[SequenceFeatureExtractor] Rewrite padding logic from pure python to numpy (#13650)

* Test np padding

* Pass feature extraction tests

* Update type hints

* Fix flaky integration tests

* Try a more stable waveform

* Add to_numpy jax support

* int32 attention masks

* Refactor normalization tests

src/transformers/feature_extraction_sequence_utils.py

@@ -27,7 +27,7 @@ from .file_utils import (
     _is_torch,
     is_tf_available,
     is_torch_available,
-    to_py_obj,
+    to_numpy,
 )
 from .utils import logging
 
@@ -151,7 +151,7 @@ class SequenceFeatureExtractor(FeatureExtractionMixin):
                 processed_features["attention_mask"] = []
             return processed_features
 
-        # If we have PyTorch/TF/NumPy tensors/arrays as inputs, we cast them as python objects
+        # If we have PyTorch/TF tensors or lists as inputs, we cast them as Numpy arrays
         # and rebuild them afterwards if no return_tensors is specified
         # Note that we lose the specific device the tensor may be on for PyTorch
 
@@ -163,14 +163,14 @@ class SequenceFeatureExtractor(FeatureExtractionMixin):
                 index += 1
             if index < len(required_input):
                 first_element = required_input[index][0]
-        # At this state, if `first_element` is still a list/tuple, it's an empty one so there is nothing to do.
-        if not isinstance(first_element, (float, int, list, tuple)):
+
+        if return_tensors is None:
             if is_tf_available() and _is_tensorflow(first_element):
-                return_tensors = "tf" if return_tensors is None else return_tensors
+                return_tensors = "tf"
             elif is_torch_available() and _is_torch(first_element):
-                return_tensors = "pt" if return_tensors is None else return_tensors
-            elif isinstance(first_element, np.ndarray):
-                return_tensors = "np" if return_tensors is None else return_tensors
+                return_tensors = "pt"
+            elif isinstance(first_element, (int, float, list, tuple, np.ndarray)):
+                return_tensors = "np"
             else:
                 raise ValueError(
                     f"type of {first_element} unknown: {type(first_element)}. "
@@ -178,13 +178,16 @@ class SequenceFeatureExtractor(FeatureExtractionMixin):
                 )
 
         for key, value in processed_features.items():
-                processed_features[key] = to_py_obj(value)
+            if isinstance(value[0], (int, float)):
+                processed_features[key] = to_numpy(value)
+            else:
+                processed_features[key] = [to_numpy(v) for v in value]
 
         # Convert padding_strategy in PaddingStrategy
         padding_strategy = self._get_padding_strategies(padding=padding, max_length=max_length)
 
         required_input = processed_features[self.model_input_names[0]]
-        if required_input and not isinstance(required_input[0], (list, tuple)):
+        if required_input and not isinstance(required_input[0], np.ndarray):
             # truncation
             processed_features = self._truncate(
                 processed_features,
@@ -203,9 +206,8 @@ class SequenceFeatureExtractor(FeatureExtractionMixin):
             return BatchFeature(processed_features, tensor_type=return_tensors)
 
         batch_size = len(required_input)
-        assert all(
-            len(v) == batch_size for v in processed_features.values()
-        ), "Some items in the output dictionary have a different batch size than others."
+        if not all(len(v) == batch_size for v in processed_features.values()):
+            raise ValueError("Some items in the output dictionary have a different batch size than others.")
 
         truncated_inputs = []
         for i in range(batch_size):
@@ -244,7 +246,7 @@ class SequenceFeatureExtractor(FeatureExtractionMixin):
 
     def _pad(
         self,
-        processed_features: Union[Dict[str, List[float]], BatchFeature],
+        processed_features: Union[Dict[str, np.ndarray], BatchFeature],
         max_length: Optional[int] = None,
         padding_strategy: PaddingStrategy = PaddingStrategy.DO_NOT_PAD,
         pad_to_multiple_of: Optional[int] = None,
@@ -254,7 +256,7 @@ class SequenceFeatureExtractor(FeatureExtractionMixin):
         Pad inputs (on left/right and up to predefined length or max length in the batch)
 
         Args:
-            processed_features: Dictionary of input values (`List[float]`) / input vectors (`List[List[float]]`) or batch of inputs values (`List[List[int]]`) / input vectors (`List[List[List[int]]]`)
+            processed_features: Dictionary of input values (`np.ndarray[float]`) / input vectors (`List[np.ndarray[float]]`) or batch of inputs values (`List[np.ndarray[int]]`) / input vectors (`List[np.ndarray[int]]`)
             max_length: maximum length of the returned list and optionally padding length (see below)
             padding_strategy: PaddingStrategy to use for padding.
 
