[Whisper] Computing features on GPU in batch mode for whisper feature extractor. (#29900)

* add _torch_extract_fbank_features_batch function in feature_extractor_whisper

* reformat feature_extraction_whisper.py file

* handle batching in single function

* add gpu test & doc

* add batch test & device in each __call__

* add device arg in doc string

---------
Co-authored-by: vaibhav.aggarwal <vaibhav.aggarwal@sprinklr.com>

src/transformers/models/whisper/feature_extraction_whisper.py

@@ -94,11 +94,19 @@ class WhisperFeatureExtractor(SequenceFeatureExtractor):
             mel_scale="slaney",
         )
 
-    def _np_extract_fbank_features(self, waveform: np.array) -> np.ndarray:
+    def _np_extract_fbank_features(self, waveform_batch: np.array, device: str) -> np.ndarray:
         """
         Compute the log-mel spectrogram of the provided audio, gives similar results to Whisper's original torch
         implementation with 1e-5 tolerance.
         """
+        if device != "cpu":
+            raise ValueError(
+                f"Got device `{device}` for feature extraction, but feature extraction on CUDA accelerator "
+                "devices requires torch, which is not installed. Either set `device='cpu'`, or "
+                "install torch according to the official instructions: https://pytorch.org/get-started/locally/"
+            )
+        log_spec_batch = []
+        for waveform in waveform_batch:
             log_spec = spectrogram(
                 waveform,
                 window_function(self.n_fft, "hann"),
@@ -111,24 +119,38 @@ class WhisperFeatureExtractor(SequenceFeatureExtractor):
             log_spec = log_spec[:, :-1]
             log_spec = np.maximum(log_spec, log_spec.max() - 8.0)
             log_spec = (log_spec + 4.0) / 4.0
-        return log_spec
+            log_spec_batch.append(log_spec)
+        log_spec_batch = np.array(log_spec_batch)
+        return log_spec_batch
 
-    def _torch_extract_fbank_features(self, waveform: np.array) -> np.ndarray:
+    def _torch_extract_fbank_features(self, waveform: np.array, device: str = "cpu") -> np.ndarray:
         """
-        Compute the log-mel spectrogram of the provided audio using the PyTorch STFT implementation.
+        Compute the log-mel spectrogram of the audio using PyTorch's GPU-accelerated STFT implementation with batching,
+        yielding results similar to cpu computing with 1e-5 tolerance.
         """
         waveform = torch.from_numpy(waveform).type(torch.float32)
 
         window = torch.hann_window(self.n_fft)
+        if device != "cpu":
+            waveform = waveform.to(device)
+            window = window.to(device)
         stft = torch.stft(waveform, self.n_fft, self.hop_length, window=window, return_complex=True)
         magnitudes = stft[..., :-1].abs() ** 2
 
         mel_filters = torch.from_numpy(self.mel_filters).type(torch.float32)
+        if device != "cpu":
+            mel_filters = mel_filters.to(device)
         mel_spec = mel_filters.T @ magnitudes
 
         log_spec = torch.clamp(mel_spec, min=1e-10).log10()
+        if waveform.dim() == 2:
+            max_val = log_spec.max(dim=2, keepdim=True)[0].max(dim=1, keepdim=True)[0]
+            log_spec = torch.maximum(log_spec, max_val - 8.0)
+        else:
             log_spec = torch.maximum(log_spec, log_spec.max() - 8.0)
         log_spec = (log_spec + 4.0) / 4.0
+        if device != "cpu":
+            log_spec = log_spec.detach().cpu()
         return log_spec.numpy()
 
     @staticmethod
@@ -165,6 +187,7 @@ class WhisperFeatureExtractor(SequenceFeatureExtractor):
         max_length: Optional[int] = None,
         sampling_rate: Optional[int] = None,
         do_normalize: Optional[bool] = None,
+        device: Optional[str] = "cpu",
         **kwargs,
     ) -> BatchFeature:
         """
@@ -211,6 +234,9 @@ class WhisperFeatureExtractor(SequenceFeatureExtractor):
             do_normalize (`bool`, *optional*, defaults to `False`):
                 Whether or not to zero-mean unit-variance normalize the input. Normalizing can help to significantly
                 improve the performance of the model.
+            device (`str`, *optional*, defaults to `'cpu'`):
+                Specifies the device for computation of the log-mel spectrogram of audio signals in the
+                `_torch_extract_fbank_features` method. (e.g., "cpu", "cuda")
         """
 
