Xi Ouyang

I also meet this issue. Have you solve it?

I refer to the torch project of their paper, I think it is same as the network they define in torch.

You should install the warpctc first. Please check this issue: https://github.com/oyxhust/CNN-LSTM-CTC-text-recognition/issues/4.

A simple way is to calculate the pixel-wise differences between the GT images and generated images. The normal images are supposed to have the smaller MSE difference.

It won't affect the speed of training. I set the max length for the "num_label", because this can reduce the error of output. When the length of prediction is longer...

I think you may not compile mxnet python when compiling mxnet. Or it can be found, I didn't have such problem.

I think you should add /path/to/mxnet/amalgamation/python in your environment.