@@ -278,42 +280,47 @@ class SequenceFeatureExtractor(FeatureExtractionMixin):
         if max_length is not None and pad_to_multiple_of is not None and (max_length % pad_to_multiple_of != 0):
             max_length = ((max_length // pad_to_multiple_of) + 1) * pad_to_multiple_of
 
-        needs_to_be_padded = padding_strategy != PaddingStrategy.DO_NOT_PAD and len(required_input) != max_length
+        needs_to_be_padded = padding_strategy != PaddingStrategy.DO_NOT_PAD and len(required_input) < max_length
 
         if needs_to_be_padded:
             difference = max_length - len(required_input)
-            padding_vector = self.feature_size * [self.padding_value] if self.feature_size > 1 else self.padding_value
             if self.padding_side == "right":
                 if return_attention_mask:
-                    processed_features["attention_mask"] = [1] * len(required_input) + [0] * difference
-                processed_features[self.model_input_names[0]] = required_input + [
-                    padding_vector for _ in range(difference)
-                ]
+                    attention_mask = np.zeros(max_length, dtype=np.int32)
+                    attention_mask[: len(required_input)] = 1
+                    processed_features["attention_mask"] = attention_mask
+                padding_shape = ((0, difference), (0, 0)) if self.feature_size > 1 else (0, difference)
+                processed_features[self.model_input_names[0]] = np.pad(
+                    required_input, padding_shape, "constant", constant_values=self.padding_value
+                )
             elif self.padding_side == "left":
                 if return_attention_mask:
-                    processed_features["attention_mask"] = [0] * difference + [1] * len(required_input)
-                processed_features[self.model_input_names[0]] = [
-                    padding_vector for _ in range(difference)
-                ] + required_input
+                    attention_mask = np.zeros(max_length, dtype=np.int32)
+                    attention_mask[-len(required_input) :] = 1
+                    processed_features["attention_mask"] = attention_mask
+                padding_shape = ((difference, 0), (0, 0)) if self.feature_size > 1 else (difference, 0)
+                processed_features[self.model_input_names[0]] = np.pad(
+                    required_input, padding_shape, "constant", constant_values=self.padding_value
+                )
             else:
                 raise ValueError("Invalid padding strategy:" + str(self.padding_side))
         elif return_attention_mask and "attention_mask" not in processed_features:
-            processed_features["attention_mask"] = [1] * len(required_input)
+            processed_features["attention_mask"] = np.ones(len(required_input), dtype=np.int32)
 
         return processed_features
 
     def _truncate(
         self,
-        processed_features: Union[Dict[str, List[float]], BatchFeature],
+        processed_features: Union[Dict[str, np.ndarray], BatchFeature],
         max_length: Optional[int] = None,
         pad_to_multiple_of: Optional[int] = None,
         truncation: Optional[bool] = None,
     ):
         """
-        Pad inputs (on left/right and up to predefined length or max length in the batch)
+        Truncate inputs to predefined length or max length in the batch
 
         Args:
-            processed_features: Dictionary of input values (`List[float]`) / input vectors (`List[List[float]]`) or batch of inputs values (`List[List[int]]`) / input vectors (`List[List[List[int]]]`)
+            processed_features: Dictionary of input values (`np.ndarray[float]`) / input vectors (`List[np.ndarray[float]]`) or batch of inputs values (`List[np.ndarray[int]]`) / input vectors (`List[np.ndarray[int]]`)
             max_length: maximum length of the returned list and optionally padding length (see below)
             pad_to_multiple_of: (optional) Integer if set will pad the sequence to a multiple of the provided value.
                 This is especially useful to enable the use of Tensor Core on NVIDIA hardware with compute capability

src/transformers/file_utils.py

@@ -1855,12 +1855,32 @@ def to_py_obj(obj):
         return obj.numpy().tolist()
     elif is_torch_available() and _is_torch(obj):
         return obj.detach().cpu().tolist()
+    elif is_flax_available() and _is_jax(obj):
+        return np.asarray(obj).tolist()
     elif isinstance(obj, np.ndarray):
         return obj.tolist()
     else:
         return obj
 
