def main():
args = input_args()
if args.mode == M_PREDICTION:
if not check_input_args(args):
return False
audio = args.audio
acoustic = Acoustic(CONFIG_PATH, operation=M_PREDICTION)
if args.predict_mode == M_PREDICT_GREEDY:
output = acoustic.run(audio=audio, greedy=True)
print(Color.red("input: {}".format(audio)))
print(Color.yellow("Greedy predict: {}".format(output)))
elif args.predict_mode == M_PREDICT_BEAM:
output = acoustic.run(audio=audio,
beam_width=args.beam,
top_paths=args.top,
greedy=False)
print(Color.yellow("input: {}".format(audio)))
for i in range(args.top):
print(Color.yellow("Beam predict: {}.{}".format(i, output[i])))
print()
else:
acoustic = Acoustic(CONFIG_PATH, operation=args.mode)
print(acoustic.model.summary())
acoustic.run()
def writer_file(file, results, mode='a', encoding='utf-8'):
"""
输入字符串、列表字符，写入文件
:param file: (str, mandatory) 文件名 或 路径 + 文件名
:param results: (str or list or tuple, mandatory) 需要写入文件的字符集或单个字符串。例如"hello word" 或['hello world']
:param mode: (str, optional, default='a') 模式。默认为 'a' ，追加模式
:param encoding: (str, optional, default='utf-8') 编码。默认为 UTF-8 编码
:return:
"""
if not isinstance(results, (str, list, tuple)):
raise ParameterError(
"result parameter must be is {}, but actually get {}".format((str, list, tuple), type(results)))
if isinstance(results, (list, tuple)):
for result in results:
if not isinstance(result, str):
raise ParameterError(
"results parameter elements must be is str, but actually get {}, elements:{}".format(
type(result), result))
Writer.check_path(file)
with open(file, mode, encoding=encoding) as f_write:
if isinstance(results, str):
f_write.writelines(results + '\n')
else:
for result in results:
f_write.writelines(result + '\n')
print("\n" + "over！File: {}, encoding: {}".format(Color.red(file), Color.red(encoding)))
def check_input_args(args):
""" 检查输入的参数 """
if args.audio is '':
print(Color.red("Please input audio file !"))
return False
else:
if not os.path.isfile(args.audio):
print(FileNotFoundException(args.audio))
return False
if args.top > args.beam:
print(Color.red("requested top:{} than the beam:{}".format(args.top, args.beam)))
return False
return True
def init_transverse(conf, input_dim, output_vocab_size):
"""
初始化 Transverse
:param conf: (obj, mandatory) 配置对象
:param input_dim: (int, mandatory) 输入维度
:param output_vocab_size: (int, mandatory) 输出词汇量大小
:return: (AcousticTransverseNet) AcousticTransverseNet 对象
"""
transverse = AcousticTransverseNet(
input_dim=input_dim,
output_vocab_size=output_vocab_size,
dn_hidden_size=conf.MODEL.TRANSVERSE.dn_hidden_size,
dilated_conv_depth=conf.MODEL.TRANSVERSE.dilated_conv_depth,
width_conv_depth=conf.MODEL.TRANSVERSE.width_conv_depth,
multi_dilated_rate=conf.MODEL.TRANSVERSE.multi_dilated_rate,
dilated_conv_filters=conf.MODEL.TRANSVERSE.dilated_conv_filters,
width_conv_filters=conf.MODEL.TRANSVERSE.width_conv_filters,
dropout_rate=conf.MODEL.TRANSVERSE.dropout_rate,
l1=conf.MODEL.TRANSVERSE.l1,
l2=conf.MODEL.TRANSVERSE.l2,
