def train(datagenes: list, load_model=None):
w, h = 1600, 200
dataset = VoiceDatasetList()
x_set, y_set = dataset.merge_load(datagenes)
pymap = PinyinMapper(sil_mode=-1)
vloader = VoiceLoader(x_set,
y_set,
batch_size=16,
n_mels=h,
feature_pad_len=w,
feature_dim=3,
cut_sub=32)
model_helper = DCBNN2D(pymap)
model_helper.compile(feature_shape=(w, h, 1),
label_max_string_length=32,
ms_output_size=1423)
if load_model is not None:
load_model = os.path.abspath(load_model)
model_helper.load(load_model)
model_helper.fit(vloader)
def train(datagenes: list, load_model=None):
w, h = 1600, 200
max_label_len = 64
dataset = VoiceDatasetList()
x_set, y_set = dataset.merge_load(datagenes)
pymap = PinyinMapper(sil_mode=-1)
vloader = VoiceLoader(
x_set,
y_set,
batch_size=16,
feature_pad_len=w,
n_mels=h,
max_label_len=max_label_len,
pymap=pymap,
melf=MelFeature5(),
all_train=False,
)
model_helper = WAVEM(pymap)
model_helper.compile(feature_shape=(w, h),
label_max_string_length=max_label_len,
ms_output_size=pymap.max_index + 1)
if load_model is not None:
load_model = os.path.abspath(load_model)
model_helper.load(load_model)
model_helper.fit(vloader, epoch=-1, save_step=100, use_ctc=True)
def train(datagenes: list, load_model=None):
w, h = 1600, 200
dataset = VoiceDatasetList()
x_set, y_set = dataset.merge_load(datagenes)
pymap = PinyinMapper(sil_mode=-1)
vloader = VoiceLoader(
x_set,
y_set,
batch_size=16,
n_mels=h,
feature_pad_len=w,
feature_dim=3,
pymap=pymap,
melf=MelFeature5(),
divide_feature_len=8,
)
model_helper = DCNN2D(pymap)
model_helper.compile(feature_shape=(w, h, 1),
ms_output_size=pymap.max_index +
1) # ctcloss 计算要求： index < num_class-1
if load_model is not None:
load_model = os.path.abspath(load_model)
model_helper.load(load_model)
model_helper.fit(vloader, epoch=-1, use_ctc=True)
def train(datagenes: list, load_model=None,**kwargs):
w, h = 1600, 200
max_label_len = 64
dataset = VoiceDatasetList()
x_set, y_set = dataset.merge_load(datagenes)
pymap = PinyinMapper(sil_mode=-1)
vloader = VoiceLoader(x_set, y_set,
batch_size=16,
feature_pad_len=w,
n_mels=h,
max_label_len=max_label_len,
pymap=pymap,
melf=MelFeature4(),
divide_feature_len=8,
all_train=False,
)
model_helper = DCBNN1D(pymap)
model_helper.compile(feature_shape=(w, h), label_max_string_length=max_label_len,
ms_output_size=pymap.max_index + 1)
if load_model is not None:
load_model = os.path.abspath(load_model)
model_helper.load(load_model)
epoch = kwargs.get("epoch",-1)
save_step = kwargs.get("save_step",1000)
model_helper.fit(vloader, epoch=epoch, save_step=save_step, use_ctc=True)
def err_count():
'''
根据音频统计错误数据集
:return:
'''
z200 = Z200(config.z200_datapath)
thchs = Thchs30(config.thu_datapath)
prime = Primewords(config.prime_datapath)
stcmd = ST_CMDS(config.stcmd_datapath)
aishell = AiShell(config.aishell_datapath)
lst = [z200, thchs, prime, stcmd, aishell]
from acoustic.ABCDNN import DCBNN1D
from util.reader import PinyinMapper,VoiceDatasetList,VoiceLoader
from feature.mel_feature import MelFeature5
w, h = 1600, 200
max_label_len = 64
pymap = PinyinMapper(sil_mode=-1)
model_helper = DCBNN1D(pymap)
model_helper.compile(feature_shape=(w, h), label_max_string_length=max_label_len,
ms_output_size=pymap.max_index + 1)
model_helper.load(os.path.join(config.model_dir,"cur_best_DCBNN1D_epoch_722_step_722000.h5"))
dataset = VoiceDatasetList()
x_set, y_set = dataset.merge_load(lst)
vloader = VoiceLoader(x_set, y_set,
batch_size=16,
feature_pad_len=w,
n_mels=h,
max_label_len=max_label_len,
pymap=pymap,
melf=MelFeature5(),
divide_feature_len=8,
all_train=True,
)
viter = vloader.create_iter(one_batch = True,return_word=True)
with open("data_err.txt","w",encoding="utf-8") as w:
for i,batch in enumerate(viter):
[_, ys, _, label_len], words = batch
py_true_b = pymap.batch_vector2pylist(ys,return_word_list=True,return_list=True)
py_pred_b = model_helper.predict(batch)
for py_true,py_pred,llen,word in zip(py_true_b,py_pred_b,label_len,words):
llen = llen[0]
# w.write(f"{pyt}")
py_true = py_true[:llen]
py_true = " ".join(py_true)
py_pred = " ".join(py_pred)
print(f"\r——{i*16}.",end="\0",flush=True)
# print(word.strip(),py_pred)
w.write(f"{word.strip()}\t{py_true}\t{py_pred}\n")
def train(datagene: list, load_model=None):
dataset = VoiceDatasetList()
_, y_set = dataset.merge_load(datagene, choose_x=False, choose_y=True)
max_label_len = 64
pinyin_map = PinyinMapper(sil_mode=0)
chs_map = ChsMapper()
tloader = TextLoader(y_set,
padding_length=max_label_len,
pinyin_map=pinyin_map,
cut_sub=16,
chs_map=chs_map)
model_helper = DCNN1D()
model_helper.compile(feature_shape=(max_label_len,
tloader.max_py_size),
ms_input_size=pinyin_map.max_index,
ms_output_size=chs_map.categores)
if load_model is not None:
model_helper.load(load_model)
model_helper.fit(tloader, -1)
stcmd = ST_CMDS(config.stcmd_datapath) # 据说还可以
thchs = Thchs30(
config.thu_datapath) # 同质性太高，不过好拟合，可以用来测试模型的效果，在这个数据上都没法得到比较好的结果的就没啥使用的必要了
prime = Primewords(config.prime_datapath)
aishell = AiShell(config.aishell_datapath) # 据说数据集很差，不用该数据训练
z200 = Z200(config.z200_datapath)
# datagenes = [thchs,stcmd,prime,aishell,z200]
datagenes = [thchs]
if __name__ == "__main__":
w, h = 1600, 200
max_label_len = 64
batch_size = 16
dataset = VoiceDatasetList()
x_set, y_set = dataset.merge_load(datagenes)
pymap = PinyinMapper(sil_mode=-1)
vloader = VoiceLoader(x_set,
y_set,
batch_size=16,
feature_pad_len=w,
n_mels=h,
max_label_len=max_label_len,
pymap=pymap,
divide_feature_len=8,
melf=MelFeature5(),
all_train=True)
viter = vloader.create_iter(one_batch=True)
for i, _ in enumerate(viter):
