RecurrentNeuralNetwork | Recurrent Neural Network from scratch using Python and Numpy | Machine Learning library

ok, I wasn't able to explain this with the 600 chars available in the comment section, therefore I will do it by answering, however ignoring your Q2.

The code to the paper you mentioned can be found at: https://github.com/bgshih/crnn It is a good starting point for handwritten text recognition. However, the CRNN implementation recognizes text on word-level, you want to do it on line-level, therefore you need larger input images, e.g. I used 800x64px and a maximum text length of 100. And as already said, stretching images to the desired size does not work very well, in my experiments the accuracy increased when using padding (randomize positions a little bit ... it's an easy way to do data augmentation).

There is a relationship between the maximum text length L and the input image width W: the Neural Network (NN) downsizes the input image by a fixed scaling factor f: L=W/f (in my example: W=800px, L=100, f=8). The illustrations attached shows the input image (800x64px) and the character probability matrix (probability for each of the 80 possible characters for each of the 100 time-steps). The NN maps the input image to this character probability matrix which serves as input for the CTC. As there are L many time-steps in the matrix, there can be at most L many characters: this of course holds for decoding, but also loss calculation must align the ground truth text somehow with this matrix, and how should a text with L+1 characters be aligned with just L time-steps contained in the matrix!? Note that inside the CTC calculation repeated characters (like in "piZZa") must be separated by a special character - therefore the possible text length decreases by 1 for each repetition.

I think with this explanation you should be able to figure out how all those length-variables in your code are related to each other.