 
+def to_numpy(obj):
+    """
+    Convert a TensorFlow tensor, PyTorch tensor, Numpy array or python list to a Numpy array.
+    """
+    if isinstance(obj, (dict, UserDict)):
+        return {k: to_numpy(v) for k, v in obj.items()}
+    elif isinstance(obj, (list, tuple)):
+        return np.array(obj)
+    elif is_tf_available() and _is_tensorflow(obj):
+        return obj.numpy()
+    elif is_torch_available() and _is_torch(obj):
+        return obj.detach().cpu().numpy()
+    elif is_flax_available() and _is_jax(obj):
+        return np.asarray(obj)
+    else:
+        return obj
+
+
 class ModelOutput(OrderedDict):
     """
     Base class for all model outputs as dataclass. Has a ``__getitem__`` that allows indexing by integer or slice (like

src/transformers/models/speech_to_text/feature_extraction_speech_to_text.py

@@ -99,15 +99,12 @@ class Speech2TextFeatureExtractor(SequenceFeatureExtractor):
         normalize_vars: Optional[bool] = True,
         padding_value: float = 0.0,
     ) -> np.ndarray:
-        # make sure we normalie float32 arrays
-        mean = x[:input_length].mean(axis=0)
-        square_sums = (x[:input_length] ** 2).sum(axis=0)
-
+        # make sure we normalize float32 arrays
         if normalize_means:
+            mean = x[:input_length].mean(axis=0)
             x = np.subtract(x, mean)
         if normalize_vars:
-            var = square_sums / x[:input_length].shape[0] - mean ** 2
-            std = np.sqrt(np.maximum(var, 1e-10))
+            std = x[:input_length].std(axis=0)
             x = np.divide(x, std)
 
         if input_length < x.shape[0]:
@@ -206,10 +203,10 @@ class Speech2TextFeatureExtractor(SequenceFeatureExtractor):
             and (isinstance(raw_speech[0], np.ndarray) or isinstance(raw_speech[0], (tuple, list)))
         )
 
-        if is_batched and not isinstance(raw_speech[0], np.ndarray):
-            raw_speech = [np.asarray(speech) for speech in raw_speech]
+        if is_batched:
+            raw_speech = [np.asarray(speech, dtype=np.float32) for speech in raw_speech]
         elif not is_batched and not isinstance(raw_speech, np.ndarray):
-            raw_speech = np.asarray(raw_speech)
+            raw_speech = np.asarray(raw_speech, dtype=np.float32)
         elif isinstance(raw_speech, np.ndarray) and raw_speech.dtype is np.dtype(np.float64):
             raw_speech = raw_speech.astype(np.float32)
 

tests/test_feature_extraction_speech_to_text.py

@@ -110,6 +110,10 @@ class Speech2TextFeatureExtractionTest(SequenceFeatureExtractionTestMixin, unitt
     def setUp(self):
         self.feat_extract_tester = Speech2TextFeatureExtractionTester(self)
 
+    def _check_zero_mean_unit_variance(self, input_vector):
+        self.assertTrue(np.all(np.mean(input_vector, axis=0) < 1e-3))
+        self.assertTrue(np.all(np.abs(np.var(input_vector, axis=0) - 1) < 1e-3))
+
     def test_call(self):
         # Tests that all call wrap to encode_plus and batch_encode_plus
         feature_extractor = self.feature_extraction_class(**self.feat_extract_tester.prepare_feat_extract_dict())
@@ -137,17 +141,9 @@ class Speech2TextFeatureExtractionTest(SequenceFeatureExtractionTestMixin, unitt
         feature_extractor = self.feature_extraction_class(**self.feat_extract_tester.prepare_feat_extract_dict())
         speech_inputs = [floats_list((1, x))[0] for x in range(800, 1400, 200)]
 
