Closed d223302 closed 4 years ago
Hi @d223302 ,
You're right, multi-head attention implementation wasn't correct in the last version, We've update our code recently, I believe it's correct now. If you're interested, please have a look at line 138-235 in src/module.py
Thanks a lot for appreciating our work!
I think there is an no attribute error in /src/asr.py line 437 : self.preprocess_mlp_dim. Should it be self.num_head? Thanks for your great and clear implementation.