M-CTC-T Model (#16402)

* added cbs to notebooks, made copy-paste error fix in generation_utils

* initial push for mctc model

* mctc feature extractor done

* added processor, tokenizer and their tests for MCTC. Have added an MCTC modeling test, adjusting model code accordingly.

* added processor, tokenizer and their tests for MCTC. Have added an MCTC modeling test, adjusting model code accordingly.

* passing attention, now struggling to figure out how attention masks make sense here

* works when excluding attention masks. ask later how one would integrate attention maskshere

* bizarre configuration error (model prefix comes first in config dict json and messes up the order)

* all passing but bizzarre config dict ordering issue when to_dict

* passing all major tests

* feature extraction, processor, tokenizer added & tests passing

* style & consistency & other logistical fixes

* copy paste fix

* model after feature extraction working

* commiting final feature extraction results; need to fix normalization

* feature extraction passing tests; probably should add tests on the specific flashlight-copied functions?

* delete print ; format code a bit

* fixing tests

* passing major tests

* fixing styles

* completed tokenization test with real example; not sure if these values are entirely correct.

* last test fixes from local

* reverting accidentally included custom setup configs

* remove load tf weights; fix config error

* testing couldnt import featureextractor

* fix docs

* fix docs

* resolving comments

* style fixes

* style fixes

* Update to MCTCConv1dSubSampler
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* relposemb fixes

* conv1d name issue; expecting config fail with paraentheses

* fix config issue

* fix config issue

* fix config issue

* change everything to MCTCT

* fixing naming change errors

* archive list

* copyrights and docs

* copyrights and docs

* copyrights and docs

* merge resolution

* move tests, fix to changed optionaldependency structure

* test directories changed

* fixing tests

* how to avoid tf tests?

* how to avoid tf tests?

* tests passing locally

* allow mctctprocessor imported any env

* allow mctctprocessor imported any env

* fixed second round of feedback, need to fix docs

* doc changes not being applied

* all fixed

* style fix

* feedback fixes

* fix copies and feature extraction style fix

* Update tests/models/visual_bert/test_modeling_visual_bert.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* copy paste huggingface:main visual bert

* added eof newline to visual bert; all tests are passing otherwise

* fix slow tests by adding attention mask

* change model id to speechbrain

* make fix-copies

* fix readme unwanted deletes

* fixing readmes, make fix-copies

* consistent M-CTC-T naming

* Update src/transformers/models/mctct/__init__.py
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* all fixed but variable naming

* adjust double quotes

* fixed variable names

* copyright and mr quilter

* Apply suggestions from code review
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* correct slow tests

* make fix-copies

* Update src/transformers/models/mctct/configuration_mctct.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/mctct/configuration_mctct.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* m-ctc-t not mctct
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

