Open feitiandemiaomi opened 4 years ago
If you use ctcloss, I advice you to check the version of pytorch, 1.2.0 is ok, other versions are not tested. Otherwise, maybe you should check whether a nan existed in your tensor before computing loss.
Yes,I use ctcloss , My verison is pytorch1.1.0 ,Just now ,I trained it with Pytorch1.2.0 ,the output are as follows: There was a problem from the beginning,Did you do something else, or could you provide the code?
you can try this ctcloss implementation https://github.com/SeanNaren/Warp-ctc
@feitiandemiaomi
Thanks for great repo , The model can be trained normally, but when I replace the lstm layer with a convolutional layer or linear layer ,the model can't converge, and the loss decreased from 270 to 4. I try to modify lr from 0.01 to 0.0000001, it also dosen't work, can you give some advice? @Sierkinhane