def __init__(self, config: dict):
"""
running_config:
batch_size: 8
num_epochs: 20
outdir: ...
log_interval_steps: 200
eval_interval_steps: 200
save_interval_steps: 200
"""
self.config = config
self.config["outdir"] = preprocess_paths(self.config["outdir"])
# Writers
self.train_writer = tf.summary.create_file_writer(
os.path.join(config["outdir"], "tensorboard", "train"))
self.eval_writer = tf.summary.create_file_writer(
os.path.join(config["outdir"], "tensorboard", "eval"))
def load_yaml(path):
with open(preprocess_paths(path), "r", encoding="utf-8") as file:
return yaml.load(file, Loader=yaml.FullLoader)
def __init__(self, config: dict, show=False):
"""
config = {
"vocabulary": str,
"blank_at_zero": bool,
"beam_width": int,
"lm_config": {
...
}
}
"""
self.config = config
self.normlizer = NSWNormalizer
self.config["vocabulary"] = preprocess_paths(self.config["vocabulary"])
self.config["spker"] = preprocess_paths(self.config["spker"])
self.config["maplist"] = preprocess_paths(self.config["maplist"])
with open(self.config['spker']) as f:
spks = f.readlines()
self.spker_map = {}
for idx, spk in enumerate(spks):
self.spker_map[spk.strip()] = idx
with open(self.config["maplist"], encoding='utf-8') as f:
data = f.readlines()
self.map_dict = {}
for line in data:
try:
a, b = line.strip().split('\t')
except:
content = line.split(' ')
a = content[0]
b = ' '.join(content[1:])
a = a.replace('[', '').replace(']', '')
b = b.split(' ')
self.map_dict[a] = b
self.num_classes = 0
lines = []
with codecs.open(self.config["vocabulary"], "r", "utf-8") as fin:
lines.extend(fin.readlines())
if show:
logging.info('load token at {}'.format(self.config['vocabulary']))
self.token_to_index = {}
self.index_to_token = {}
self.vocab_array = []
self.tf_vocab_array = tf.constant([], dtype=tf.string)
self.index_to_unicode_points = tf.constant([], dtype=tf.int32)
index = 0
if self.config["blank_at_zero"]:
self.blank = 0
index = 1
self.tf_vocab_array = tf.concat([self.tf_vocab_array, [""]],
axis=0)
self.index_to_unicode_points = tf.concat(
[self.index_to_unicode_points, [0]], axis=0)
for line in lines:
line = line.strip() # Strip the '\n' char
if line.startswith("#") or not line or line == "\n":
continue
self.token_to_index[line] = index
self.index_to_token[index] = line
self.vocab_array.append(line)
self.tf_vocab_array = tf.concat([self.tf_vocab_array, [line]],
axis=0)
upoint = tf.strings.unicode_decode(line, "UTF-8")
self.index_to_unicode_points = tf.concat(
[self.index_to_unicode_points, upoint], axis=0)
index += 1
self.num_classes = index
if not self.config["blank_at_zero"]:
self.blank = index
self.num_classes += 1
self.tf_vocab_array = tf.concat([self.tf_vocab_array, [""]],
axis=0)
self.index_to_unicode_points = tf.concat(
[self.index_to_unicode_points, [0]], axis=0)
self.stop = self.endid()
self.pad = self.blank
self.stop = -1
def __init__(self, decoder_config: dict, show=False):
"""
decoder_config = {
"vocabulary": str,
"blank_at_zero": bool,
"beam_width": int,
"lm_config": {
...
}
}
"""
self.decoder_config = decoder_config
self.decoder_config["vocabulary"] = preprocess_paths(
self.decoder_config["vocabulary"])
self.scorer = None
self.num_classes = 0
lines = []
with codecs.open(self.decoder_config["vocabulary"], "r",
"utf-8") as fin:
lines.extend(fin.readlines())
if show:
print('load token at {}'.format(self.decoder_config['vocabulary']))
self.token_to_index = {}
self.index_to_token = {}
self.vocab_array = []
self.tf_vocab_array = tf.constant([], dtype=tf.string)
self.index_to_unicode_points = tf.constant([], dtype=tf.int32)
index = 0
if self.decoder_config["blank_at_zero"]:
self.blank = 0
index = 1
self.tf_vocab_array = tf.concat([self.tf_vocab_array, [""]],
axis=0)
self.index_to_unicode_points = tf.concat(
[self.index_to_unicode_points, [0]], axis=0)
for line in lines:
line = line.strip() # Strip the '\n' char
# Skip comment line, empty line
if line.startswith("#") or not line or line == "\n":
continue
self.token_to_index[line] = index
self.index_to_token[index] = line
self.vocab_array.append(line)
self.tf_vocab_array = tf.concat([self.tf_vocab_array, [line]],
axis=0)
upoint = tf.strings.unicode_decode(line, "UTF-8")
self.index_to_unicode_points = tf.concat(
[self.index_to_unicode_points, upoint], axis=0)
index += 1
self.num_classes = index
if not self.decoder_config["blank_at_zero"]:
self.blank = index
self.num_classes += 1
self.tf_vocab_array = tf.concat([self.tf_vocab_array, [""]],
axis=0)
self.index_to_unicode_points = tf.concat(
[self.index_to_unicode_points, [0]], axis=0)
if self.decoder_config['model_type'] == 'Transducer':
self.stop = self.endid()
self.pad = self.blank
self.start = self.startid()
elif self.decoder_config['model_type'] == 'LAS':
self.stop = self.endid()
self.pad = 0
self.start = self.startid()
elif self.decoder_config['model_type'] == 'LM':
self.stop = self.endid()
self.pad = 0
self.start = self.startid()
else:
self.pad = 0
self.stop = -1
