config"]["tfrecords_dir"],
speech_featurizer, text_featurizer, "train",
augmentations=config["learning_config"]["augmentations"], shuffle=True,
)
eval_dataset = ASRTFRecordDataset(
config["learning_config"]["dataset_config"]["eval_paths"],
config["learning_config"]["dataset_config"]["tfrecords_dir"],
speech_featurizer, text_featurizer, "eval", shuffle=False
)
else:
train_dataset = ASRSliceDataset(
stage="train", speech_featurizer=speech_featurizer,
text_featurizer=text_featurizer,
data_paths=config["learning_config"]["dataset_config"]["train_paths"],
augmentations=config["learning_config"]["augmentations"], shuffle=True,
)
eval_dataset = ASRSliceDataset(
stage="eval", speech_featurizer=speech_featurizer,
text_featurizer=text_featurizer,
data_paths=config["learning_config"]["dataset_config"]["eval_paths"],
shuffle=False
)
# Build DS2 model
f, c = speech_featurizer.compute_feature_dim()
with ctc_trainer.strategy.scope():
satt_ds2_model = SelfAttentionDS2(input_shape=[None, f, c],
arch_config=config["model_config"],
num_classes=text_featurizer.num_classes)
satt_ds2_model._build([1, 50, f, c])
optimizer = create_optimizer(
name=config["learning_config"]["optimizer_config"]["name"],
d_model=config["model_config"]["att"]["head_size"],
**config["learning_config"]["optimizer_config"]["config"]
)
# Compile
ctc_trainer.compile(satt_ds2_model, optimizer, max_to_keep=args.max_ckpts)
ctc_trainer.fit(train_dataset, eval_dataset, args.eval_train_ratio)
if args.export:
if args.from_weights:
ctc_trainer.model.save_weights(args.export)
else:
ctc_trainer.model.save(args.export)
elif args.mode == "test":
tf.random.set_seed(0)
assert args.export
text_featurizer.add_scorer(
Scorer(**text_featurizer.decoder_config["lm_config"],
vocabulary=text_featurizer.vocab_array))
# Build DS2 model
f, c = speech_featurizer.compute_feature_dim()
satt_ds2_model = SelfAttentionDS2(input_shape=[None, f, c],
arch_config=config["model_config"],
num_classes=text_featurizer.num_classes)
satt_ds2_model._build([1, 50, f, c])
satt_ds2_model.summary(line_length=100)
optimizer = create_optimizer(
name=config["learning_config"]["optimizer_config"]["name"],
d_model=config["model_config"]["att"]["head_size"],
**config["learning_config"]["optimizer_config"]["config"]
)
batch_size = config["learning_config"]["running_config"]["batch_size"]
if args.tfrecords:
test_dataset = ASRTFRecordDataset(
config["learning_config"]["dataset_config"]["test_paths"],
config["learning_config"]["dataset_config"]["tfrecords_dir"],
speech_featurizer, text_featurizer, "test",
augmentations=config["learning_config"]["augmentations"], shuffle=False
).create(batch_size * args.eval_train_ratio)
else:
test_dataset = ASRSliceDataset(
stage="test", speech_featurizer=speech_featurizer,
text_featurizer=text_featurizer,
data_paths=config["learning_config"]["dataset_config"]["test_paths"],
augmentations=config["learning_config"]["augmentations"], shuffle=False
).create(batch_size * args.eval_train_ratio)
ctc_tester = BaseTester(
config=config["learning_config"]["running_config"],
saved_path=args.export, from_weights=args.from_weights
)
ctc_tester.compile(satt_ds2_model, speech_featurizer, text_featurizer)
ctc_tester.run(test_dataset)
else:
assert args.export
# Build DS2 model
f, c = speech_featurizer.compute_feature_dim()
satt_ds2_model = SelfAttentionDS2(input_shape=[None, f, c],
arch_config=config["model_config"],
num_classes=text_featurizer.num_classes)