activation=conf.MODEL.TRANSVERSE.activation,
learning_rate=conf.MODEL.TRANSVERSE.learning_rate,
warmup_steps=conf.MODEL.TRANSVERSE.warmup_steps,
optimizer_beta_1=conf.MODEL.TRANSVERSE.optimizer_beta_1,
optimizer_beta_2=conf.MODEL.TRANSVERSE.optimizer_beta_2,
optimizer_epsilon=conf.MODEL.TRANSVERSE.optimizer_epsilon,
ckpt_dir=conf.MODEL.TRANSVERSE.ckpt_dir,
ckpt_max_to_keep=conf.MODEL.TRANSVERSE.ckpt_max_to_keep)
print(Color.green('Initialization transverse from scratch'))
return transverse
def init_or_restore_audio_feature(dataset, file, f_type, frame_length,
frame_shift, mfcc_dim):
"""
初始化或还原音频特征器
:param dataset: (list, mandatory) 数据集
:param file: (str, mandatory) 保存或者还原audio_feature的文件
:param f_type: (str, mandatory) 处理音频特征的类型。"spectrogram" or "mfcc"
:param frame_length: (int, mandatory) 帧长
:param frame_shift: (int, mandatory) 帧移
:param mfcc_dim: (int, mandatory) mfcc 特征维度
:return: (AudioFeature) 音频特征对象
"""
if os.path.isfile(file):
audio_feat = AudioFeatures.load(file)
output = "Audio feature: restore from : {}".format(file)
else:
audio_feat = AudioFeatures(f_type=f_type,
frame_length=frame_length,
frame_shift=frame_shift,
mfcc_dim=mfcc_dim)
audio_feat.fit(train_data=dataset)
output = "Audio feature: Not found file:{} Initializing from scratch".format(
file)
audio_feat.save(file)
print(Color.green(output))
return audio_feat
def init_transformer(input_vocab_size, output_vocab_size, padding_index, conf):
"""
初始化 transformer
:param input_vocab_size: (int, mandatory) 输入词汇大小
:param output_vocab_size: (int, mandatory) 输出词汇大小
:param padding_index: (int, mandatory) 填充字符索引
:param conf: (object) 配置信息
:return: (LanguageTransformer) transformer 对象
"""
transformer = LanguageTransformer(
input_vocab_size=input_vocab_size,
output_vocab_size=output_vocab_size,
d_model=conf.MODEL.TRANSFORMER.d_model,
heads_num=conf.MODEL.TRANSFORMER.heads_num,
forward_hidden=conf.MODEL.TRANSFORMER.forward_hidden,
num_layers=conf.MODEL.TRANSFORMER.num_layers,
input_max_positional=conf.MODEL.TRANSFORMER.input_max_positional,
target_max_positional=conf.MODEL.TRANSFORMER.target_max_positional,
dropout_rate=conf.MODEL.TRANSFORMER.dropout_rate,
ckpt_dir=conf.MODEL.TRANSFORMER.ckpt_dir,
ckpt_max_to_keep=conf.MODEL.TRANSFORMER.ckpt_max_to_keep,
lr_warmup_steps=conf.MODEL.TRANSFORMER.lr_warmup_steps,
optimizer_beta_1=conf.MODEL.TRANSFORMER.optimizer_beta_1,
optimizer_beta_2=conf.MODEL.TRANSFORMER.optimizer_beta_2,
optimizer_epsilon=conf.MODEL.TRANSFORMER.optimizer_epsilon,
padding_index=padding_index,
pred_max_length=conf.MODEL.TRANSFORMER.pred_max_length)
print(Color.green('Initialization transformer from scratch'))
return transformer
def check_input_sentence(sent):
"""
检查输入的pinyin句子是否满足要求
:param sent: (str, mandatory) 拼音句子。例如 'lv4 shi4 yang2'
:return: (bool) True or False
"""
if sent is None:
print(Color.red("Please input the prediction sentence !"))
return False
sent = regular(sent)
if len(sent) == 0:
print(
Color.red(
"Please input the prediction sentence ! sentence cannot be empty!"