-        # TODO(Patrick, Suraj, Anton) - It's surprising that "non-padded/non-numpified" padding
-        # results in quite inaccurate variance computation after (see 5e-1 tolerance)
-        # Issue is filed and PR is underway: https://github.com/huggingface/transformers/issues/13539
-        #        paddings = ["longest", "max_length", "do_not_pad"]
-        #        max_lengths = [None, 16, None]
-        #        var_tolerances = [1e-3, 1e-3, 5e-1]
-        paddings = ["longest", "max_length"]
-        max_lengths = [None, 16]
-        var_tolerances = [1e-3, 1e-3]
-        for max_length, padding, var_tol in zip(max_lengths, paddings, var_tolerances):
-
+        paddings = ["longest", "max_length", "do_not_pad"]
+        max_lengths = [None, 16, None]
+        for max_length, padding in zip(max_lengths, paddings):
             inputs = feature_extractor(
                 speech_inputs, padding=padding, max_length=max_length, return_attention_mask=True
             )
@@ -155,28 +151,17 @@ class Speech2TextFeatureExtractionTest(SequenceFeatureExtractionTestMixin, unitt
             attention_mask = inputs.attention_mask
             fbank_feat_lengths = [np.sum(x) for x in attention_mask]
 
-            def _check_zero_mean_unit_variance(input_vector, var_tol=1e-3):
-                self.assertTrue(np.all(np.mean(input_vector, axis=0) < 1e-3))
-                self.assertTrue(np.all(np.abs(np.var(input_vector, axis=0) - 1) < var_tol))
-
-            _check_zero_mean_unit_variance(input_features[0][: fbank_feat_lengths[0]], var_tol)
-            _check_zero_mean_unit_variance(input_features[1][: fbank_feat_lengths[1]], var_tol)
-            _check_zero_mean_unit_variance(input_features[2][: fbank_feat_lengths[2]], var_tol)
+            self._check_zero_mean_unit_variance(input_features[0][: fbank_feat_lengths[0]])
+            self._check_zero_mean_unit_variance(input_features[1][: fbank_feat_lengths[1]])
+            self._check_zero_mean_unit_variance(input_features[2][: fbank_feat_lengths[2]])
 
     def test_cepstral_mean_and_variance_normalization_np(self):
         feature_extractor = self.feature_extraction_class(**self.feat_extract_tester.prepare_feat_extract_dict())
         speech_inputs = [floats_list((1, x))[0] for x in range(800, 1400, 200)]
 
-        # TODO(Patrick, Suraj, Anton) - It's surprising that "non-padded/non-numpified" padding
-        # results in quite inaccurate variance computation after (see 5e-1 tolerance)
-        # Issue is filed and PR is underway: https://github.com/huggingface/transformers/issues/13539
-        #        paddings = ["longest", "max_length", "do_not_pad"]
-        #        max_lengths = [None, 16, None]
-        #        var_tolerances = [1e-3, 1e-3, 5e-1]
-        paddings = ["longest", "max_length"]
-        max_lengths = [None, 16]
-        var_tolerances = [1e-3, 1e-3]
-        for max_length, padding, var_tol in zip(max_lengths, paddings, var_tolerances):
+        paddings = ["longest", "max_length", "do_not_pad"]
+        max_lengths = [None, 16, None]
+        for max_length, padding in zip(max_lengths, paddings):
             inputs = feature_extractor(
                 speech_inputs, max_length=max_length, padding=padding, return_tensors="np", return_attention_mask=True
             )
@@ -184,15 +169,11 @@ class Speech2TextFeatureExtractionTest(SequenceFeatureExtractionTestMixin, unitt
             attention_mask = inputs.attention_mask
             fbank_feat_lengths = [np.sum(x) for x in attention_mask]
 
-            def _check_zero_mean_unit_variance(input_vector, var_tol=1e-3):
-                self.assertTrue(np.all(np.mean(input_vector, axis=0) < 1e-3))
-                self.assertTrue(np.all(np.abs(np.var(input_vector, axis=0) - 1) < var_tol))
-
-            _check_zero_mean_unit_variance(input_features[0][: fbank_feat_lengths[0]], var_tol)
+            self._check_zero_mean_unit_variance(input_features[0][: fbank_feat_lengths[0]])
             self.assertTrue(input_features[0][fbank_feat_lengths[0] :].sum() < 1e-6)
-            _check_zero_mean_unit_variance(input_features[1][: fbank_feat_lengths[1]], var_tol)
+            self._check_zero_mean_unit_variance(input_features[1][: fbank_feat_lengths[1]])
             self.assertTrue(input_features[0][fbank_feat_lengths[1] :].sum() < 1e-6)
-            _check_zero_mean_unit_variance(input_features[2][: fbank_feat_lengths[2]], var_tol)
+            self._check_zero_mean_unit_variance(input_features[2][: fbank_feat_lengths[2]])
 
     def test_cepstral_mean_and_variance_normalization_trunc_max_length(self):
         feature_extractor = self.feature_extraction_class(**self.feat_extract_tester.prepare_feat_extract_dict())
@@ -209,13 +190,9 @@ class Speech2TextFeatureExtractionTest(SequenceFeatureExtractionTestMixin, unitt
         attention_mask = inputs.attention_mask
         fbank_feat_lengths = np.sum(attention_mask == 1, axis=1)
 
