Hotfix chunk_length_s instead of _ms.
#15029
Conversation
| The input length for in each chunk. If `0` then chunking is disabled (default). Only available for CTC | ||
| models. | ||
| stride_length_ms (`int`, *optional*, defaults to `chunk_length_ms / 6`): | ||
| The length of stride on the left and right of each chunk. Used only with `chunk_length_ms > 0`. This | ||
| stride_length_s (`int`, *optional*, defaults to `chunk_length_s / 6`): |
There was a problem hiding this comment.
type should be float IMO
| inputs_len = len(inputs) | ||
| chunk_len = chunk_length_ms * self.feature_extractor.sampling_rate // 1000 | ||
| stride_len = stride_length_ms * self.feature_extractor.sampling_rate // 1000 | ||
| chunk_len = chunk_length_s * self.feature_extractor.sampling_rate |
There was a problem hiding this comment.
| chunk_len = chunk_length_s * self.feature_extractor.sampling_rate | |
| chunk_len = int(chunk_length_s * self.feature_extractor.sampling_rate) |
is safer no?
There was a problem hiding this comment.
True, need to round too actually !
|
Feel free to merge as soon as the tests are passing :-) |
* Hotfix `chunk_length_s` instead of `_ms`. * Adding fix of `pad_token` which should be last/previous token for CTC proper decoding * Fixing ChunkPipeline unwrapping. * Adding a PackIterator specific test.
|
|
||
| def chunk_iter(inputs, feature_extractor, chunk_len, stride_left, stride_right): | ||
| inputs_len = inputs.shape[0] | ||
| step = chunk_len - stride_left - stride_right |
There was a problem hiding this comment.
What does stride mean in this context? Because step doesn't equal chunk_size + stride_left, it is kind of hard to interpret what exactly stride_left or stride_right represent.
There was a problem hiding this comment.
If you calculate the step in this way, the actual overlap with the left chunk would be 2 * (stride_left + stride_right).
Consider, for example, chunk_len = 15, stride_left = 3, stride_right = 3, and step = 15-3-3 = 9.
0 chunk: start=0, end=15,
1 chunk: start=9, end=24,
2 chunk: start=18, end=33
For 1 chunk, the overlap with the 0 chunk is 6, and the overlap with the 2 chunk is 6, totaling 12 of overlap, but the intended overlap was 3 with the left chunk and 3 with the right chunk, totaling 5 of overlap.
And fixes issue that the filled in token was the padded_token which could lead to incorrect decoding.
Using the same token for CTC is better (prevent extra repetition)
What does this PR do?
Fixes # (issue)
@patrickvonplaten
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 who may be interested in your PR.