Fixes to BahdanauAttnDecoderRNN

spro / practical-pytorch

Go to https://github.com/pytorch/tutorials - this repo is deprecated and no longer maintained

MIT License

4.52k stars 1.11k forks source link

Open anantzoid opened 5 years ago

anantzoid commented 5 years ago

This PR fixes all the issue with BahdanauAttnDecoderRNN class and makes the NMT trainable with it as well. The changes consist of the following:

Syntax fixes (wrong class in super),
Removed unused variable max_length,
Handled dot product for batch-size=1 using bmm in Attn class for 'dot' and 'general'
Tested both the decoders under the section Testing the models