-        def _check_zero_mean_unit_variance(input_vector):
-            self.assertTrue(np.all(np.mean(input_vector, axis=0) < 1e-3))
-            self.assertTrue(np.all(np.abs(np.var(input_vector, axis=0) - 1) < 1e-3))
-
-        _check_zero_mean_unit_variance(input_features[0, : fbank_feat_lengths[0]])
-        _check_zero_mean_unit_variance(input_features[1])
-        _check_zero_mean_unit_variance(input_features[2])
+        self._check_zero_mean_unit_variance(input_features[0, : fbank_feat_lengths[0]])
+        self._check_zero_mean_unit_variance(input_features[1])
+        self._check_zero_mean_unit_variance(input_features[2])
 
     def test_cepstral_mean_and_variance_normalization_trunc_longest(self):
         feature_extractor = self.feature_extraction_class(**self.feat_extract_tester.prepare_feat_extract_dict())
@@ -232,13 +209,9 @@ class Speech2TextFeatureExtractionTest(SequenceFeatureExtractionTestMixin, unitt
         attention_mask = inputs.attention_mask
         fbank_feat_lengths = np.sum(attention_mask == 1, axis=1)
 
-        def _check_zero_mean_unit_variance(input_vector):
-            self.assertTrue(np.all(np.mean(input_vector, axis=0) < 1e-3))
-            self.assertTrue(np.all(np.abs(np.var(input_vector, axis=0) - 1) < 1e-3))
-
-        _check_zero_mean_unit_variance(input_features[0, : fbank_feat_lengths[0]])
-        _check_zero_mean_unit_variance(input_features[1, : fbank_feat_lengths[1]])
-        _check_zero_mean_unit_variance(input_features[2])
+        self._check_zero_mean_unit_variance(input_features[0, : fbank_feat_lengths[0]])
+        self._check_zero_mean_unit_variance(input_features[1, : fbank_feat_lengths[1]])
+        self._check_zero_mean_unit_variance(input_features[2])
 
         # make sure that if max_length < longest -> then pad to max_length
         self.assertEqual(input_features.shape, (3, 4, 24))
@@ -256,9 +229,9 @@ class Speech2TextFeatureExtractionTest(SequenceFeatureExtractionTestMixin, unitt
         attention_mask = inputs.attention_mask
         fbank_feat_lengths = np.sum(attention_mask == 1, axis=1)
 
-        _check_zero_mean_unit_variance(input_features[0, : fbank_feat_lengths[0]])
-        _check_zero_mean_unit_variance(input_features[1, : fbank_feat_lengths[1]])
-        _check_zero_mean_unit_variance(input_features[2])
+        self._check_zero_mean_unit_variance(input_features[0, : fbank_feat_lengths[0]])
+        self._check_zero_mean_unit_variance(input_features[1, : fbank_feat_lengths[1]])
+        self._check_zero_mean_unit_variance(input_features[2])
 
         # make sure that if max_length < longest -> then pad to max_length
         self.assertEqual(input_features.shape, (3, 6, 24))

tests/test_feature_extraction_wav2vec2.py

@@ -102,6 +102,10 @@ class Wav2Vec2FeatureExtractionTest(SequenceFeatureExtractionTestMixin, unittest
     def setUp(self):
         self.feat_extract_tester = Wav2Vec2FeatureExtractionTester(self)
 
+    def _check_zero_mean_unit_variance(self, input_vector):
+        self.assertTrue(np.all(np.mean(input_vector, axis=0) < 1e-3))
+        self.assertTrue(np.all(np.abs(np.var(input_vector, axis=0) - 1) < 1e-3))
+
     def test_call(self):
         # Tests that all call wrap to encode_plus and batch_encode_plus
         feat_extract = self.feature_extraction_class(**self.feat_extract_tester.prepare_feat_extract_dict())
@@ -130,15 +134,11 @@ class Wav2Vec2FeatureExtractionTest(SequenceFeatureExtractionTestMixin, unittest
             processed = feat_extract(speech_inputs, padding=padding, max_length=max_length, return_tensors="np")
             input_values = processed.input_values
 