         if sampling_rate is not None:
@@ -272,7 +298,7 @@ class WhisperFeatureExtractor(SequenceFeatureExtractor):
         extract_fbank_features = (
             self._torch_extract_fbank_features if is_torch_available() else self._np_extract_fbank_features
         )
-        input_features = [extract_fbank_features(waveform) for waveform in input_features[0]]
+        input_features = extract_fbank_features(input_features[0], device)
 
         if isinstance(input_features[0], List):
             padded_inputs["input_features"] = [np.asarray(feature, dtype=np.float32) for feature in input_features]

tests/models/whisper/test_feature_extraction_whisper.py

@@ -24,7 +24,7 @@ import numpy as np
 from datasets import load_dataset
 
 from transformers import WhisperFeatureExtractor
-from transformers.testing_utils import check_json_file_has_correct_format, require_torch
+from transformers.testing_utils import check_json_file_has_correct_format, require_torch, require_torch_gpu
 from transformers.utils.import_utils import is_torch_available
 
 from ...test_sequence_feature_extraction_common import SequenceFeatureExtractionTestMixin
@@ -207,6 +207,7 @@ class WhisperFeatureExtractionTest(SequenceFeatureExtractionTestMixin, unittest.
 
         return [x["array"] for x in speech_samples]
 
+    @require_torch_gpu
     @require_torch
     def test_torch_integration(self):
         # fmt: off
@@ -223,6 +224,7 @@ class WhisperFeatureExtractionTest(SequenceFeatureExtractionTestMixin, unittest.
         input_speech = self._load_datasamples(1)
         feature_extractor = WhisperFeatureExtractor()
         input_features = feature_extractor(input_speech, return_tensors="pt").input_features
+
         self.assertEqual(input_features.shape, (1, 80, 3000))
         self.assertTrue(torch.allclose(input_features[0, 0, :30], EXPECTED_INPUT_FEATURES, atol=1e-4))
 
@@ -253,3 +255,37 @@ class WhisperFeatureExtractionTest(SequenceFeatureExtractionTestMixin, unittest.
 
         self.assertTrue(np.all(np.mean(audio) < 1e-3))
         self.assertTrue(np.all(np.abs(np.var(audio) - 1) < 1e-3))
+
+    @require_torch_gpu
+    @require_torch
+    def test_torch_integration_batch(self):
+        # fmt: off
+        EXPECTED_INPUT_FEATURES = torch.tensor(
+            [
+                [
+                    0.1193, -0.0946, -0.1098, -0.0196, 0.0225, -0.0690, -0.1736, 0.0951,
+                    0.0971, -0.0817, -0.0702, 0.0162, 0.0260, 0.0017, -0.0192, -0.1678,
+                    0.0709, -0.1867, -0.0655, -0.0274, -0.0234, -0.1884, -0.0516, -0.0554,
+                    -0.0274, -0.1425, -0.1423, 0.0837, 0.0377, -0.0854
+                ],
+                [
+                    -0.4696, -0.0751, 0.0276, -0.0312, -0.0540, -0.0383, 0.1295, 0.0568,
+                    -0.2071, -0.0548, 0.0389, -0.0316, -0.2346, -0.1068, -0.0322, 0.0475,
+                    -0.1709, -0.0041, 0.0872, 0.0537, 0.0075, -0.0392, 0.0371, 0.0189,
+                    -0.1522, -0.0270, 0.0744, 0.0738, -0.0245, -0.0667
+                ],
+                [
+                    -0.2337, -0.0060, -0.0063, -0.2353, -0.0431, 0.1102, -0.1492, -0.0292,
+                     0.0787, -0.0608, 0.0143, 0.0582, 0.0072, 0.0101, -0.0444, -0.1701,
+                     -0.0064, -0.0027, -0.0826, -0.0730, -0.0099, -0.0762, -0.0170, 0.0446,
+                     -0.1153, 0.0960, -0.0361, 0.0652, 0.1207, 0.0277
+                ]
+            ]
+        )
+        # fmt: on
+
+        input_speech = self._load_datasamples(3)
+        feature_extractor = WhisperFeatureExtractor()
+        input_features = feature_extractor(input_speech, return_tensors="pt").input_features
+        self.assertEqual(input_features.shape, (3, 80, 3000))
+        self.assertTrue(torch.allclose(input_features[:, 0, :30], EXPECTED_INPUT_FEATURES, atol=1e-4))