I see you using open source dataset and label pinyin for every dataset, but we all know label pinyin has muti-tone issue. So how can I reduce error when labeling pinyin for other sentences.

ctc decode阶段，会不会出现内存泄漏的问题，网上有人说r1 = tfk.get_value(r[0][0])这个存在内存泄漏，有没有遇到相似问题？求解，多谢！

RT

您好,在v1 版本 加载模型时:cnn卷基层结束,在进入fsmn 报错TypeError: 'NoneType' object is not subscriptable

初步定为在CFSMN类的 call方法中的tf.tensordot(inputs, self._V, axes=1) + self._bias_V函数,能力有限没能解决,还请指教

报错位置:

x.shape=(none, none, none)

output.shape = (1/2, 1/2, 1/2)

def cnn_cell (size, x, pool=True):
x = norm(conv2d(size)(x))
x = norm(conv2d(size)(x))
if pool:
x = maxpool(x)
return x

load your trained model file,but not right

如何修改得到汉字的结果呢

如题，经过代码分析，默认情况下
self.batch_size =self.q_size；
gpu_num=1；
sample_nums = self.batch_size * gpu_num；

而遍历循环代码for j in range(self.q_size // sample_nums)中：
self.q_size // sample_nums 不就等价于1//gpu_num么？那这个循环j始终为range(1)啊,

求解是不是应该改成 sample_nums = self.batch_size // gpu_num？

附上代码片段：
‘‘’
“#” 开始迭代，使用Adam优化的随机梯度下降法，并将结果输出到日志文件
print('training------')
index_list = [i for i in range(self.q_size)]
sample_nums = self.batch_size * gpu_num
for i in range(self.epoch_steps):
total_loss = 0
total_accuracy = 0
for j in range(self.q_size // sample_nums): # 默认情况：batch_size == q_size
if self.is_sample:
sample_list = simple_sampling(index_list, sample_nums)
q_set = []
t_set = []
q_actual_length = []
t_actual_length = []
for index in sample_list:
q_set.append(self.q_set[index])
t_set.append(self.t_set[index])
q_actual_length.append(self.q_actual_length[index])
t_actual_length.append(self.t_actual_length[index])
‘’’