-            def _check_zero_mean_unit_variance(input_vector):
-                self.assertTrue(np.abs(np.mean(input_vector)) < 1e-3)
-                self.assertTrue(np.abs(np.var(input_vector) - 1) < 1e-3)
-
-            _check_zero_mean_unit_variance(input_values[0][:800])
+            self._check_zero_mean_unit_variance(input_values[0][:800])
             self.assertTrue(input_values[0][800:].sum() < 1e-6)
-            _check_zero_mean_unit_variance(input_values[1][:1000])
+            self._check_zero_mean_unit_variance(input_values[1][:1000])
             self.assertTrue(input_values[0][1000:].sum() < 1e-6)
-            _check_zero_mean_unit_variance(input_values[2][:1200])
+            self._check_zero_mean_unit_variance(input_values[2][:1200])
 
     def test_zero_mean_unit_variance_normalization(self):
         feat_extract = self.feature_extraction_class(**self.feat_extract_tester.prepare_feat_extract_dict())
@@ -152,13 +152,9 @@ class Wav2Vec2FeatureExtractionTest(SequenceFeatureExtractionTestMixin, unittest
             processed = feat_extract(speech_inputs, max_length=max_length, padding=padding)
             input_values = processed.input_values
 
-            def _check_zero_mean_unit_variance(input_vector):
-                self.assertTrue(np.abs(np.mean(input_vector)) < 1e-3)
-                self.assertTrue(np.abs(np.var(input_vector) - 1) < 1e-3)
-
-            _check_zero_mean_unit_variance(input_values[0][:800])
-            _check_zero_mean_unit_variance(input_values[1][:1000])
-            _check_zero_mean_unit_variance(input_values[2][:1200])
+            self._check_zero_mean_unit_variance(input_values[0][:800])
+            self._check_zero_mean_unit_variance(input_values[1][:1000])
+            self._check_zero_mean_unit_variance(input_values[2][:1200])
 
     def test_zero_mean_unit_variance_normalization_trunc_np_max_length(self):
         feat_extract = self.feature_extraction_class(**self.feat_extract_tester.prepare_feat_extract_dict())
@@ -168,13 +164,9 @@ class Wav2Vec2FeatureExtractionTest(SequenceFeatureExtractionTestMixin, unittest
         )
         input_values = processed.input_values
 
-        def _check_zero_mean_unit_variance(input_vector):
-            self.assertTrue(np.abs(np.mean(input_vector)) < 1e-3)
-            self.assertTrue(np.abs(np.var(input_vector) - 1) < 1e-3)
-
-        _check_zero_mean_unit_variance(input_values[0, :800])
-        _check_zero_mean_unit_variance(input_values[1])
-        _check_zero_mean_unit_variance(input_values[2])
+        self._check_zero_mean_unit_variance(input_values[0, :800])
+        self._check_zero_mean_unit_variance(input_values[1])
+        self._check_zero_mean_unit_variance(input_values[2])
 
     def test_zero_mean_unit_variance_normalization_trunc_np_longest(self):
         feat_extract = self.feature_extraction_class(**self.feat_extract_tester.prepare_feat_extract_dict())
@@ -184,13 +176,9 @@ class Wav2Vec2FeatureExtractionTest(SequenceFeatureExtractionTestMixin, unittest
         )
         input_values = processed.input_values
 
-        def _check_zero_mean_unit_variance(input_vector):
-            self.assertTrue(np.abs(np.mean(input_vector)) < 1e-3)
-            self.assertTrue(np.abs(np.var(input_vector) - 1) < 1e-3)
-
-        _check_zero_mean_unit_variance(input_values[0, :800])
-        _check_zero_mean_unit_variance(input_values[1, :1000])
-        _check_zero_mean_unit_variance(input_values[2])
+        self._check_zero_mean_unit_variance(input_values[0, :800])
+        self._check_zero_mean_unit_variance(input_values[1, :1000])
+        self._check_zero_mean_unit_variance(input_values[2])
 
         # make sure that if max_length < longest -> then pad to max_length
         self.assertTrue(input_values.shape == (3, 1000))
@@ -201,9 +189,9 @@ class Wav2Vec2FeatureExtractionTest(SequenceFeatureExtractionTestMixin, unittest
         )
         input_values = processed.input_values
 
