Wav2Vec2 (#9659)

* add raw scaffold * implement feat extract layers * make style * remove + * correctly convert weights * make feat extractor work * make feature extraction proj work * run forward pass * finish forward pass * Succesful decoding example * remove unused files * more changes * add wav2vec tokenizer * add new structure * fix run forward * add other layer norm architecture * finish 2nd structure * add model tests * finish tests for tok and model * clean-up * make style * finish docstring for model and config * make style * correct docstring * correct tests * change checkpoints to fairseq * fix examples * finish wav2vec2 * make style * apply sylvains suggestions * apply lysandres suggestions * change print to log.info * re-add assert statement * add input_values as required input name * finish wav2vec2 tokenizer * Update tests/test_tokenization_wav2vec2.py Co-authored-by: Lysandre Debut <lysandre@huggingface.co> * apply sylvains suggestions Co-authored-by: Lysandre Debut <lysandre@huggingface.co>
2021-02-02 15:52:10 +03:00
parent d996024af7
commit d6217fb30c
20 changed files with 2233 additions and 5 deletions
--- a/tests/test_modeling_wav2vec2.py
+++ b/tests/test_modeling_wav2vec2.py
@@ -0,0 +1,354 @@
+# coding=utf-8
+# Copyright 2021 The HuggingFace Inc. 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.
+""" Testing suite for the PyTorch Wav2Vec2 model. """
+
+
+import math
+import unittest
+
+from tests.test_modeling_common import floats_tensor
+from transformers import is_torch_available
+from transformers.testing_utils import require_datasets, require_soundfile, require_torch, slow, torch_device
+
+from .test_configuration_common import ConfigTester
+from .test_modeling_common import ModelTesterMixin, _config_zero_init
+
+
+if is_torch_available():
+    import torch
+
+    from transformers import Wav2Vec2Config, Wav2Vec2ForMaskedLM, Wav2Vec2Model, Wav2Vec2Tokenizer
+
+
+class Wav2Vec2ModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=13,
+        seq_length=1024,  # speech is longer
+        is_training=False,
+        hidden_size=16,
+        feat_extract_norm="group",
+        feat_extract_dropout=0.0,
+        feat_extract_activation="gelu",
+        conv_dim=(32, 32, 32),
+        conv_stride=(4, 4, 4),
+        conv_kernel=(8, 8, 8),
+        conv_bias=False,
+        num_conv_pos_embeddings=16,
+        num_conv_pos_embedding_groups=2,
+        num_hidden_layers=4,
+        num_attention_heads=2,
+        hidden_dropout_prob=0.1,  # this is most likely not correctly set yet
+        intermediate_size=20,
+        layer_norm_eps=1e-5,
+        hidden_act="gelu",
+        initializer_range=0.02,
+        vocab_size=32,
+        do_stable_layer_norm=False,
+        scope=None,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.seq_length = seq_length
+        self.is_training = is_training
+        self.hidden_size = hidden_size
+        self.feat_extract_norm = feat_extract_norm
+        self.feat_extract_dropout = feat_extract_dropout
+        self.feat_extract_activation = feat_extract_activation
+        self.conv_dim = conv_dim
+        self.conv_stride = conv_stride
+        self.conv_kernel = conv_kernel
+        self.conv_bias = conv_bias
+        self.num_conv_pos_embeddings = num_conv_pos_embeddings
+        self.num_conv_pos_embedding_groups = num_conv_pos_embedding_groups
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.intermediate_size = intermediate_size
+        self.layer_norm_eps = layer_norm_eps
+        self.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+        self.vocab_size = vocab_size
+        self.do_stable_layer_norm = do_stable_layer_norm
+        self.scope = scope
+
+        output_seq_length = self.seq_length
+        for kernel, stride in zip(self.conv_kernel, self.conv_stride):
+            output_seq_length = (output_seq_length - (kernel - 1)) / stride
+        self.output_seq_length = int(math.ceil(output_seq_length))
+        self.encoder_seq_length = self.output_seq_length
+
+    def prepare_config_and_inputs(self):
+        input_values = floats_tensor([self.batch_size, self.seq_length], self.vocab_size)
+
+        config = Wav2Vec2Config(
+            hidden_size=self.hidden_size,
+            feat_extract_norm=self.feat_extract_norm,
+            feat_extract_dropout=self.feat_extract_dropout,
+            feat_extract_activation=self.feat_extract_activation,
+            conv_dim=self.conv_dim,
+            conv_stride=self.conv_stride,
+            conv_kernel=self.conv_kernel,
+            conv_bias=self.conv_bias,
+            num_conv_pos_embeddings=self.num_conv_pos_embeddings,
