Created by: F0REacH

I've tried to compare warp-ctc torch7 bindings with Paddle CTC -> test repo see logs Warp-ctc torch implementation - works just fine. Paddle with the same data and the same network structure can't successfully minimize the training error.

I am unable to isolate the problem by changing hyperparameters, network type/size, optim. functions, switching between CPU/GPU. It seems the only solution that works is to truncate sequences to ~500 timesteps.

Any possible solutions/suggestions how to make this work with large varying length sequences (3000-5000 timesteps each)?

Using latest master as:

cmake -DWITH_GPU=ON -DWITH_DOC=OFF -DCMAKE_INSTALL_PREFIX=/opt/paddle ..