-        _check_zero_mean_unit_variance(input_values[0, :800])
-        _check_zero_mean_unit_variance(input_values[1, :1000])
-        _check_zero_mean_unit_variance(input_values[2])
+        self._check_zero_mean_unit_variance(input_values[0, :800])
+        self._check_zero_mean_unit_variance(input_values[1, :1000])
+        self._check_zero_mean_unit_variance(input_values[2])
 
         # make sure that if max_length > longest -> then pad to longest
         self.assertTrue(input_values.shape == (3, 1200))

tests/test_modeling_speech_to_text.py

@@ -724,7 +724,7 @@ class Speech2TextModelIntegrationTests(unittest.TestCase):
             return batch
 
         ds = load_dataset("patrickvonplaten/librispeech_asr_dummy", "clean", split="validation")
-        ds = ds.select(range(num_samples)).map(map_to_array)
+        ds = ds.sort("id").select(range(num_samples)).map(map_to_array)
 
         return ds["speech"][:num_samples]
 
@@ -740,7 +740,9 @@ class Speech2TextModelIntegrationTests(unittest.TestCase):
         generated_ids = model.generate(input_features)
         generated_transcript = processor.batch_decode(generated_ids, skip_special_tokens=True)
 
-        EXPECTED_TRANSCRIPTIONS = ["a man said to the universe sir i exist"]
+        EXPECTED_TRANSCRIPTIONS = [
+            "mister quilter is the apostle of the middle classes and we are glad to welcome his gospel"
+        ]
         self.assertListEqual(generated_transcript, EXPECTED_TRANSCRIPTIONS)
 
     def test_generation_librispeech_batched(self):
@@ -759,10 +761,10 @@ class Speech2TextModelIntegrationTests(unittest.TestCase):
         generated_transcripts = processor.batch_decode(generated_ids, skip_special_tokens=True)
 
         EXPECTED_TRANSCRIPTIONS = [
-            "a man said to the universe sir i exist",
-            "sweat covered brion's body trickling into the titleing cloth that was the only garment he wore",
-            "the cut on his chest still dripping blood the ache of his overstrained eyes even the soaring arena around him with the thousands of spectators were trivialities not worth thinking about",
-            "his instant of panic was followed by a small sharp blow high on his chest",
+            "mister quilter is the apostle of the middle classes and we are glad to welcome his gospel",
+            "nor is mister cultar's manner less interesting than his matter",
+            "he tells us that at this festive season of the year with christmas and roast beef looming before us similes drawn from eating and its results occur most readily to the mind",
+            "he has grave doubts whether sir frederick leyton's work is really greek after all and can discover in it but little of rocky ithaca",
         ]
 
         self.assertListEqual(generated_transcripts, EXPECTED_TRANSCRIPTIONS)

tests/test_pipelines_automatic_speech_recognition.py

@@ -42,9 +42,9 @@ class AutomaticSpeechRecognitionPipelineTests(unittest.TestCase):
             tokenizer="facebook/s2t-small-mustc-en-fr-st",
             framework="pt",
         )
-        waveform = np.zeros((34000,))
+        waveform = np.tile(np.arange(1000, dtype=np.float32), 34)
         output = speech_recognizer(waveform)
-        self.assertEqual(output, {"text": "C'est ce que j'ai fait à ce moment-là."})
+        self.assertEqual(output, {"text": "(Applaudissements)"})
 
     @require_torch
     def test_torch_small_no_tokenizer_files(self):
@@ -68,14 +68,14 @@ class AutomaticSpeechRecognitionPipelineTests(unittest.TestCase):
             tokenizer="facebook/wav2vec2-base-960h",
             framework="pt",
         )
-        waveform = np.zeros((34000,))
+        waveform = np.tile(np.arange(1000, dtype=np.float32), 34)
         output = speech_recognizer(waveform)
         self.assertEqual(output, {"text": ""})
 
         from datasets import load_dataset
 
-        ds = load_dataset("patrickvonplaten/librispeech_asr_dummy", "clean", split="validation")
-        filename = ds[0]["file"]
+        ds = load_dataset("patrickvonplaten/librispeech_asr_dummy", "clean", split="validation").sort("id")
+        filename = ds[40]["file"]
         output = speech_recognizer(filename)
         self.assertEqual(output, {"text": "A MAN SAID TO THE UNIVERSE SIR I EXIST"})
 
