Skip to content
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

Revert "Incorrect Whisper long-form decoding timestamps " #32148

Merged
merged 1 commit into from
Jul 23, 2024
Merged

Revert "Incorrect Whisper long-form decoding timestamps " #32148

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
1 change: 1 addition & 0 deletions src/transformers/models/clvp/processing_clvp.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -73,6 +73,7 @@ def __call__(self, *args, **kwargs):
inputs["attention_mask"] = encodings["attention_mask"]
return inputs

# Copied from transformers.models.whisper.processing_whisper.WhisperProcessor.batch_decode with Whisper->Clvp
def batch_decode(self, *args, **kwargs):
"""
This method forwards all its arguments to ClvpTokenizer's [`~PreTrainedTokenizer.batch_decode`]. Please
Expand Down
7 changes: 0 additions & 7 deletions src/transformers/models/whisper/processing_whisper.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -84,13 +84,6 @@ def batch_decode(self, *args, **kwargs):
This method forwards all its arguments to WhisperTokenizer's [`~PreTrainedTokenizer.batch_decode`]. Please
refer to the docstring of this method for more information.
"""

# If segments are present in args, we are performing long-form generation and need to return long form timestamps.
# The long-form timestamps are already present in segments and should be passed as kwargs to batch_decode.
if isinstance(args[0], dict) and "segments" in args[0]:
kwargs["longform_timestamps"] = args[0].pop("segments")
args = tuple(args[0]["sequences"].unsqueeze(0))

return self.tokenizer.batch_decode(*args, **kwargs)

def decode(self, *args, **kwargs):
Expand Down
42 changes: 12 additions & 30 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 @@ -558,7 +558,7 @@ def _decode_with_timestamps(self, token_ids, skip_special_tokens=False, time_pre
]
return "".join(outputs)

def _compute_offsets(self, token_ids, time_precision=0.02, longform_timestamps=None):
def _compute_offsets(self, token_ids, time_precision=0.02):
"""
Compute offsets for a given tokenized input
Expand All @@ -567,8 +567,6 @@ def _compute_offsets(self, token_ids, time_precision=0.02, longform_timestamps=N
List of tokenized input ids. Can be obtained using the `__call__` method.
time_precision (`float`, `optional`, defaults to 0.02):
The time ratio to convert from token to time.
longform_timestamps (List[dict], *optional*):
Timestamps obtained using long form generation in Whisper, to be used to replace predicted timestamps in token_ids.
"""
offsets = []
# ensure torch tensor of token ids is placed on cpu
Expand All @@ -589,7 +587,7 @@ def _compute_offsets(self, token_ids, time_precision=0.02, longform_timestamps=N
consecutive = np.append(consecutive, np.where(timestamp_tokens)[0][-1] + 1)

last_slice = np.where(timestamp_tokens)[0][0]
for i, current_slice in enumerate(consecutive):
for current_slice in consecutive:
sliced_tokens = token_ids[last_slice:current_slice]
if len(sliced_tokens) > 1:
start_timestamp_position = sliced_tokens[0].item() - timestamp_begin
Expand All @@ -598,27 +596,15 @@ def _compute_offsets(self, token_ids, time_precision=0.02, longform_timestamps=N
sliced_tokens = self._preprocess_token_ids(sliced_tokens)
text = self._decode(sliced_tokens)
text = self._filter_timestamp_ids(text)

if longform_timestamps is not None:
offsets.append(
{
"text": text,
"timestamp": (
longform_timestamps[0][i]["start"].item(),
longform_timestamps[0][i]["end"].item(),
),
}
)
else:
offsets.append(
{
"text": text,
"timestamp": (
start_timestamp_position * time_precision,
end_timestamp_position * time_precision,
),
}
)
offsets.append(
{
"text": text,
"timestamp": (
start_timestamp_position * time_precision,
end_timestamp_position * time_precision,
),
}
)
last_slice = current_slice

return offsets
Expand Down Expand Up @@ -727,11 +713,7 @@ def decode(

# retrieve offsets
if output_offsets:
longform_timestamps = kwargs.get("longform_timestamps")
offsets = self._compute_offsets(
token_ids, time_precision=time_precision, longform_timestamps=longform_timestamps
)

offsets = self._compute_offsets(token_ids, time_precision=time_precision)
return {"text": text, "offsets": offsets}
return text

Expand Down
42 changes: 12 additions & 30 deletions src/transformers/models/whisper/tokenization_whisper_fast.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -200,7 +200,7 @@ def _decode_with_timestamps(self, token_ids, skip_special_tokens=False, time_pre
return "".join(outputs)

# Copied from transformers.models.whisper.tokenization_whisper.WhisperTokenizer._compute_offsets
def _compute_offsets(self, token_ids, time_precision=0.02, longform_timestamps=None):
def _compute_offsets(self, token_ids, time_precision=0.02):
"""
Compute offsets for a given tokenized input
Expand All @@ -209,8 +209,6 @@ def _compute_offsets(self, token_ids, time_precision=0.02, longform_timestamps=N
List of tokenized input ids. Can be obtained using the `__call__` method.
time_precision (`float`, `optional`, defaults to 0.02):
The time ratio to convert from token to time.
longform_timestamps (List[dict], *optional*):
Timestamps obtained using long form generation in Whisper, to be used to replace predicted timestamps in token_ids.
"""
offsets = []
# ensure torch tensor of token ids is placed on cpu
Expand All @@ -231,7 +229,7 @@ def _compute_offsets(self, token_ids, time_precision=0.02, longform_timestamps=N
consecutive = np.append(consecutive, np.where(timestamp_tokens)[0][-1] + 1)

