type=str,
default=None,
help="Path to saved model")
parser.add_argument("output",
type=str,
default=None,
help="TFLite file path to be exported")
args = parser.parse_args()
assert args.saved and args.output
config = UserConfig(DEFAULT_YAML, args.config, learning=True)
speech_featurizer = TFSpeechFeaturizer(config["speech_config"])
text_featurizer = CharFeaturizer(config["decoder_config"])
# build model
conformer = Conformer(**config["model_config"],
vocabulary_size=text_featurizer.num_classes)
conformer._build(speech_featurizer.shape)
conformer.load_weights(args.saved)
conformer.summary(line_length=150)
conformer.add_featurizers(speech_featurizer, text_featurizer)
concrete_func = conformer.make_tflite_function(
greedy=True).get_concrete_function()
converter = tf.lite.TFLiteConverter.from_concrete_functions([concrete_func])
converter.optimizations = [tf.lite.Optimize.DEFAULT]
converter.target_spec.supported_ops = [
tf.lite.OpsSet.TFLITE_BUILTINS, tf.lite.OpsSet.SELECT_TF_OPS
"num_masks": 10,
"mask_factor": 100,
"p_upperbound": 0.05
},
"freq_masking": {
"mask_factor": 27
}
},
"include_original": False
}
data = "/mnt/Data/ML/ASR/Raw/LibriSpeech/train-clean-100/transcripts.tsv"
text_featurizer = CharFeaturizer({
"vocabulary": None,
"blank_at_zero": True,
"beam_width": 5,
"norm_score": True
})
speech_featurizer = TFSpeechFeaturizer({
"sample_rate": 16000,
"frame_ms": 25,
"stride_ms": 10,
"num_feature_bins": 80,
"feature_type": "log_mel_spectrogram",
"preemphasis": 0.97,
"normalize_signal": True,
"normalize_feature": True,
"normalize_per_feature": False
})
from tiramisu_asr.featurizers.text_featurizers import CharFeaturizer
txf = CharFeaturizer(None, blank_at_zero=True)
a = txf.extract("fkaff aksfbfnak kcjhoiu")
print(a)
from tiramisu_asr.models.ctc import CtcModel
from tiramisu_asr.featurizers.text_featurizers import CharFeaturizer
from tiramisu_asr.featurizers.speech_featurizers import TFSpeechFeaturizer, read_raw_audio
from tiramisu_asr.utils.utils import bytes_to_string, merge_two_last_dims
decoder_config = {
"vocabulary": "/mnt/Projects/asrk16/TiramisuASR/vocabularies/vietnamese.txt",
"beam_width": 100,
"blank_at_zero": False,
"lm_config": {
"model_path": "/mnt/Data/ML/NLP/vntc_asrtrain_5gram_trie.binary",
"alpha": 2.0,
"beta": 2.0
}
}
text_featurizer = CharFeaturizer(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 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)
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()