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 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