tests/models/mctct/test_feature_extraction_mctct.py

+# coding=utf-8
+# Copyright 2022 HuggingFace Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import itertools
+import random
+import unittest
+
+import numpy as np
+
+from transformers import is_speech_available
+from transformers.testing_utils import require_torch, require_torchaudio
+
+from ...test_sequence_feature_extraction_common import SequenceFeatureExtractionTestMixin
+
+
+if is_speech_available():
+    from transformers import MCTCTFeatureExtractor
+
+global_rng = random.Random()
+
+
+def floats_list(shape, scale=1.0, rng=None, name=None):
+    """Creates a random float32 tensor"""
+    if rng is None:
+        rng = global_rng
+
+    values = []
+    for _batch_idx in range(shape[0]):
+        values.append([])
+        for _ in range(shape[1]):
+            values[-1].append(rng.random() * scale)
+
+    return values
+
+
+@require_torch
+@require_torchaudio
+class MCTCTFeatureExtractionTester(unittest.TestCase):
+    def __init__(
+        self,
+        parent,
+        batch_size=7,
+        min_seq_length=400,
+        max_seq_length=2000,
+        feature_size=24,
+        num_mel_bins=24,
+        padding_value=0.0,
+        sampling_rate=16_000,
+        return_attention_mask=True,
+        do_normalize=True,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.min_seq_length = min_seq_length
+        self.max_seq_length = max_seq_length
+        self.seq_length_diff = (self.max_seq_length - self.min_seq_length) // (self.batch_size - 1)
+        self.feature_size = feature_size
+        self.num_mel_bins = num_mel_bins
+        self.padding_value = padding_value
+        self.sampling_rate = sampling_rate
+        self.return_attention_mask = return_attention_mask
+        self.do_normalize = do_normalize
+
+    def prepare_feat_extract_dict(self):
+        return {
+            "feature_size": self.feature_size,
+            "num_mel_bins": self.num_mel_bins,
+            "padding_value": self.padding_value,
+            "sampling_rate": self.sampling_rate,
+            "return_attention_mask": self.return_attention_mask,
+            "do_normalize": self.do_normalize,
+        }
+
+    def prepare_inputs_for_common(self, equal_length=False, numpify=False):
+        def _flatten(list_of_lists):
+            return list(itertools.chain(*list_of_lists))
+
+        if equal_length:
+            speech_inputs = [floats_list((self.max_seq_length, self.feature_size)) for _ in range(self.batch_size)]
+        else:
+            # make sure that inputs increase in size
+            speech_inputs = [
+                floats_list((x, self.feature_size))
+                for x in range(self.min_seq_length, self.max_seq_length, self.seq_length_diff)
+            ]
+        if numpify:
+            speech_inputs = [np.asarray(x) for x in speech_inputs]
+        return speech_inputs
+
+
+@require_torch
+@require_torchaudio
+class MCTCTFeatureExtractionTest(SequenceFeatureExtractionTestMixin, unittest.TestCase):
+
+    feature_extraction_class = MCTCTFeatureExtractor if is_speech_available() else None
+
+    def setUp(self):
+        self.feat_extract_tester = MCTCTFeatureExtractionTester(self)
+
+    def _check_zero_mean_unit_variance(self, input_vector):
+        self.assertTrue(np.all(np.mean(input_vector) < 1e-3))
+        self.assertTrue(np.all(np.abs(np.var(input_vector) - 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())
+        # create three inputs of length 800, 1000, and 12000
+        speech_inputs = [floats_list((1, x))[0] for x in range(8000, 14000, 2000)]
+        np_speech_inputs = [np.asarray(speech_input) for speech_input in speech_inputs]
+
+        # Test feature size
+        input_features = feature_extractor(np_speech_inputs, padding=True, return_tensors="np").input_features
+        self.assertTrue(input_features.ndim == 3)
+        self.assertTrue(input_features.shape[-1] == feature_extractor.feature_size)
+
+        # Test not batched input
+        encoded_sequences_1 = feature_extractor(speech_inputs[0], return_tensors="np").input_features
+        encoded_sequences_2 = feature_extractor(np_speech_inputs[0], return_tensors="np").input_features
+        self.assertTrue(np.allclose(encoded_sequences_1, encoded_sequences_2, atol=1e-3))
+
+        # Test batched
+        encoded_sequences_1 = feature_extractor(speech_inputs, return_tensors="np").input_features
+        encoded_sequences_2 = feature_extractor(np_speech_inputs, return_tensors="np").input_features
