Question about CTC Loss

nl8590687 / ASRT_SpeechRecognition

A Deep-Learning-Based Chinese Speech Recognition System 基于深度学习的中文语音识别系统

https://asrt.ailemon.net

GNU General Public License v3.0

7.71k stars 1.89k forks source link

Question about CTC Loss #173

Open pierre-zhao opened 4 years ago

pierre-zhao commented 4 years ago

Hi, I used my collected data to train the model, but unfortunately, there is an error during the training process. Here is the log information. bug Can u give me some suggestions>_>

pierre-zhao commented 4 years ago

I guess the problem is caused by the error between input_length and label_length. Why you divide 8 in your code for calculating input_length (203 line in readdate24.py)?

nl8590687 commented 4 years ago

Yes, Because the size of the data stream is 1/8 of the input layer according to Pooling Layer(2*2) .

pierre-zhao commented 4 years ago

Get it! I still try to solve the CTC loss error above.

pierre-zhao commented 4 years ago

Can I add a fixed number to input_length to insure that input_length > label_length?

nl8590687 commented 4 years ago

Maybe can, I think.

pierre-zhao commented 4 years ago

I follow your annotation (没有余数，只留整除。有余数就整除再+1). But another error appears

nl8590687 commented 4 years ago

This is because of a problem with your tensorflow installation. your python is version 3.6 but your tensorflow is for python 3.5.

pierre-zhao commented 4 years ago

Maybe. I will try it again. Thx.