Attention-based models do not work

[hirofumi0810]

I'm refactoring attention-based models, so they do not work now!
Please wait for a few days.

[wolverineq]

Good to hear! and hope a multi-gpu training version will be added since training of attention-based model is quite time-consuming...

[hirofumi0810]

now finished (greedy decoding)

TODO:

• beam search
• multi GPUs version

[CyanHillFox]

@hirofumi0810 @hlthu
Does it work now? It seems that the preprocessed data created by this repo doesn't meet the requirement.

I tried the timit experiment with attention-based model, but it seems that the structure of input data doesn't match. For example, in timit/data/load_dataset_attention.py, it expects that label and input data are in "label" and "input" folder, but in the preprocessed data, they are in "dataset" and "feature". And that the labels are in .csv file rather than .npy file.
Maybe there should be an update?

Sincerely.