satt_ds2_model._build([1, 50, f, c])
optimizer = create_optimizer(
name=config["learning_config"]["optimizer_config"]["name"],
d_model=config["model_config"]["att"]["head_size"],
**config["learning_config"]["optimizer_config"]["config"]
)
def save_func(**kwargs):
if args.from_weights:
kwargs["model"].save_weights(args.export)
else:
kwargs["model"].save(args.export)
save_from_checkpoint(func=save_func,
outdir=config["learning_config"]["running_config"]["outdir"],
model=satt_ds2_model, optimizer=optimizer)
def main():
parser = argparse.ArgumentParser(prog="SelfAttentionDS2 Histogram")
parser.add_argument("--config", type=str, default=None,
help="Config file")
parser.add_argument("--audio", type=str, default=None,
help="Audio file")
parser.add_argument("--saved_model", type=str, default=None,
help="Saved model")
parser.add_argument("--from_weights", type=bool, default=False,
help="Load from weights")
parser.add_argument("--output", type=str, default=None,
help="Output dir storing histograms")
args = parser.parse_args()
config = UserConfig(args.config, args.config, learning=False)
speech_featurizer = SpeechFeaturizer(config["speech_config"])
text_featurizer = TextFeaturizer(config["decoder_config"])
text_featurizer.add_scorer(Scorer(**text_featurizer.decoder_config["lm_config"],
vocabulary=text_featurizer.vocab_array))
f, c = speech_featurizer.compute_feature_dim()
satt_ds2_model = SelfAttentionDS2(input_shape=[None, f, c],
arch_config=config["model_config"],
num_classes=text_featurizer.num_classes)
satt_ds2_model._build([1, 50, f, c])
if args.from_weights:
satt_ds2_model.load_weights(args.saved_model)
else:
saved_model = tf.keras.models.load_model(args.saved_model)
satt_ds2_model.set_weights(saved_model.get_weights())
satt_ds2_model.summary(line_length=100)
satt_ds2_model.add_featurizers(speech_featurizer, text_featurizer)
signal = read_raw_audio(args.audio, speech_featurizer.sample_rate)
features = speech_featurizer.extract(signal)
decoded = satt_ds2_model.recognize_beam(tf.expand_dims(features, 0), lm=True)
print(bytes_to_string(decoded.numpy()))
for i in range(1, len(satt_ds2_model.base_model.layers)):
func = tf.keras.backend.function([satt_ds2_model.base_model.input],
[satt_ds2_model.base_model.layers[i].output])
data = func([np.expand_dims(features, 0), 1])[0][0]
print(data.shape)
data = data.flatten()
plt.hist(data, 200, color='green', histtype="stepfilled")
plt.title(f"Output of {satt_ds2_model.base_model.layers[i].name}", fontweight="bold")
plt.savefig(os.path.join(
args.output, f"{i}_{satt_ds2_model.base_model.layers[i].name}.png"))
plt.clf()
plt.cla()
plt.close()
fc = satt_ds2_model(tf.expand_dims(features, 0), training=False)
plt.hist(fc[0].numpy().flatten(), 200, color="green", histtype="stepfilled")
plt.title(f"Output of {satt_ds2_model.layers[-1].name}", fontweight="bold")
plt.savefig(os.path.join(args.output, f"{satt_ds2_model.layers[-1].name}.png"))
plt.clf()
plt.cla()
plt.close()
fc = tf.nn.softmax(fc)
plt.hist(fc[0].numpy().flatten(), 10, color="green", histtype="stepfilled")
plt.title("Output of softmax", fontweight="bold")
plt.savefig(os.path.join(args.output, "softmax_hist.png"))
plt.clf()
plt.cla()
plt.close()
plt.hist(features.flatten(), 200, color="green", histtype="stepfilled")
plt.title("Log Mel Spectrogram", fontweight="bold")
plt.savefig(os.path.join(args.output, "log_mel_spectrogram.png"))
plt.clf()
plt.cla()
plt.close()
def main():