))
return False
return True
def init_or_restore_chinese_dict(dataset, file):
"""
初始化或还原中文字典
:param dataset: (list, mandatory) 数据集
:param file: (str, mandatory) 保存或者还原chinese的文件
:return: (WordDict) 字典对象
"""
if os.path.isfile(file):
chinese = WordDict.load(file)
output = 'Restore chinese dict from file: {}'.format(file)
else:
chinese = WordDict()
chinese.fit(dataset=dataset)
chinese.save(file)
output = 'Initialize chinese dict from dataset'
print(Color.green(output))
return chinese
def init_or_restore_pinyin_dict(dataset, file):
"""
初始化或还原拼音字典
:param dataset: (list, mandatory) 数据集
:param file: (str, mandatory) 保存或者还原pinyin字典的文件
:return: (WordDict) 字典对象
"""
if os.path.isfile(file):
pinyin = PinYin.load(file)
output = 'Restore pinyin dict from file:{}'.format(file)
else:
pinyin = PinYin()
pinyin.fit(dataset=dataset)
pinyin.save(file)
output = 'Initialize pinyin dict form dataset'
print(Color.green(output))
return pinyin
def restore_transformer(file, ckpt_dir):
"""
还原 transformer
:param file: (str, mandatory) transformer 配置文件
:param ckpt_dir: (str, mandatory) 检查点保存目录
:return: (LanguageTransformer) transformer 对象
"""
try:
language_transformer = LanguageTransformer.restore(file, ckpt_dir)
except Exception as e:
print(
Color.red(
"restore transformer model failed！config file: {}, checkpoint: {}"
.format(file, ckpt_dir)))
print(e)
return False
return language_transformer
def load_dataset(source_data_path, data_path):
"""
加载数据集
:param data_path: (str, mandatory) 训练数据所在的路径
:param source_data_path: (str, mandatory) 实际数据路径
:return: (dict of list) 列表字典。返回音频文件路径，标签文件路径，标签数据路径，拼音字符，汉字
"""
pinyin_sent_data = list()
chinese_sent_data = list()
chinese_data = list()
pinyin_data = list()
data_file = os.listdir(data_path)
for file in tqdm(data_file):
# 检查文件名后缀是不是 wav 音频类型文件。
if file[-3:] == 'wav':
# 标签文件。根据音频文件名得到标签文件名
label_file = file + '.trn'
# 判断标签文件是否存在。不存在则报错
assert label_file in data_file, FileNotFoundException(label_file)
# 音频文件路径和标签文件路径
label_file_path = os.path.join(data_path, label_file)
# 读取标签文件内容。找到对应标签数据路径
label_data_file_path = DataUtils.read_text_data(
label_file_path, show_progress_bar=False)
assert len(label_data_file_path) == 1, \
'Get redundant data path: {}, label_file_path:{}'.format(label_data_file_path, label_file_path)
# 重新拼接路径
label_data_file_path = os.path.join(
source_data_path, label_data_file_path[0].split('/')[-1])
assert os.path.isfile(label_data_file_path), FileNotFoundException(
label_data_file_path)
# 读取标签数据。包括拼音、文本
text_data = DataUtils.read_text_data(label_data_file_path,
handle_func=handle,
show_progress_bar=False)
chinese = text_data[0]
pinyin = text_data[1]
# 检查是否存在重复的数据，如果存在则跳过
chinese_sent = ''.join(chinese)
pinyin_sent = ' '.join(pinyin)
if chinese_sent in chinese_sent_data and pinyin_sent in pinyin_sent_data:
# print("repeat sent: {}, {}".format(chinese_sent, pinyin_sent))
continue
else:
chinese_sent_data.append(chinese_sent)
pinyin_sent_data.append(pinyin_sent)
# 检查pinyin字符是否正确。主要判断每个拼音的最后一个字符是不是数字。不是数字则报错
for py in pinyin:
assert py[-1].isdigit(), "the last character:{} of Pinyin is not a number! " \
"pinyin_str:{}, " \
"label_data_file_path:{}, " \
"label_file_path:{}".format(py, pinyin, label_data_file_path,
label_file_path)
# 将由多个中文字符组成的词转换成单个字
new_chinese = list()
for ch in chinese:
new_chinese += list(ch)
chinese = new_chinese
# 检查是否是一个拼音对应中文字符， 如果不是则报错
assert len(chinese) == len(pinyin), "the number of pinyin:{} and chinese:{} is different, " \
"chinese:{}, pinyin:{}, file:{}".format(len(pinyin), len(chinese),
chinese, pinyin,
label_data_file_path)
chinese_data.append(chinese)
pinyin_data.append(pinyin)
print(Color.red("load to {} of data!".format(len(chinese_data))))
return chinese_data, pinyin_data