def convert_to_TF(deep_speech_dataset):
data_entries = deep_speech_dataset.entries
num_feature_bins = deep_speech_dataset.num_feature_bins
audio_featurizer = deep_speech_dataset.audio_featurizer
feature_normalize = deep_speech_dataset.config.audio_config.normalize
text_featurizer = deep_speech_dataset.text_featurizer
filename = "E:/TUDA/german-speechdata-package-v2/test.tfrecords"
print('Writing', filename)
with tf.python_io.TFRecordWriter(filename) as writer:
for audio_file, _, transcript in data_entries:
features = dataset._preprocess_audio(audio_file, audio_featurizer,
feature_normalize)
labels = featurizer.compute_label_feature(
transcript, text_featurizer.token_to_index)
flattened_features = [
item for sublist_20ms in features for item in sublist_20ms
]
example = tf.train.Example(features=tf.train.Features(
feature={
'dim1':
_int64_feature(len(features)),
'dim2':
_int64_feature(num_feature_bins),
'dim3':
_int64_feature(1),
'labels':
_bytes_feature(np.asarray(labels).tostring()),
'features':
_bytes_feature(np.asarray(flattened_features).tostring())
}))
writer.write(example.SerializeToString())
def gen_TFRecord(deep_speech_dataset):
data_entries = deep_speech_dataset.entries
num_feature_bins = deep_speech_dataset.num_feature_bins
audio_featurizer = deep_speech_dataset.audio_featurizer
feature_normalize = deep_speech_dataset.config.audio_config.normalize
text_featurizer = deep_speech_dataset.text_featurizer
writers = []
records_csv = []
for data_entry in data_entries:
#TODO BUCKETT
writers.append(
tf.python_io.TFRecordWriter(data_entry[0][:-4] + ".tfrecord"))
records_csv.append(data_entry[0][:-4] + ".tfrecord")
"""Dataset generator function."""
writer_index = 0
for audio_file, _, transcript in data_entries:
features = dataset._preprocess_audio(audio_file, audio_featurizer,
feature_normalize)
labels = featurizer.compute_label_feature(
transcript, text_featurizer.token_to_index)
flattened_features = [
item for sublist_20ms in features for item in sublist_20ms
]
ff = tf.reshape(flattened_features, [161, 1])
feature_dict = collections.OrderedDict()
feature_dict["features"] = _bytes_feature(
tf.compat.as_bytes(np.asarray(flattened_features).tostring()))
# create_float_feature(flattened_features)
feature_dict["shape"] = _bytes_feature(
tf.compat.as_bytes(
np.asarray([len(features), num_feature_bins, 1]).tostring()))
#create_int_feature([len(features),num_feature_bins,1])
feature_dict["labels"] = _bytes_feature(
tf.compat.as_bytes(np.asarray(labels).tostring()))
#create_int_feature(labels)
tf_example = tf.train.Example(features=tf.train.Features(
feature=feature_dict))
writers[writer_index].write(tf_example.SerializeToString())
writer_index += 1
print("WRITING: " + str(writer_index + 1) + "/" + str(len(writers)),
end='\r')
for writer in writers:
writer.close()
with open("E:/TUDA/german-speechdata-package-v2/records_test.csv",
'w') as f:
for record in records_csv:
f.write(record + '\n')
def _gen_data():
"""Dataset generator function."""
for audio_file, _, transcript in data_entries:
features = _preprocess_audio(audio_file, audio_featurizer,
feature_normalize)
labels = featurizer.compute_label_feature(
transcript, text_featurizer.token_to_index)
input_length = [features.shape[0]]
label_length = [len(labels)]
# Yield a tuple of (features, labels) where features is a dict containing
# all info about the actual data features.
yield ({
"features": features,
"input_length": input_length,
"label_length": label_length
}, labels)
def _gen_data():
"""Dataset generator function."""
for audio_file, _, transcript in data_entries:
features = _preprocess_audio(
audio_file, audio_featurizer, feature_normalize)
labels = featurizer.compute_label_feature(
transcript, text_featurizer.token_to_index)
input_length = [features.shape[0]]
label_length = [len(labels)]
# Yield a tuple of (features, labels) where features is a dict containing
# all info about the actual data features.
yield (
{
"features": features,
"input_length": input_length,
"label_length": label_length
},
labels)
def check_reshape(deep_speech_dataset):
data_entries = deep_speech_dataset.entries
num_feature_bins = deep_speech_dataset.num_feature_bins
audio_featurizer = deep_speech_dataset.audio_featurizer
feature_normalize = deep_speech_dataset.config.audio_config.normalize
text_featurizer = deep_speech_dataset.text_featurizer
"""Dataset generator function."""
for audio_file, _, transcript in data_entries:
features = dataset._preprocess_audio(audio_file, audio_featurizer,
feature_normalize)
labels = featurizer.compute_label_feature(
transcript, text_featurizer.token_to_index)
flattened_features = [
item for sublist_20ms in features for item in sublist_20ms
]
#ff = tf.reshape(flattened_features,[features.shape[0],features.shape[1]])
ff = np.reshape(flattened_features,
[features.shape[0], features.shape[1], 1])
#for i in range(len(features)):
# for j in range(len(features[i])):
#print(str(ff[i][j]) +"#############"+str( features[i][j]))
#print(ff == features)
print(ff[0])
print("################")
print(features[0])
#print(features.shape)
break
#ds = generate_dataset("E:/TUDA/german-speechdata-package-v2/test.csv")
#convert_to_TF(ds)
#check_reshape(ds)
#gen_TFRecord(ds)
#input_fn(128, "E:/TUDA/german-speechdata-package-v2/records_test.csv", 1)
#input_fn(128, "E:/TUDA/german-speechdata-package-v2/records_test.csv", 1)
#read_tfRecord()
def _preprocess_transcript(transcript, token_to_index):
"""Process transcript as label features."""
return featurizer.compute_label_feature(transcript, token_to_index)