+            num_conv_pos_embedding_groups=self.num_conv_pos_embedding_groups,
+            num_hidden_layers=self.num_hidden_layers,
+            num_attention_heads=self.num_attention_heads,
+            hidden_dropout_prob=self.hidden_dropout_prob,
+            intermediate_size=self.intermediate_size,
+            layer_norm_eps=self.layer_norm_eps,
+            hidden_act=self.hidden_act,
+            initializer_range=self.initializer_range,
+            vocab_size=self.vocab_size,
+        )
+
+        return config, input_values
+
+    def create_and_check_model(self, config, input_values):
+        model = Wav2Vec2Model(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(input_values)
+        self.parent.assertEqual(
+            result.last_hidden_state.shape, (self.batch_size, self.output_seq_length, self.hidden_size)
+        )
+
+    def prepare_config_and_inputs_for_common(self):
+        config, input_values = self.prepare_config_and_inputs()
+        inputs_dict = {"input_values": input_values}
+        return config, inputs_dict
+
+
+@require_torch
+class Wav2Vec2ModelTest(ModelTesterMixin, unittest.TestCase):
+    all_model_classes = (
+        (
+            Wav2Vec2Model,
+            Wav2Vec2ForMaskedLM,
+        )
+        if is_torch_available()
+        else ()
+    )
+    test_pruning = False
+    test_headmasking = False
+    test_torchscript = False
+
+    def setUp(self):
+        self.model_tester = Wav2Vec2ModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=Wav2Vec2Config, hidden_size=37)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_model(*config_and_inputs)
+
+    # Wav2Vec2 has no inputs_embeds
+    def test_inputs_embeds(self):
+        pass
+
+    # `input_ids` is renamed to `input_values`
+    def test_forward_signature(self):
+        pass
+
+    # Wav2Vec2 cannot resize token embeddings
+    # since it has no tokens embeddings
+    def test_resize_tokens_embeddings(self):
+        pass
+
+    # Wav2Vec2 has no inputs_embeds
+    # and thus the `get_input_embeddings` fn
+    # is not implemented
+    def test_model_common_attributes(self):
+        pass
+
+    def test_initialization(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        configs_no_init = _config_zero_init(config)
+        for model_class in self.all_model_classes:
+            model = model_class(config=configs_no_init)
+            for name, param in model.named_parameters():
+                if param.requires_grad:
+                    if "conv.weight" in name:
+                        self.assertTrue(
+                            -1.0 <= ((param.data.mean() * 1e9).round() / 1e9).item() <= 1.0,
+                            msg="Parameter {} of model {} seems not properly initialized".format(name, model_class),
+                        )
+                    else:
+                        self.assertIn(
+                            ((param.data.mean() * 1e9).round() / 1e9).item(),
+                            [0.0, 1.0],
+                            msg="Parameter {} of model {} seems not properly initialized".format(name, model_class),
+                        )
+
+    @slow
+    def test_model_from_pretrained(self):
+        model = Wav2Vec2Model.from_pretrained("facebook/wav2vec2-base-960h")
+        self.assertIsNotNone(model)
+
+
+@require_torch
+class Wav2Vec2RobustModelTest(ModelTesterMixin, unittest.TestCase):
+    all_model_classes = (Wav2Vec2Model, Wav2Vec2ForMaskedLM) if is_torch_available() else ()
+    test_pruning = False
+    test_headmasking = False
+    test_torchscript = False
+
+    def setUp(self):
+        self.model_tester = Wav2Vec2ModelTester(
+            self, conv_stride=(3, 3, 3), feat_extract_norm="layer", do_stable_layer_norm=True
+        )
+        self.config_tester = ConfigTester(self, config_class=Wav2Vec2Config, hidden_size=37)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_model(*config_and_inputs)
+
+    # Wav2Vec2 has no inputs_embeds
+    def test_inputs_embeds(self):
+        pass
+
+    # `input_ids` is renamed to `input_values`
+    def test_forward_signature(self):
+        pass
+
+    # Wav2Vec2 cannot resize token embeddings
+    # since it has no tokens embeddings
+    def test_resize_tokens_embeddings(self):
+        pass
+
+    # Wav2Vec2 has no inputs_embeds
+    # and thus the `get_input_embeddings` fn
+    # is not implemented
+    def test_model_common_attributes(self):
+        pass
+
+    def test_initialization(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        configs_no_init = _config_zero_init(config)
+        for model_class in self.all_model_classes:
