Inference again

[facebookresearch]

Inference again

[glample]

Because the target is just used to evaluate the BLEU score. But if you are just interested in inference, this is enough to do this. It's clearly not very efficient, but it's hacky workaround that should work until I add a translate.py file for inference.

Yes, best-bleu_xxx.pth only contains the model, this is what you need to use. checkpoint.pth contains the last model in the training experiment (not necessarily the best), along with the optimizer, the training parameters, etc.

[glample]

Hi,

Yes, the above way seems correct. If you don't have parallel data, you can simply run the same thing and use an arbitrary random text file with the same number of lines for the target file to create a fake parallel data.
Also, note that the checkpoint.pth is to restart the experiment, it contains a lot of things like training epoch, best scores, etc. In your case you want to reload a file like best-bleu_en_fr_valid.pth

[shawnzaru]

Thanks for reply. I have been a bit hesitant to use fake target data as the training is done using back-translation (A -> B -> A and B -> A -> B). Could you explain to me why using fake target data for inference doesn't matter? Thanks.

[shawnzaru]

Also, I am trying to understand the difference between checkpoint.pth and best-bleu_xxx.pth. Do you mean checkpoint.pth is more redundant than best-bleu_xxx.pth or is it that in the inference case I have to reload best-bleu_xxx.pth instead of checkpoint.pth, because loading checkpoint.pth will make the program do things differently? Thanks.

[shawnzaru]

Thanks 👍