Skip to content

Soft error whisper. #22475

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
Narsil merged 2 commits into from
Apr 4, 2023
Merged

Soft error whisper. #22475

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
c384494
Soft error whisper.
Mar 30, 2023
5a94f9c
Fix format.
Narsil Mar 30, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 1 addition & 3 deletions src/transformers/models/whisper/tokenization_whisper.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -877,9 +877,7 @@ def new_chunk():

if previous_tokens:
if return_timestamps:
# Last token should always be timestamps, so there shouldn't be
# leftover
raise ValueError(
logger.warning(
"There was an error while processing timestamps, we haven't found a timestamp as last token. Was"
" WhisperTimeStampLogitsProcessor used?"
)
Comment on lines 881 to 883

@ArthurZucker ArthurZucker Mar 30, 2023

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Since you suspect that the model was probably not finetuned with timestamps we can also add this

@Narsil Narsil Mar 30, 2023

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I hesitated to put it here. While I have 1 example where I think it's true, I don't think it's enough to make it the first and foremost recommendation.

@sanchit-gandhi sanchit-gandhi Apr 18, 2023

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

(I occasionally hit this error when using the pre-trained model in Flax, so it's not specific to fine-tuned ones!)

@Narsil Narsil Apr 18, 2023

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Interesting to know !

Expand Down
33 changes: 32 additions & 1 deletion tests/pipelines/test_pipelines_automatic_speech_recognition.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -17,7 +17,7 @@
import numpy as np
import pytest
from datasets import load_dataset
from huggingface_hub import snapshot_download
from huggingface_hub import hf_hub_download, snapshot_download

from transformers import (
MODEL_FOR_CTC_MAPPING,
Expand All @@ -39,6 +39,7 @@
require_pyctcdecode,
require_tf,
require_torch,
require_torch_gpu,
require_torchaudio,
slow,
)
Expand Down Expand Up @@ -1158,6 +1159,36 @@ def test_stride(self):
output = speech_recognizer({"raw": waveform, "stride": (1000, 8000), "sampling_rate": 16_000})
self.assertEqual(output, {"text": "XB"})

@slow
@require_torch_gpu
def test_slow_unfinished_sequence(self):
from transformers import GenerationConfig

pipe = pipeline(
"automatic-speech-recognition",
model="vasista22/whisper-hindi-large-v2",
device="cuda:0",
)
# Original model wasn't trained with timestamps and has incorrect generation config
pipe.model.generation_config = GenerationConfig.from_pretrained("openai/whisper-large-v2")

audio = hf_hub_download("Narsil/asr_dummy", filename="hindi.ogg", repo_type="dataset")

out = pipe(
audio,
return_timestamps=True,
)
self.assertEqual(
out,
{
"chunks": [
{"text": "", "timestamp": (18.94, 0.0)},
{"text": "मिर्ची में कितने विभिन्न प्रजातियां हैं", "timestamp": (None, None)},
],
"text": "मिर्ची में कितने विभिन्न प्रजातियां हैं",
},
)


def require_ffmpeg(test_case):
"""
Expand Down