+            model = model_class(config=configs_no_init)
+            for name, param in model.named_parameters():
+                if param.requires_grad:
+                    if "conv.weight" in name:
+                        self.assertTrue(
+                            -1.0 <= ((param.data.mean() * 1e9).round() / 1e9).item() <= 1.0,
+                            msg="Parameter {} of model {} seems not properly initialized".format(name, model_class),
+                        )
+                    else:
+                        self.assertIn(
+                            ((param.data.mean() * 1e9).round() / 1e9).item(),
+                            [0.0, 1.0],
+                            msg="Parameter {} of model {} seems not properly initialized".format(name, model_class),
+                        )
+
+    @slow
+    def test_model_from_pretrained(self):
+        model = Wav2Vec2Model.from_pretrained("facebook/wav2vec2-base-960h")
+        self.assertIsNotNone(model)
+
+
+@require_torch
+@slow
+@require_datasets
+@require_soundfile
+class Wav2Vec2ModelIntegrationTest(unittest.TestCase):
+    def _load_datasamples(self, num_samples):
+        from datasets import load_dataset
+
+        import soundfile as sf
+
+        # map files to raw
+        def map_to_array(batch):
+            speech, _ = sf.read(batch["file"])
+            batch["speech"] = speech
+            return batch
+
+        ds = load_dataset("patrickvonplaten/librispeech_asr_dummy", "clean", split="validation")
+        ds = ds.select(range(num_samples)).map(map_to_array)
+
+        return ds["speech"][:num_samples]
+
+    def test_inference_masked_lm_normal(self):
+        model = Wav2Vec2ForMaskedLM.from_pretrained("facebook/wav2vec2-base-960h")
+        model.to(torch_device)
+        tokenizer = Wav2Vec2Tokenizer.from_pretrained("facebook/wav2vec2-base-960h", do_lower_case=True)
+
+        input_speech = self._load_datasamples(1)
+
+        input_values = tokenizer(input_speech, return_tensors="pt").input_values.to(torch_device)
+
+        with torch.no_grad():
+            logits = model(input_values).logits
+
+        predicted_ids = torch.argmax(logits, dim=-1)
+        predicted_trans = tokenizer.batch_decode(predicted_ids)
+
+        EXPECTED_TRANSCRIPTIONS = ["a man said to the universe sir i exist"]
+        self.assertListEqual(predicted_trans, EXPECTED_TRANSCRIPTIONS)
+
+    def test_inference_masked_lm_normal_batched(self):
+        model = Wav2Vec2ForMaskedLM.from_pretrained("facebook/wav2vec2-base-960h")
+        model.to(torch_device)
+        tokenizer = Wav2Vec2Tokenizer.from_pretrained("facebook/wav2vec2-base-960h", do_lower_case=True)
+
+        input_speech = self._load_datasamples(2)
+
+        input_values = tokenizer(input_speech, return_tensors="pt", padding=True, truncation=True).input_values.to(
+            torch_device
+        )
+
+        with torch.no_grad():
+            logits = model(input_values).logits
+
+        predicted_ids = torch.argmax(logits, dim=-1)
+        predicted_trans = tokenizer.batch_decode(predicted_ids)
+
+        EXPECTED_TRANSCRIPTIONS = [
+            "a man said to the universe sir i exist",
+            "sweat covered brion's body trickling into the tight lowing cloth that was the only garment he wore",
+        ]
+        self.assertListEqual(predicted_trans, EXPECTED_TRANSCRIPTIONS)
+
+    def test_inference_masked_lm_robust_batched(self):
+        model = Wav2Vec2ForMaskedLM.from_pretrained("facebook/wav2vec2-large-960h-lv60-self").to(torch_device)
+        tokenizer = Wav2Vec2Tokenizer.from_pretrained("facebook/wav2vec2-large-960h-lv60-self", do_lower_case=True)
+
+        input_speech = self._load_datasamples(4)
+
+        input_values = tokenizer(input_speech, return_tensors="pt", padding=True, truncation=True).input_values.to(
+            torch_device
+        )
+
+        with torch.no_grad():
+            logits = model(input_values).logits
+
+        predicted_ids = torch.argmax(logits, dim=-1)
+        predicted_trans = tokenizer.batch_decode(predicted_ids)
+
+        EXPECTED_TRANSCRIPTIONS = [
+            "a man said to the universe sir i exist",
+            "sweat covered brion's body trickling into the tight loin 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 panic was followed by a small sharp blow high on his chest",
+        ]
+        self.assertListEqual(predicted_trans, EXPECTED_TRANSCRIPTIONS)
--- a/tests/test_tokenization_wav2vec2.py
+++ b/tests/test_tokenization_wav2vec2.py
@@ -0,0 +1,301 @@
+# coding=utf-8
+# Copyright 2021 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.
