What does 'check length of chars and wordChars: 0<len(wordChars)<=len(chars)' mean?

[hasansalimkanmaz]

If you create a new issue, please provide the following information:

1. Which program causes the problem

   • Custom TF operation

2. Versions

   • TensorFlow version: Version: 1.15.0
   • Python version: 3.7
   • C++ compiler: gcc (Ubuntu 5.5.0-12ubuntu1~16.04) 5.5.0 20171010
   • Operating system: Linux

3. Issue I couldn't understand this error message in your code. I got this error during evalution of 1st epoch in the training.

   
   Validate NN
   terminate called after throwing an instance of 'std::invalid_argument'
   what():  check length of chars and wordChars: 0<len(wordChars)<=len(chars)```

My task is to recognize handwritten dates, such as "19/10/1993" it only contains slashes and numbers. Does it result from this?

How can I solve the issue?

[githubharald]

Hi,

can you show what you pass as chars and wordChars to the algorithm?

Just as a reminder:

• chars: all the characters that can be recognized by the neural network
• wordChars: when parsing the corpus, words are strings containing wordChars

In your case (just guessing) this could look like this:

• chars: "/0123456789"
• wordChars: "0123456789"
• and your corpus should contain all allowed numbers like "1 2 3 [...] 30 31 1990 1991 [...] 2020 [...]"

So, wordChars is a subset of chars and therefore, there can't be more characters in wordChars than chars, therefore 0<len(wordChars)<=len(chars) must hold.

One more thing: if possible, don't use word beam search while training, as it is slower than best path decoding. I would train using best path decoding, and if best path decoding gives good results, than word beam search should also give good results.

[hasansalimkanmaz]

Thanks for crystal clear explanation.

It worked and improved my accuracy.