+        for enc_seq_1, enc_seq_2 in zip(encoded_sequences_1, encoded_sequences_2):
+            self.assertTrue(np.allclose(enc_seq_1, enc_seq_2, atol=1e-3))
+
+    def test_cepstral_mean_and_variance_normalization(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(8000, 14000, 2000)]
+
+        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,
+                truncation=max_length is not None,  # reference to #16419
+            )
+            input_features = inputs.input_features
+            attention_mask = inputs.attention_mask
+            fbank_feat_lengths = [np.sum(x) for x in attention_mask]
+            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(8000, 14000, 2000)]
+
+        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,
+                truncation=max_length is not None,
+            )
+            input_features = inputs.input_features
+            attention_mask = inputs.attention_mask
+            fbank_feat_lengths = [np.sum(x) for x in attention_mask]
+
+            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)
+            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)
+            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())
+        speech_inputs = [floats_list((1, x))[0] for x in range(8000, 14000, 2000)]
+        inputs = feature_extractor(
+            speech_inputs,
+            padding="max_length",
+            max_length=4,
+            truncation=True,
+            return_tensors="np",
+            return_attention_mask=True,
+        )
+        input_features = inputs.input_features
+        attention_mask = inputs.attention_mask
+        fbank_feat_lengths = np.sum(attention_mask == 1, axis=1)
+
+        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())
+        speech_inputs = [floats_list((1, x))[0] for x in range(8000, 14000, 2000)]
+        inputs = feature_extractor(
+            speech_inputs,
+            padding="longest",
+            max_length=4,
+            truncation=True,
+            return_tensors="np",
+            return_attention_mask=True,
+        )
+        input_features = inputs.input_features
+        attention_mask = inputs.attention_mask
+        fbank_feat_lengths = np.sum(attention_mask == 1, axis=1)
+
+        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))
+
+        speech_inputs = [floats_list((1, x))[0] for x in range(8000, 14000, 2000)]
+        inputs = feature_extractor(
+            speech_inputs,
+            padding="longest",
+            max_length=16,
+            truncation=True,
+            return_tensors="np",
+            return_attention_mask=True,
+        )
+        input_features = inputs.input_features
+        attention_mask = inputs.attention_mask
+        fbank_feat_lengths = np.sum(attention_mask == 1, axis=1)
+
+        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, 16, 24))
+
+    def test_double_precision_pad(self):
+        import torch
+
+        feature_extractor = self.feature_extraction_class(**self.feat_extract_tester.prepare_feat_extract_dict())
+        np_speech_inputs = np.random.rand(100, 32).astype(np.float64)
+        py_speech_inputs = np_speech_inputs.tolist()
+
+        for inputs in [py_speech_inputs, np_speech_inputs]:
+            np_processed = feature_extractor.pad([{"input_features": inputs}], return_tensors="np")
+            self.assertTrue(np_processed.input_features.dtype == np.float32)
+            pt_processed = feature_extractor.pad([{"input_features": inputs}], return_tensors="pt")
+            self.assertTrue(pt_processed.input_features.dtype == torch.float32)
+
+    def test_different_window(self):
+        import torch
+
+        init_dict = self.feat_extract_tester.prepare_feat_extract_dict()
+        init_dict["win_function"] = "hann_window"
+
+        feature_extractor = self.feature_extraction_class(**init_dict)
+        np_speech_inputs = np.random.rand(100, 32).astype(np.float64)
+        py_speech_inputs = np_speech_inputs.tolist()
+
+        for inputs in [py_speech_inputs, np_speech_inputs]:
+            np_processed = feature_extractor.pad([{"input_features": inputs}], return_tensors="np")
+            self.assertTrue(np_processed.input_features.dtype == np.float32)
+            pt_processed = feature_extractor.pad([{"input_features": inputs}], return_tensors="pt")
+            self.assertTrue(pt_processed.input_features.dtype == torch.float32)