+"""Tests for the Wav2Vec2 tokenizer."""
+import inspect
+import json
+import os
+import random
+import shutil
+import tempfile
+import unittest
+
+import numpy as np
+
+from transformers.models.wav2vec2.tokenization_wav2vec2 import VOCAB_FILES_NAMES, Wav2Vec2Tokenizer
+
+
+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
+
+
+class Wav2Vec2TokenizerTest(unittest.TestCase):
+    tokenizer_class = Wav2Vec2Tokenizer
+
+    def setUp(self):
+        super().setUp()
+
+        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.special_tokens_map = {"pad_token": "<pad>", "unk_token": "<unk>", "bos_token": "<s>", "eos_token": "</s>"}
+
+        self.tmpdirname = tempfile.mkdtemp()
+        self.vocab_file = os.path.join(self.tmpdirname, VOCAB_FILES_NAMES["vocab_file"])
+        with open(self.vocab_file, "w", encoding="utf-8") as fp:
+            fp.write(json.dumps(vocab_tokens) + "\n")
+
+    def get_tokenizer(self, **kwargs):
+        kwargs.update(self.special_tokens_map)
+        return Wav2Vec2Tokenizer.from_pretrained(self.tmpdirname, **kwargs)
+
+    def test_tokenizer_decode(self):
+        # TODO(PVP) - change to facebook
+        tokenizer = Wav2Vec2Tokenizer.from_pretrained("facebook/wav2vec2-base-960h")
+
+        sample_ids = [
+            [11, 5, 15, tokenizer.pad_token_id, 15, 8, 98],
+            [24, 22, 5, tokenizer.word_delimiter_token_id, 24, 22, 5, 77],
+        ]
+        tokens = tokenizer.decode(sample_ids[0])
+        batch_tokens = tokenizer.batch_decode(sample_ids)
+        self.assertEqual(tokens, batch_tokens[0])
+        self.assertEqual(batch_tokens, ["HELLO<unk>", "BYE BYE<unk>"])
+
+    def test_tokenizer_decode_special(self):
+        # TODO(PVP) - change to facebook
+        tokenizer = Wav2Vec2Tokenizer.from_pretrained("facebook/wav2vec2-base-960h")
+
+        sample_ids = [
+            [11, 5, 15, tokenizer.pad_token_id, 15, 8, 98],
+            [24, 22, 5, tokenizer.word_delimiter_token_id, 24, 22, 5, 77],
+        ]
+        sample_ids_2 = [
+            [11, 5, 5, 5, 5, 5, 15, 15, 15, tokenizer.pad_token_id, 15, 8, 98],
+            [
+                24,
+                22,
+                5,
+                tokenizer.pad_token_id,
+                tokenizer.pad_token_id,
+                tokenizer.pad_token_id,
+                tokenizer.word_delimiter_token_id,
+                24,
+                22,
+                5,
+                77,
+                tokenizer.word_delimiter_token_id,
+            ],
+        ]
+
+        batch_tokens = tokenizer.batch_decode(sample_ids)
+        batch_tokens_2 = tokenizer.batch_decode(sample_ids_2)
+        self.assertEqual(batch_tokens, batch_tokens_2)
+        self.assertEqual(batch_tokens, ["HELLO<unk>", "BYE BYE<unk>"])
+
+    def test_tokenizer_decode_added_tokens(self):
+        tokenizer = Wav2Vec2Tokenizer.from_pretrained("facebook/wav2vec2-base-960h")
+        tokenizer.add_tokens(["!", "?"])
