cuda issue

[DivyaWadehra]

will it not work without cuda, on cpu?

[DivyaWadehra]

AttributeError: module 'config' has no attribute 'max_seq_len'

[seanie12]

It will work both on cpu and cuda.
Maybe there is typo. Change config.max_seq_len into config.max_len

[DivyaWadehra]

It is not working on cpu since there is a statement which is gpu specificl "line 20 trainer.py"

[DivyaWadehra]

TypeError: expected Byte (got Bool)

[seanie12]

@DivyaWadehra
If you want to run the model on cpu, change it to device = "cpu"
For the type error, check this link

[DivyaWadehra]

segmentation fault while training and the torch_scatter is installed

[seanie12]

Please describe your python and pytorch version.

[pzxbjx]

My python version is 3.7 and torch version is 1.4.0, when i installed torch-scatter, i got an error "command 'gcc' failed with exit status 1", and i can't fix it , is there a substitude for torch-scatter?

I can train the model with use_pointer=False, the val loss didn't drop until it reached around 4.42, is it normal? What should the val loss be with pointer network?

Thank you!

[seanie12]

@pzxbjx Regarding the installation of torch-scatter, please refer to this repo.
The pointer network is an important component for quality question generation. There are several variants of how to copy words from source sequences. Try out this. You can efficiently implement the copy mechanism with scatter_add from native pytorch library.