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

Revert "Incorrect Whisper long-form decoding timestamps  (#32003)"

This reverts commit cd48553f.

src/transformers/models/clvp/processing_clvp.py

@@ -73,6 +73,7 @@ class ClvpProcessor(ProcessorMixin):
             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

src/transformers/models/whisper/processing_whisper.py

@@ -84,13 +84,6 @@ class WhisperProcessor(ProcessorMixin):
         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):

src/transformers/models/whisper/tokenization_whisper.py

@@ -558,7 +558,7 @@ class WhisperTokenizer(PreTrainedTokenizer):
         ]
         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
 
@@ -567,8 +567,6 @@ class WhisperTokenizer(PreTrainedTokenizer):
                 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
@@ -589,7 +587,7 @@ class WhisperTokenizer(PreTrainedTokenizer):
             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
@@ -598,18 +596,6 @@ class WhisperTokenizer(PreTrainedTokenizer):
                 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,
@@ -727,11 +713,7 @@ class WhisperTokenizer(PreTrainedTokenizer):
 
         # 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
 

src/transformers/models/whisper/tokenization_whisper_fast.py

@@ -200,7 +200,7 @@ class WhisperTokenizerFast(PreTrainedTokenizerFast):
         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
 
@@ -209,8 +209,6 @@ class WhisperTokenizerFast(PreTrainedTokenizerFast):
                 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
@@ -231,7 +229,7 @@ class WhisperTokenizerFast(PreTrainedTokenizerFast):
             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
@@ -240,18 +238,6 @@ class WhisperTokenizerFast(PreTrainedTokenizerFast):
                 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,
@@ -373,11 +359,7 @@ class WhisperTokenizerFast(PreTrainedTokenizerFast):
 
         # 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
 

tests/models/whisper/test_modeling_whisper.py

@@ -2243,72 +2243,6 @@ class WhisperModelIntegrationTests(unittest.TestCase):
         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)