+        tokenizer.add_special_tokens({"cls_token": "$$$"})
+
+        sample_ids = [
+            [
+                11,
+                5,
+                15,
+                tokenizer.pad_token_id,
+                15,
+                8,
+                98,
+                32,
+                32,
+                33,
+                tokenizer.word_delimiter_token_id,
+                32,
+                32,
+                33,
+                34,
+                34,
+            ],
+            [24, 22, 5, tokenizer.word_delimiter_token_id, 24, 22, 5, 77, tokenizer.pad_token_id, 34, 34],
+        ]
+        batch_tokens = tokenizer.batch_decode(sample_ids)
+
+        self.assertEqual(batch_tokens, ["HELLO<unk>!?!?$$$", "BYE BYE<unk>$$$"])
+
+    def test_call(self):
+        # Tests that all call wrap to encode_plus and batch_encode_plus
+        tokenizer = self.get_tokenizer()
+        # create three inputs of length 800, 1000, and 1200
+        speech_inputs = [floats_list((1, x))[0] for x in range(800, 1400, 200)]
+        np_speech_inputs = [np.asarray(speech_input) for speech_input in speech_inputs]
+
+        # Test not batched input
+        encoded_sequences_1 = tokenizer(speech_inputs[0], return_tensors="np").input_values
+        encoded_sequences_2 = tokenizer(np_speech_inputs[0], return_tensors="np").input_values
+        self.assertTrue(np.allclose(encoded_sequences_1, encoded_sequences_2, atol=1e-3))
+
+        # Test batched
+        encoded_sequences_1 = tokenizer(speech_inputs, return_tensors="np").input_values
+        encoded_sequences_2 = tokenizer(np_speech_inputs, return_tensors="np").input_values
+        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_padding(self, max_length=50):
+        def _input_values_have_equal_length(input_values):
+            length = len(input_values[0])
+            for input_values_slice in input_values[1:]:
+                if len(input_values_slice) != length:
+                    return False
+            return True
+
+        def _input_values_are_equal(input_values_1, input_values_2):
+            if len(input_values_1) != len(input_values_2):
+                return False
+
+            for input_values_slice_1, input_values_slice_2 in zip(input_values_1, input_values_2):
+                if not np.allclose(np.asarray(input_values_slice_1), np.asarray(input_values_slice_2), atol=1e-3):
+                    return False
+            return True
+
+        tokenizer = self.get_tokenizer()
+        speech_inputs = [floats_list((1, x))[0] for x in range(800, 1400, 200)]
+
+        input_values_1 = tokenizer(speech_inputs).input_values
+        input_values_2 = tokenizer(speech_inputs, padding="longest").input_values
+        input_values_3 = tokenizer(speech_inputs, padding="longest", max_length=1600).input_values
+
+        self.assertFalse(_input_values_have_equal_length(input_values_1))
+        self.assertTrue(_input_values_have_equal_length(input_values_2))
+        self.assertTrue(_input_values_have_equal_length(input_values_3))
+        self.assertTrue(_input_values_are_equal(input_values_2, input_values_3))
+        self.assertTrue(len(input_values_1[0]) == 800)
+        self.assertTrue(len(input_values_2[0]) == 1200)
+        # padding should be 0.0
+        self.assertTrue(abs(sum(np.asarray(input_values_2[0])[800:])) < 1e-3)
+        self.assertTrue(abs(sum(np.asarray(input_values_2[1])[1000:])) < 1e-3)
+
+        input_values_4 = tokenizer(speech_inputs, padding="max_length").input_values
+        input_values_5 = tokenizer(speech_inputs, padding="max_length", max_length=1600).input_values
+
+        self.assertTrue(_input_values_are_equal(input_values_1, input_values_4))
+        self.assertTrue(input_values_5.shape, (3, 1600))
+        # padding should be 0.0
+        self.assertTrue(abs(sum(np.asarray(input_values_5[0])[800:1200])) < 1e-3)
+
+        input_values_6 = tokenizer(speech_inputs, pad_to_multiple_of=500).input_values
+        input_values_7 = tokenizer(speech_inputs, padding="longest", pad_to_multiple_of=500).input_values
+        input_values_8 = tokenizer(
+            speech_inputs, padding="max_length", pad_to_multiple_of=500, max_length=2400
+        ).input_values
+
+        self.assertTrue(_input_values_are_equal(input_values_1, input_values_6))
+        self.assertTrue(input_values_7.shape, (3, 1500))
