All chunks except the first one ignore num_beams in Whisper long-form transcription

[cifkao]

System Info

• transformers version: 4.39.0.dev0 (dabe855)

Who can help?

@patrickvonplaten @sanchit-gandhi

Reproduction

In WhisperGenerationMixin.generate_with_fallback(), the kwargs argument is modified in-place by kwargs.pop("num_beams", 1), which results in the following problems:

1. Subsequent iterations of the fallback loop will use num_beams=1 (although this has no effect provided that temperature 0 appears only once in the temperature list).
2. More importantly, since kwargs is passed by reference (and not via the **kwargs syntax), modifications to it propagate to the caller and to subsequent calls of generate_with_fallback, so every chunk except for the first one will always use num_beams=1.

Expected behavior

The specified num_beams should be used for all chunks.

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.

Please note that issues that do not follow the contributing guidelines are likely to be ignored.

[ArthurZucker]

This sounds like a pretty huge bug, do you have a reproducer?

[cifkao]

@ArthurZucker How is this for a reproducer?

import datasets
from transformers import AutoProcessor, GenerationMixin, WhisperForConditionalGeneration
import numpy as np

processor = AutoProcessor.from_pretrained("openai/whisper-tiny")
model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-tiny")

orig_generate = GenerationMixin.generate


def generate(self, *args, **kwargs):
    print("num_beams:", args[1].num_beams)
    return orig_generate(self, *args, **kwargs)


GenerationMixin.generate = generate

ds = datasets.load_dataset(
    "google/fleurs", "en_us", split="test", trust_remote_code=True
)
ds = ds = ds.cast_column("audio", datasets.Audio(sampling_rate=16000))
raw_audio = np.concatenate([x["array"].astype(np.float32) for x in ds[:16]["audio"]])

inputs = processor(
    [raw_audio],
    return_tensors="pt",
    truncation=False,
    padding="longest",
    return_attention_mask=True,
    sampling_rate=16_000,
)

model.generate(
    **inputs,
    num_beams=2,
    task="transcribe",
    language="en",
)

Before fix:

num_beams: 2
num_beams: 1
num_beams: 1
num_beams: 1
num_beams: 1
num_beams: 1

After fix:

num_beams: 2
num_beams: 2
num_beams: 2
num_beams: 2
num_beams: 2
num_beams: 2

I've tried to modify one of the long-form transcription tests to check that GenerationMixin.generate() is being called with the correct arguments, but I haven't figured out how to mock it out correctly.

[ArthurZucker]

IT's perfect, thanks!