@@ -92,8 +92,8 @@ class AutomaticSpeechRecognitionPipelineTests(unittest.TestCase):
 
         from datasets import load_dataset
 
-        ds = load_dataset("patrickvonplaten/librispeech_asr_dummy", "clean", split="validation")
-        filename = ds[0]["file"]
+        ds = load_dataset("patrickvonplaten/librispeech_asr_dummy", "clean", split="validation").sort("id")
+        filename = ds[40]["file"]
         output = speech_recognizer(filename)
         self.assertEqual(output, {"text": 'Ein Mann sagte zum Universum : " Sir, ich existiert! "'})
 
@@ -110,16 +110,16 @@ class AutomaticSpeechRecognitionPipelineTests(unittest.TestCase):
 
         asr = AutomaticSpeechRecognitionPipeline(model=model, tokenizer=tokenizer, feature_extractor=feature_extractor)
 
-        waveform = np.zeros((34000,))
+        waveform = np.tile(np.arange(1000, dtype=np.float32), 34)
         output = asr(waveform)
         self.assertEqual(output, {"text": ""})
 
-        ds = load_dataset("patrickvonplaten/librispeech_asr_dummy", "clean", split="validation")
-        filename = ds[0]["file"]
+        ds = load_dataset("patrickvonplaten/librispeech_asr_dummy", "clean", split="validation").sort("id")
+        filename = ds[40]["file"]
         output = asr(filename)
         self.assertEqual(output, {"text": "A MAN SAID TO THE UNIVERSE SIR I EXIST"})
 
-        filename = ds[0]["file"]
+        filename = ds[40]["file"]
         with open(filename, "rb") as f:
             data = f.read()
         output = asr(data)
@@ -139,17 +139,17 @@ class AutomaticSpeechRecognitionPipelineTests(unittest.TestCase):
 
         asr = AutomaticSpeechRecognitionPipeline(model=model, tokenizer=tokenizer, feature_extractor=feature_extractor)
 
-        waveform = np.zeros((34000,))
+        waveform = np.tile(np.arange(1000, dtype=np.float32), 34)
 
         output = asr(waveform)
-        self.assertEqual(output, {"text": "E questo è il motivo per cui non ci siamo mai incontrati."})
+        self.assertEqual(output, {"text": "(Applausi)"})
 
-        ds = load_dataset("patrickvonplaten/librispeech_asr_dummy", "clean", split="validation")
-        filename = ds[0]["file"]
+        ds = load_dataset("patrickvonplaten/librispeech_asr_dummy", "clean", split="validation").sort("id")
+        filename = ds[40]["file"]
         output = asr(filename)
         self.assertEqual(output, {"text": "Un uomo disse all'universo: \"Signore, io esisto."})
 
-        filename = ds[0]["file"]
+        filename = ds[40]["file"]
         with open(filename, "rb") as f:
             data = f.read()
         output = asr(data)

tests/test_sequence_feature_extraction_common.py

@@ -372,7 +372,7 @@ class SequenceFeatureExtractionTestMixin(FeatureExtractionSavingTestMixin):
         input_np = feat_extract.pad(processed_features, padding="longest", return_tensors="np")[input_name]
         input_pt = feat_extract.pad(processed_features, padding="longest", return_tensors="pt")[input_name]
 
-        self.assertTrue(abs(input_np.astype(np.float32).sum() - input_pt.numpy().sum()) < 1e-2)
+        self.assertTrue(abs(input_np.astype(np.float32).sum() - input_pt.numpy().astype(np.float32).sum()) < 1e-2)
 
     @require_tf
     def test_padding_accepts_tensors_tf(self):
@@ -385,7 +385,7 @@ class SequenceFeatureExtractionTestMixin(FeatureExtractionSavingTestMixin):
         input_np = feat_extract.pad(processed_features, padding="longest", return_tensors="np")[input_name]
         input_tf = feat_extract.pad(processed_features, padding="longest", return_tensors="tf")[input_name]
 
-        self.assertTrue(abs(input_np.astype(np.float32).sum() - input_tf.numpy().sum()) < 1e-2)
+        self.assertTrue(abs(input_np.astype(np.float32).sum() - input_tf.numpy().astype(np.float32).sum()) < 1e-2)
 
     def test_attention_mask(self):
         feat_dict = self.feat_extract_dict