tests/models/mctct/test_processor_mctct.py

+# Copyright 2022 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import json
+import os
+import shutil
+import tempfile
+import unittest
+
+from transformers import MCTCTProcessor, is_speech_available, is_torch_available
+from transformers.file_utils import FEATURE_EXTRACTOR_NAME
+from transformers.models.wav2vec2.tokenization_wav2vec2 import VOCAB_FILES_NAMES, Wav2Vec2CTCTokenizer
+from transformers.testing_utils import require_torch, require_torchaudio
+
+
+if is_speech_available() and is_torch_available():
+    from transformers import MCTCTFeatureExtractor
+
+    from .test_feature_extraction_mctct import floats_list
+
+
+@require_torch
+@require_torchaudio
+class MCTCTProcessorTest(unittest.TestCase):
+    def setUp(self):
+        vocab = "<pad> <s> </s> <unk> | E T A O N I H S R D L U M W C F G Y P B V K ' X J Q Z".split(" ")
+        vocab_tokens = dict(zip(vocab, range(len(vocab))))
+
+        self.add_kwargs_tokens_map = {
+            "pad_token": "<pad>",
+            "unk_token": "<unk>",
+            "bos_token": "<s>",
+            "eos_token": "</s>",
+        }
+        feature_extractor_map = {
+            "feature_size": 1,
+            "padding_value": 0.0,
+            "sampling_rate": 16000,
+            "return_attention_mask": False,
+            "do_normalize": True,
+        }
+
+        self.tmpdirname = tempfile.mkdtemp()
+        self.vocab_file = os.path.join(self.tmpdirname, VOCAB_FILES_NAMES["vocab_file"])
+        self.feature_extraction_file = os.path.join(self.tmpdirname, FEATURE_EXTRACTOR_NAME)
+        with open(self.vocab_file, "w", encoding="utf-8") as fp:
+            fp.write(json.dumps(vocab_tokens) + "\n")
+
+        with open(self.feature_extraction_file, "w", encoding="utf-8") as fp:
+            fp.write(json.dumps(feature_extractor_map) + "\n")
+
+    def get_tokenizer(self, **kwargs_init):
+        kwargs = self.add_kwargs_tokens_map.copy()
+        kwargs.update(kwargs_init)
+        return Wav2Vec2CTCTokenizer.from_pretrained(self.tmpdirname, **kwargs)
+
+    def get_feature_extractor(self, **kwargs):
+        return MCTCTFeatureExtractor.from_pretrained(self.tmpdirname, **kwargs)
+
+    def tearDown(self):
+        shutil.rmtree(self.tmpdirname)
+
+    def test_save_load_pretrained_default(self):
+        tokenizer = self.get_tokenizer()
+        feature_extractor = self.get_feature_extractor()
+
+        processor = MCTCTProcessor(tokenizer=tokenizer, feature_extractor=feature_extractor)
+
+        processor.save_pretrained(self.tmpdirname)
+        processor = MCTCTProcessor.from_pretrained(self.tmpdirname)
+
+        self.assertEqual(processor.tokenizer.get_vocab(), tokenizer.get_vocab())
+        self.assertIsInstance(processor.tokenizer, Wav2Vec2CTCTokenizer)
+
+        self.assertEqual(processor.feature_extractor.to_json_string(), feature_extractor.to_json_string())
+        self.assertIsInstance(processor.feature_extractor, MCTCTFeatureExtractor)
+
+    def test_save_load_pretrained_additional_features(self):
+        processor = MCTCTProcessor(tokenizer=self.get_tokenizer(), feature_extractor=self.get_feature_extractor())
+        processor.save_pretrained(self.tmpdirname)
+
+        tokenizer_add_kwargs = self.get_tokenizer(bos_token="(BOS)", eos_token="(EOS)")
+        feature_extractor_add_kwargs = self.get_feature_extractor(do_normalize=False, padding_value=1.0)
+
+        processor = MCTCTProcessor.from_pretrained(
+            self.tmpdirname, bos_token="(BOS)", eos_token="(EOS)", do_normalize=False, padding_value=1.0
+        )
+
+        self.assertEqual(processor.tokenizer.get_vocab(), tokenizer_add_kwargs.get_vocab())
+        self.assertIsInstance(processor.tokenizer, Wav2Vec2CTCTokenizer)
+
+        self.assertEqual(processor.feature_extractor.to_json_string(), feature_extractor_add_kwargs.to_json_string())
+        self.assertIsInstance(processor.feature_extractor, MCTCTFeatureExtractor)
+
+    def test_feature_extractor(self):
+        feature_extractor = self.get_feature_extractor()
+        tokenizer = self.get_tokenizer()
+
+        processor = MCTCTProcessor(tokenizer=tokenizer, feature_extractor=feature_extractor)
+
+        raw_speech = floats_list((3, 1000))
+
+        input_feat_extract = feature_extractor(raw_speech, return_tensors="np")
+        input_processor = processor(raw_speech, return_tensors="np")
+
+        for key in input_feat_extract.keys():
+            self.assertAlmostEqual(input_feat_extract[key].sum(), input_processor[key].sum(), delta=1e-2)
+
+    def test_tokenizer(self):
+        feature_extractor = self.get_feature_extractor()
+        tokenizer = self.get_tokenizer()
+
+        processor = MCTCTProcessor(tokenizer=tokenizer, feature_extractor=feature_extractor)
+
+        input_str = "This is a test string"
+
+        with processor.as_target_processor():
+            encoded_processor = processor(input_str)
+
+        encoded_tok = tokenizer(input_str)
+
+        for key in encoded_tok.keys():
+            self.assertListEqual(encoded_tok[key], encoded_processor[key])
+
+    def test_tokenizer_decode(self):
+        feature_extractor = self.get_feature_extractor()
+        tokenizer = self.get_tokenizer()
+
+        processor = MCTCTProcessor(tokenizer=tokenizer, feature_extractor=feature_extractor)
+
+        predicted_ids = [[1, 4, 5, 8, 1, 0, 8], [3, 4, 3, 1, 1, 8, 9]]
+
+        decoded_processor = processor.batch_decode(predicted_ids)
+        decoded_tok = tokenizer.batch_decode(predicted_ids)
+
+        self.assertListEqual(decoded_tok, decoded_processor)

utils/check_repo.py

@@ -59,6 +59,7 @@ IGNORE_NON_TESTED = PRIVATE_MODELS.copy() + [
     "DetrDecoderWrapper",  # Building part of bigger (tested) model.
     "M2M100Encoder",  # Building part of bigger (tested) model.
     "M2M100Decoder",  # Building part of bigger (tested) model.
+    "MCTCTEncoder",  # Building part of bigger (tested) model.
     "Speech2TextEncoder",  # Building part of bigger (tested) model.
     "Speech2TextDecoder",  # Building part of bigger (tested) model.
     "LEDEncoder",  # Building part of bigger (tested) model.