When using your project with just a few modification in my lisence plate recognition work, the CTC loss always gave nan.

[jiayi-wei]

Could you give me any suggestion?

[airaria]

If you are using your own data, you may check if the value of x is in the right range [-1,1] ?
and if y starts with a SOS symbol(for example, '^'), ends with a EOS symbol('$') and paddings (' ') ?
The value of SOS symbol in the dictionary (vocab['token2id']) should be 0 ( i.e. , vocab['token2id']['^']=0),
and besides,
vocab['token2id']['$']= len(vocab['token2id'])-1,
vocab['token2id'][' ']=len(vocab['token2id'])-2

If your data is OK, I suggest you to follow and print the variables step by step, and check the values of the variables to see if any anomalies happen before the CTC loss.