parser = argparse.ArgumentParser(prog="SelfAttentionDS2 Histogram")
parser.add_argument("--config", type=str, default=None, help="Config file")
parser.add_argument("--audio", type=str, default=None, help="Audio file")
parser.add_argument("--saved_model",
type=str,
default=None,
help="Saved model")
parser.add_argument("--from_weights",
type=bool,
default=False,
help="Load from weights")
parser.add_argument("--output",
type=str,
default=None,
help="Output dir storing histograms")
args = parser.parse_args()
config = UserConfig(args.config, args.config, learning=False)
speech_featurizer = SpeechFeaturizer(config["speech_config"])
text_featurizer = TextFeaturizer(config["decoder_config"])
text_featurizer.add_scorer(
Scorer(**text_featurizer.decoder_config["lm_config"],
vocabulary=text_featurizer.vocab_array))
f, c = speech_featurizer.compute_feature_dim()
satt_ds2_model = SelfAttentionDS2(input_shape=[None, f, c],
arch_config=config["model_config"],
num_classes=text_featurizer.num_classes)
satt_ds2_model._build([1, 50, f, c])
if args.from_weights:
satt_ds2_model.load_weights(args.saved_model)
else:
saved_model = tf.keras.models.load_model(args.saved_model)
satt_ds2_model.set_weights(saved_model.get_weights())
satt_ds2_model.summary(line_length=100)
satt_ds2_model.add_featurizers(speech_featurizer, text_featurizer)
signal = read_raw_audio(args.audio, speech_featurizer.sample_rate)
features = speech_featurizer.extract(signal)
decoded = satt_ds2_model.recognize_beam(tf.expand_dims(features, 0),
lm=True)
print(bytes_to_string(decoded.numpy()))
# for i in range(1, len(satt_ds2_model.base_model.layers)):
# func = tf.keras.backend.function([satt_ds2_model.base_model.input],
# [satt_ds2_model.base_model.layers[i].output])
# data = func([np.expand_dims(features, 0), 1])[0][0]
# print(data.shape)
# plt.figure(figsize=(16, 5))
# ax = plt.gca()
# im = ax.imshow(data.T, origin="lower", aspect="auto")
# ax.set_title(f"{satt_ds2_model.base_model.layers[i].name}", fontweight="bold")
# divider = make_axes_locatable(ax)
# cax = divider.append_axes("right", size="5%", pad=0.05)
# plt.colorbar(im, cax=cax)
# plt.savefig(os.path.join(
# args.output, f"{i}_{satt_ds2_model.base_model.layers[i].name}.png"))
# plt.clf()
# plt.cla()
# plt.close()
fc = satt_ds2_model(tf.expand_dims(features, 0), training=False)
plt.figure(figsize=(16, 5))
ax = plt.gca()
ax.set_title(f"{satt_ds2_model.layers[-1].name}", fontweight="bold")
im = ax.imshow(fc[0].numpy().T, origin="lower", aspect="auto")
divider = make_axes_locatable(ax)
cax = divider.append_axes("right", size="5%", pad=0.05)
plt.colorbar(im, cax=cax)
plt.savefig(
os.path.join(args.output, f"{satt_ds2_model.layers[-1].name}.png"))
plt.clf()
plt.cla()
plt.close()
fc = tf.nn.softmax(fc)
plt.figure(figsize=(16, 5))
ax = plt.gca()
ax.set_title("Softmax", fontweight="bold")
im = ax.imshow(fc[0].numpy().T, origin="lower", aspect="auto")
divider = make_axes_locatable(ax)
cax = divider.append_axes("right", size="5%", pad=0.05)
plt.colorbar(im, cax=cax)
plt.savefig(os.path.join(args.output, "softmax.png"))
plt.clf()
plt.cla()
plt.close()
plt.figure(figsize=(16, 5))
ax = plt.gca()
ax.set_title("Log Mel Spectrogram", fontweight="bold")
im = ax.imshow(features[:, :, 0].T, origin="lower", aspect="auto")
divider = make_axes_locatable(ax)
cax = divider.append_axes("right", size="5%", pad=0.05)
plt.colorbar(im, cax=cax)
plt.savefig(os.path.join(args.output, "features.png"))
plt.clf()
plt.cla()
plt.close()