[Wav2Vec2] Improve SpecAugment function by converting numpy based fun… #10494
Conversation
punitvara
force-pushed
the
feature
branch
2 times, most recently
from
53a39ae to
0af1ec9
Compare
| mask_idcs.append(np.unique(mask_idc[mask_idc < sz])) | ||
| mask_idc = torch.randperm(sz - min_len)[:num_mask] | ||
| mask_idc = torch.from_numpy( | ||
| np.asarray([mask_idc[j] + offset for j in range(len(mask_idc)) for offset in range(lengths[j])]) |
There was a problem hiding this comment.
Let's try to not use np here. If we use torch.from_numpy(...) - it's not GPU-friendly
There was a problem hiding this comment.
mask_idc = torch.tensor([mask_idc[j] + offset for j in range(len(mask_idc)) for offset in range(lengths[j])])
I just did this. Is this right ?
|
|
||
| min_len = min([len(m) for m in mask_idcs]) | ||
| min_len = torch.min(mask_idcs) | ||
| for i, mask_idc in enumerate(mask_idcs): |
There was a problem hiding this comment.
Let's try to get rid of the for-loop and do tensor operations only
There was a problem hiding this comment.
I am not sure how to do this. Something like following but not getting how should I put it ? Can you help here.
mask[i, mask_idc] = [True, torch.randperm(mask_idc)[:min_len] if torch.tensor(mask_idcs).size() > min_len]
|
We need to run benchmark tests to see by how much the speed improved both on CPU and GPU |
…ction to pytorch based function Implements huggingface#10459 fixes some style changes
|
@patrickvonplaten Can you please help with above comments ? |
|
Hey @punitvara, At the moment, I sadly don't have the time to handle the big chunk of the PR. It would be great if you could try to:
Taking a look at those PRs should help you: #9600, #9453, #6064 |
|
Closing due to inactivity. Sorry @punitvara, I saw a lot of interest from other people to open a PR and this one seems to have stalled. Feel free to re-open it and give it a second shot if you want :-) |
|
I got busy into some other work. I will try to work on different issue. If you get any PR, feel free to merge it |
…ction to pytorch based function
Implements #10459
What does this PR do?
Fixes # (issue)
Before submitting
Pull Request section?
to it if that's the case.
documentation guidelines, and
here are tips on formatting docstrings.
Who can review?
Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors which may be interested in your PR.