CharizardAcademy / convtransformer

Code for the ACL2020 paper Character-Level Translation with Self-Attention
32 stars 7 forks source link

Several training issues in the current version #1

Closed jungokasai closed 3 years ago

jungokasai commented 4 years ago

I'm running into several issues in running the current version. Could you confirm that you can run training without an error? It looks like that the current version is missing options and has a syntax error. This doesn't seem like an issue with my environment. First, I found a syntax error https://github.com/CharizardAcademy/convtransformer/blob/master/convtransformer/fairseq/models/transformer.py#L941

Even after fixing this, I got the following error with an options:

    result = self.forward(*input, **kwargs)
  File "/private/home/jkasai/projects/convtransformer/convtransformer/fairseq/models/transformer.py", line 1343, in forward
    if self.enable_pooling:
  File "/private/home/jkasai/.conda/envs/my_pytorch_env/lib/python3.7/site-packages/torch/nn/modules/module.py", line 576, in __getattr__
    type(self).__name__, name))
AttributeError: 'ConvTransformerEncoderLayer' object has no attribute 'enable_pooling'`
CharizardAcademy commented 4 years ago

thanks for your feedback. We have removed some errors in the code and updated onto the repo. Pls try our latest version

jungokasai commented 4 years ago

Thank you! I was able to run it.

CharizardAcademy commented 3 years ago

issue closed