search

HawkAaron / warp-transducer

A fast parallel implementation of RNN Transducer.
Apache License 2.0
306 stars 124 forks source link

RNNT loss is constantly zero #92

Open Rajratnpranesh opened 2 years ago

Rajratnpranesh commented 2 years ago

Hey

I am trying to train the model but the loss is constant = '0'. I found that the loss output from the PyTorch binding of WarpRNNT is a zero tensor. I am getting an error of 'src/binding.cpp:151: unsupported data type'. Is this is the reason why WarpRNNT is not able to calculate loss (I charged the datatype from float16 to 32 as sum() operation was not possible with float16 then after summing back to float16). I appreciate any pointer.

Ty

ccc0117 commented 2 years ago

I have same issue. Can someone help?