def dfsmn_model_train(train_data_path, label_data_path):
# 0.准备训练所需数据------------------------------
data_args = data_hparams()
data_args.data_type = 'train'
data_args.data_path = train_data_path
data_args.label_data_path = label_data_path
data_args.thchs30 = True
data_args.aishell = True
data_args.batch_size = 16
# data_args.data_length = 100
data_args.data_length = None
data_args.shuffle = True
train_data = get_data(data_args)
# 0.准备验证所需数据------------------------------
data_args = data_hparams()
data_args.data_type = 'dev'
data_args.data_path = train_data_path
data_args.label_data_path = label_data_path
data_args.thchs30 = True
data_args.aishell = True
data_args.batch_size = 16
# data_args.data_length = 16
data_args.data_length = None
data_args.shuffle = True
dev_data = get_data(data_args)
# 1.声学模型训练-----------------------------------
from .Model.cnn_dfsmn_ctc import Am, am_hparams
am_args = am_hparams()
am_args.vocab_size = len(train_data.am_vocab)
am_args.gpu_nums = 1
am_args.lr = 0.001
am_args.is_training = True
am = Am(am_args)
if os.path.exists('ModelMemory/cnn_dfsmn_ctc_model.h5'):
print('load acoustic model...')
am.ctc_model.load_weights('ModelMemory/cnn_dfsmn_ctc_model.h5')
batch_num = len(train_data.wav_lst) // train_data.batch_size
# checkpoint
# ckpt = "model_{epoch:02d}.hdf5"
# checkpoint = ModelCheckpoint(os.path.join('./logs_am/checkpoint', ckpt), monitor='val_loss', save_weights_only=False, verbose=1, save_best_only=True)
batch = train_data.get_am_batch()
dev_batch = dev_data.get_am_batch()
epochs = 150
# am.ctc_model.fit_generator(batch, steps_per_epoch=batch_num, epochs=epochs, callbacks=[checkpoint], workers=1, use_multiprocessing=False, validation_data=dev_batch, validation_steps=200)
am.ctc_model.fit_generator(batch,
steps_per_epoch=batch_num,
epochs=epochs,
workers=1,
use_multiprocessing=False,
validation_data=dev_batch,
validation_steps=500)
am.ctc_model.save_weights('ModelMemory/cnn_dfsmn_ctc_model.h5')
def dfsmn_model_decode (wav_file_path):
# 1.语料加载-----------------------------------
print('loading Lang...')
pinyin_list = []
with open('Lang/pinyin.txt', 'r', encoding='utf-8') as file_object:
for pinyin in tqdm(file_object):
pinyin_list.append(pinyin.rstrip('\n'))
pinyin_list.append('_')
# 2.声学模型加载-----------------------------------
print('loading acoustic model...')
am_args = am_hparams()
am_args.vocab_size = len(pinyin_list)
am_args.is_training = False
am = Am(am_args)
am.ctc_model.load_weights('ModelMemory/cnn_dfsmn_ctc_model.h5')
# 3. 启动在线识别-------------------------------------------
print('start online decode...')
x = get_online_data(wav_file_path)
pinyin_id = am.model.predict(x, steps=1)
_, pinyin = decode_ctc(pinyin_id, pinyin_list)