Aku Rouhe
Aku Rouhe
The items' keys are meant to be unique. You should not both take and provide the same key. It might be cleanest to just change the key of your DynamicItem...
Hey, should we implement the top-k hypothesis output (as discussed in PR #761) in this PR instead?
I'm ok with using `(scorer_func, weight)` -tuples too, but maybe I vote for separate arguments (Mirco's second suggestion). Tuples would force specifying weights for everything, but I think it could...
I also like the idea of always returning with the top-k dimension: > topk_tokens, topk_lens, scores = self.hparams.test_search(x, wav_lens) but this does changes in a _lot_ of files, including files...
Alright, yeah this does make sense. We need to just communicate this clearly.
This was added intentionally. Instead, the file methods in the Pretrained interfaces should take the savedir option.
Perhaps we could instead finally change the fetch behaviour. This has been requested for other reasons, as well, e.g. issue #1055. We could instead do the following: 1. In fetch,...
I would be happy to remove the magic audio fetching in `load_audio` altogether. I think the convenience is only there for example code - but not really in actual usage....
I'm not familiar with how that recipe is supposed to work but it seems something is broken here, other issues #1300 (also yours), and #1285, #954. I think PR #1223...
Actually, now that I think about it, it's a recipe, so pip installed version does not matter here, and there's no hotfix to make.