def main(argv):
speech_file = argv[1]
feature_type = argv[2]
augments = {
# "after": {
# "time_masking": {
# "num_masks": 10,
# "mask_factor": 100,
# "p_upperbound": 0.05
# },
# "freq_masking": {
# "mask_factor": 27
# }
# },
}
au = UserAugmentation(augments)
speech_conf = {
"sample_rate": 16000,
"frame_ms": 25,
"stride_ms": 10,
"feature_type": feature_type,
"preemphasis": 0.97,
"normalize_signal": True,
"normalize_feature": True,
"normalize_per_feature": False,
"num_feature_bins": 80,
}
signal = read_raw_audio(speech_file, speech_conf["sample_rate"])
sf = NumpySpeechFeaturizer(speech_conf)
ft = sf.extract(signal)
ft = au["after"].augment(ft)[:, :, 0]
plt.figure(figsize=(16, 2.5))
ax = plt.gca()
ax.set_title(f"{feature_type}", fontweight="bold")
librosa.display.specshow(ft.T, cmap="magma")
v1 = np.linspace(ft.min(), ft.max(), 8, endpoint=True)
plt.colorbar(pad=0.01, fraction=0.02, ax=ax, format="%.2f", ticks=v1)
plt.tight_layout()
plt.savefig(argv[3])
plt.show()
def preprocess(self, audio, transcript):
with tf.device("/CPU:0"):
signal = read_raw_audio(audio,
self.speech_featurizer_lms.sample_rate)
signal = self.augmentations["before"].augment(signal)
lms = self.speech_featurizer_lms.extract(signal)
lgs = self.speech_featurizer_lgs.extract(signal)
lms = self.augmentations["after"].augment(lms)
lgs = self.augmentations["after"].augment(lgs)
label = self.text_featurizer.extract(transcript.decode("utf-8"))
label_length = tf.cast(tf.shape(label)[0], tf.int32)
pred_inp = self.text_featurizer.prepand_blank(label)
lms = tf.convert_to_tensor(lms, tf.float32)
lgs = tf.convert_to_tensor(lgs, tf.float32)
input_length = tf.cast(tf.shape(lms)[0], tf.int32)
return lms, lgs, input_length, label, label_length, pred_inp
def main(argv):
speech_file = argv[1]
feature_type = argv[2]
speech_conf = {
"sample_rate": 16000,
"frame_ms": 25,
"stride_ms": 10,
"feature_type": feature_type,
"preemphasis": 0.97,
"normalize_signal": True,
"normalize_feature": True,
"normalize_per_feature": False,
"num_feature_bins": 80,
}
signal = read_raw_audio(speech_file, speech_conf["sample_rate"])
sf = TFSpeechFeaturizer(speech_conf)
ft = sf.extract(signal)[:, :, 0]
plt.figure(figsize=(15, 5))
plt.imshow(ft.T, origin="lower")
plt.colorbar()
plt.tight_layout()
plt.show()
parser.add_argument("--blank",
type=int,
default=0,
help="Path to conformer tflite")
parser.add_argument("--statesize",
type=int,
default=320,
help="Path to conformer tflite")
args = parser.parse_args()
tflitemodel = tf.lite.Interpreter(model_path=args.tflite)
signal = read_raw_audio(args.filename)
input_details = tflitemodel.get_input_details()
output_details = tflitemodel.get_output_details()
tflitemodel.resize_tensor_input(input_details[0]["index"], signal.shape)
tflitemodel.allocate_tensors()
tflitemodel.set_tensor(input_details[0]["index"], signal)
tflitemodel.set_tensor(input_details[1]["index"],
tf.constant(args.blank, dtype=tf.int32))
tflitemodel.set_tensor(input_details[2]["index"],
tf.zeros([1, 2, 1, args.statesize], dtype=tf.float32))
tflitemodel.invoke()
hyp = tflitemodel.get_tensor(output_details[0]["index"])
print("".join([chr(u) for u in hyp]))
model.save_weights("/tmp/transducer.h5")
model.add_featurizers(speech_featurizer=speech_featurizer,
text_featurizer=text_featurizer)
features = tf.zeros(shape=[5, 50, 80, 1], dtype=tf.float32)
pred = model.recognize(features)
print(pred)
pred = model.recognize_beam(features)
print(pred)
# stamp = datetime.datetime.now().strftime("%Y%m%d-%H%M%S")
# logdir = '/tmp/logs/func/%s' % stamp
# writer = tf.summary.create_file_writer(logdir)
#
signal = read_raw_audio("/home/nlhuy/Desktop/test/11003.wav",
speech_featurizer.sample_rate)
#
# tf.summary.trace_on(graph=True, profiler=True)
# hyps = model.recognize_tflite(signal, 0, tf.zeros([1, 2, 1, 320], dtype=tf.float32))
# with writer.as_default():
# tf.summary.trace_export(
# name="recognize_tflite",
# step=0,
# profiler_outdir=logdir)
#
# print(hyps[0])
#
# # hyps = model.recognize_beam(features)
#
#
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 = CharFeaturizer(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()