+        self.assertTrue(input_values_8.shape, (3, 2500))
+        # padding should be 0.0
+        self.assertTrue(abs(sum(np.asarray(input_values_7[0])[800:])) < 1e-3)
+        self.assertTrue(abs(sum(np.asarray(input_values_7[1])[1000:])) < 1e-3)
+        self.assertTrue(abs(sum(np.asarray(input_values_7[2])[1200:])) < 1e-3)
+        self.assertTrue(abs(sum(np.asarray(input_values_8[0])[800:])) < 1e-3)
+        self.assertTrue(abs(sum(np.asarray(input_values_8[1])[1000:])) < 1e-3)
+        self.assertTrue(abs(sum(np.asarray(input_values_8[2])[1200:])) < 1e-3)
+
+    def test_save_pretrained(self):
+        pretrained_name = list(self.tokenizer_class.pretrained_vocab_files_map["vocab_file"].keys())[0]
+        tokenizer = self.tokenizer_class.from_pretrained(pretrained_name)
+        tmpdirname2 = tempfile.mkdtemp()
+
+        tokenizer_files = tokenizer.save_pretrained(tmpdirname2)
+        self.assertSequenceEqual(
+            sorted(tuple(VOCAB_FILES_NAMES.values()) + ("special_tokens_map.json", "added_tokens.json")),
+            sorted(tuple(x.split("/")[-1] for x in tokenizer_files)),
+        )
+
+        # Checks everything loads correctly in the same way
+        tokenizer_p = self.tokenizer_class.from_pretrained(tmpdirname2)
+
+        # Check special tokens are set accordingly on Rust and Python
+        for key in tokenizer.special_tokens_map:
+            self.assertTrue(key in tokenizer_p.special_tokens_map)
+
+        shutil.rmtree(tmpdirname2)
+
+    def test_get_vocab(self):
+        tokenizer = self.get_tokenizer()
+        vocab_dict = tokenizer.get_vocab()
+        self.assertIsInstance(vocab_dict, dict)
+        self.assertGreaterEqual(len(tokenizer), len(vocab_dict))
+
+        vocab = [tokenizer.convert_ids_to_tokens(i) for i in range(len(tokenizer))]
+        self.assertEqual(len(vocab), len(tokenizer))
+
+        tokenizer.add_tokens(["asdfasdfasdfasdf"])
+        vocab = [tokenizer.convert_ids_to_tokens(i) for i in range(len(tokenizer))]
+        self.assertEqual(len(vocab), len(tokenizer))
+
+    def test_save_and_load_tokenizer(self):
+        tokenizer = self.get_tokenizer()
+        # Isolate this from the other tests because we save additional tokens/etc
+        tmpdirname = tempfile.mkdtemp()
+
+        sample_ids = [0, 1, 4, 8, 9, 0, 12]
+        before_tokens = tokenizer.decode(sample_ids)
+        before_vocab = tokenizer.get_vocab()
+        tokenizer.save_pretrained(tmpdirname)
+
+        after_tokenizer = tokenizer.__class__.from_pretrained(tmpdirname)
+        after_tokens = after_tokenizer.decode(sample_ids)
+        after_vocab = after_tokenizer.get_vocab()
+
+        self.assertEqual(before_tokens, after_tokens)
+        self.assertDictEqual(before_vocab, after_vocab)
+
+        shutil.rmtree(tmpdirname)
+
+        tokenizer = self.get_tokenizer()
+
+        # Isolate this from the other tests because we save additional tokens/etc
+        tmpdirname = tempfile.mkdtemp()
+
+        before_len = len(tokenizer)
+        sample_ids = [0, 1, 4, 8, 9, 0, 12, before_len, before_len + 1, before_len + 2]
+        tokenizer.add_tokens(["?", "!"])
+        additional_special_tokens = tokenizer.additional_special_tokens
+        additional_special_tokens.append("&")
+        tokenizer.add_special_tokens({"additional_special_tokens": additional_special_tokens})
+        before_tokens = tokenizer.decode(sample_ids)
+        before_vocab = tokenizer.get_vocab()
+        tokenizer.save_pretrained(tmpdirname)
+
+        after_tokenizer = tokenizer.__class__.from_pretrained(tmpdirname)
+        after_tokens = after_tokenizer.decode(sample_ids)
+        after_vocab = after_tokenizer.get_vocab()
+
+        self.assertEqual(before_tokens, after_tokens)
+        self.assertDictEqual(before_vocab, after_vocab)
+
+        self.assertTrue(len(tokenizer), before_len + 3)
+        self.assertTrue(len(tokenizer), len(after_tokenizer))
+        shutil.rmtree(tmpdirname)
+
+    def test_tokenizer_slow_store_full_signature(self):
+        signature = inspect.signature(self.tokenizer_class.__init__)
+        tokenizer = self.get_tokenizer()
+
+        for parameter_name, parameter in signature.parameters.items():
+            if parameter.default != inspect.Parameter.empty:
+                self.assertIn(parameter_name, tokenizer.init_kwargs)