Is there a way to do back-propagation through time using CTC loss?
I’ve seen implementations of BPTT where you partition your outputs and labels into chunks, calculate loss while retaining graphs. With CTC, partitioning labels into equal chunks and assigning each one to an output block is ill-defined since by design, the labels may be misaligned.
Is there a way to make BPTT or MRBP (https://arxiv.org/pdf/2010.06891.pdf) work with CTC loss?
Thank you