last_slice = np.where(timestamp_tokens)[0][0]
for i, current_slice in enumerate(consecutive):
for current_slice in consecutive:
sliced_tokens = token_ids[last_slice:current_slice]
if len(sliced_tokens) > 1:
start_timestamp_position = sliced_tokens[0].item() - timestamp_begin
Expand All @@ -240,27 +238,15 @@ def _compute_offsets(self, token_ids, time_precision=0.02, longform_timestamps=N
sliced_tokens = self._preprocess_token_ids(sliced_tokens)
text = self._decode(sliced_tokens)
text = self._filter_timestamp_ids(text)

if longform_timestamps is not None:
offsets.append(
{
"text": text,
"timestamp": (
longform_timestamps[0][i]["start"].item(),
longform_timestamps[0][i]["end"].item(),
),
}
)
else:
offsets.append(
{
"text": text,
"timestamp": (
start_timestamp_position * time_precision,
end_timestamp_position * time_precision,
),
}
)
offsets.append(
{
"text": text,
"timestamp": (
start_timestamp_position * time_precision,
end_timestamp_position * time_precision,
),
}
)
last_slice = current_slice

return offsets
Expand Down Expand Up @@ -373,11 +359,7 @@ def decode(

# retrieve offsets
if output_offsets:
longform_timestamps = kwargs.get("longform_timestamps")
offsets = self._compute_offsets(
token_ids, time_precision=time_precision, longform_timestamps=longform_timestamps
)

offsets = self._compute_offsets(token_ids, time_precision=time_precision)
return {"text": text, "offsets": offsets}
return text

Expand Down
66 changes: 0 additions & 66 deletions tests/models/whisper/test_modeling_whisper.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2245,72 +2245,6 @@ def test_tiny_timestamp_generation(self):
transcript = processor.batch_decode(generated_ids, skip_special_tokens=True, output_offsets=True)
self.assertEqual(transcript, EXPECTED_TRANSCRIPT)

@slow
def test_tiny_longform_timestamps_generation(self):
set_seed(0)
processor = WhisperProcessor.from_pretrained("openai/whisper-tiny")
model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-tiny")
model.to(torch_device)

sample = self._load_datasamples(1)
input_speech = np.concatenate(sample * 10)

input_features = processor(input_speech, return_tensors="pt", truncation=False, sampling_rate=16_000)
input_features = input_features.to(torch_device)

generated_ids = model.generate(**input_features, return_timestamps=True, return_segments=True)

EXPECTED_TRANSCRIPT = [
{
"text": " Mr. Quilter is the apostle of the middle classes and we are glad to welcome his gospel. Mr. Quilter is the apostle of the middle classes and we are glad to welcome his gospel. Mr. Quilter is the apostle of the middle classes and we are glad to welcome his gospel. Mr. Quilter is the apostle of the middle classes and we are glad to welcome his gospel. Mr. Quilter is the apostle of the middle classes and we are glad to welcome his gospel. Mr. Quilter is the apostle of the middle classes and we are glad to welcome his gospel. Mr. Quilter is the apostle of the middle classes and we are glad to welcome his gospel. Mr. Quilter is the apostle of the middle classes and we are glad to welcome his gospel. Mr. Quilter is the apostle of the middle classes and we are glad to welcome his gospel. Mr. Quilter is the apostle of the middle classes and we are glad to welcome his gospel.",
"offsets": [
{
"text": " Mr. Quilter is the apostle of the middle classes and we are glad to welcome his gospel.",
"timestamp": (0.0, 6.0),
},
{
"text": " Mr. Quilter is the apostle of the middle classes and we are glad to welcome his gospel.",
"timestamp": (6.0, 12.0),
},
{
"text": " Mr. Quilter is the apostle of the middle classes and we are glad to welcome his gospel.",
"timestamp": (12.0, 18.0),
},
{
"text": " Mr. Quilter is the apostle of the middle classes and we are glad to welcome his gospel.",
"timestamp": (18.0, 24.0),
},
{
"text": " Mr. Quilter is the apostle of the middle classes and we are glad to welcome his gospel.",
"timestamp": (24.0, 29.0),
},
{
"text": " Mr. Quilter is the apostle of the middle classes and we are glad to welcome his gospel.",
"timestamp": (29.0, 35.0),
},
{
"text": " Mr. Quilter is the apostle of the middle classes and we are glad to welcome his gospel.",
"timestamp": (35.0, 41.0),
},
{
"text": " Mr. Quilter is the apostle of the middle classes and we are glad to welcome his gospel.",
"timestamp": (41.0, 47.0),
},
{
"text": " Mr. Quilter is the apostle of the middle classes and we are glad to welcome his gospel.",
"timestamp": (47.0, 53.0),
},
{
"text": " Mr. Quilter is the apostle of the middle classes and we are glad to welcome his gospel.",
"timestamp": (53.0, 58.20000076293945),
},
],
}
]

transcript = processor.batch_decode(generated_ids, skip_special_tokens=True, output_offsets=True)
self.assertEqual(transcript, EXPECTED_TRANSCRIPT)

@slow
def test_large_timestamp_generation(self):
set_seed(0)
Expand Down
Loading