[tests] deflake dither test #36284
[tests] deflake dither test #36284
Conversation
| # compute features | ||
| input_features_no_dither = feature_extractor_no_dither( | ||
| np_speech_inputs, padding=True, return_tensors="np" | ||
| np_speech_inputs, padding=True, return_tensors="np", sampling_rate=dict_no_dither["sampling_rate"] |
There was a problem hiding this comment.
(the sampling_rate arg is redundant when it is manually set at init time, but we throw a warning nevertheless. warnings make flake-finder runs much slower :D)
|
I would love to leave this PR for @eustlb 🤗 but feel free to ping me if anything gets blocked for too long |
eustlb
left a comment
eustlb
left a comment
There was a problem hiding this comment.
LGTM! Thanks @gante.
The sampling_rate arg is indeed redundant. Nevertheless, it is well designed this way IMO to the extent that we want it as a feature-extractor level instance, to reflect the one on which the model has been trained, and we also want to have it as an argument when the feature extractor is called to ensure that the user is warned in the (very likely) case where he forgot to resample the audio.
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
|
(@eustlb perhaps I didn't convey my message correctly -- I agree with the warning 🤗 I wanted to clarify that this PR's diff in the lines where we call the processor exists so that we don't throw a warning at test time, which ultimately makes the |
4 participants
What does this PR do?
test_dither, added in Add dithering to theSpeech2TextFeatureExtractorAPI. #34638. On a--flake-finder --flake-runs 1000runs we now get 0 failures.sampling_ratewarning message to specify what"this function"means (it was unclear from the warning message).