This commit is contained in:
2025-10-09 16:47:16 +08:00
parent c8feb4deb5
commit e27e3f16bb
5248 changed files with 1778505 additions and 0 deletions

View File

@@ -0,0 +1,356 @@
# Copyright 2021 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 os
import random
import tempfile
import unittest
import numpy as np
from transformers import Speech2TextFeatureExtractor
from transformers.testing_utils import check_json_file_has_correct_format, require_torch, require_torchaudio
from ...test_sequence_feature_extraction_common import SequenceFeatureExtractionTestMixin
global_rng = random.Random()
# Copied from tests.models.whisper.test_feature_extraction_whisper.floats_list
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 Speech2TextFeatureExtractionTester:
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 Speech2TextFeatureExtractionTest(SequenceFeatureExtractionTestMixin, unittest.TestCase):
feature_extraction_class = Speech2TextFeatureExtractor
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())
# 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 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))
# Test 2-D numpy arrays are batched.
speech_inputs = [floats_list((1, x))[0] for x in (800, 800, 800)]
np_speech_inputs = np.asarray(speech_inputs)
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_dither(self):
np.random.seed(42) # seed the dithering randn()
# Tests that features with and without little dithering are similar, but not the same
dict_no_dither = self.feat_extract_tester.prepare_feat_extract_dict()
dict_no_dither["dither"] = 0.0
dict_dither = self.feat_extract_tester.prepare_feat_extract_dict()
dict_dither["dither"] = 1.0
feature_extractor_no_dither = self.feature_extraction_class(**dict_no_dither)
feature_extractor_dither = self.feature_extraction_class(**dict_dither)
# 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]
# compute features
input_features_no_dither = feature_extractor_no_dither(
np_speech_inputs, padding=True, return_tensors="np", sampling_rate=dict_no_dither["sampling_rate"]
).input_features
input_features_dither = feature_extractor_dither(
np_speech_inputs, padding=True, return_tensors="np", sampling_rate=dict_dither["sampling_rate"]
).input_features
# test there is a difference between features (there's added noise to input signal)
diff = input_features_dither - input_features_no_dither
# features are not identical
self.assertTrue(np.abs(diff).mean() > 1e-5)
# features are not too different
self.assertTrue(np.abs(diff).mean() <= 1e-3)
self.assertTrue(np.abs(diff).max() <= 5e-2)
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(800, 1400, 200)]
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
)
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(800, 1400, 200)]
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
)
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(800, 1400, 200)]
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(800, 1400, 200)]
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(800, 1400, 200)]
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, 6, 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 _load_datasamples(self, num_samples):
from datasets import load_dataset
ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
# automatic decoding with librispeech
speech_samples = ds.sort("id")[:num_samples]["audio"]
return [x["array"] for x in speech_samples]
def test_integration(self):
# fmt: off
expected = np.array([
-1.5745, -1.7713, -1.7020, -1.6069, -1.2250, -1.1105, -0.9072, -0.8241,
-1.2310, -0.8098, -0.3320, -0.4101, -0.7985, -0.4996, -0.8213, -0.9128,
-1.0420, -1.1286, -1.0440, -0.7999, -0.8405, -1.2275, -1.5443, -1.4625,
])
# fmt: on
input_speech = self._load_datasamples(1)
feature_extractor = self.feature_extraction_class(**self.feat_extract_tester.prepare_feat_extract_dict())
input_features = feature_extractor(input_speech, return_tensors="pt").input_features
self.assertEqual(input_features.shape, (1, 584, 24))
self.assertTrue(np.allclose(input_features[0, 0, :30], expected, atol=1e-4))
def test_feat_extract_from_and_save_pretrained(self):
feat_extract_first = self.feature_extraction_class(**self.feat_extract_dict)
with tempfile.TemporaryDirectory() as tmpdirname:
saved_file = feat_extract_first.save_pretrained(tmpdirname)[0]
check_json_file_has_correct_format(saved_file)
feat_extract_second = self.feature_extraction_class.from_pretrained(tmpdirname)
dict_first = feat_extract_first.to_dict()
dict_second = feat_extract_second.to_dict()
self.assertDictEqual(dict_first, dict_second)
def test_feat_extract_to_json_file(self):
feat_extract_first = self.feature_extraction_class(**self.feat_extract_dict)
with tempfile.TemporaryDirectory() as tmpdirname:
json_file_path = os.path.join(tmpdirname, "feat_extract.json")
feat_extract_first.to_json_file(json_file_path)
feat_extract_second = self.feature_extraction_class.from_json_file(json_file_path)
dict_first = feat_extract_first.to_dict()
dict_second = feat_extract_second.to_dict()
self.assertEqual(dict_first, dict_second)
# exact same tests than before, except that we simulate that torchaudio is not available
@require_torch
@unittest.mock.patch(
"transformers.models.speech_to_text.feature_extraction_speech_to_text.is_speech_available", lambda: False
)
class Speech2TextFeatureExtractionWithoutTorchaudioTest(Speech2TextFeatureExtractionTest):
def test_using_audio_utils(self):
# Tests that it uses audio_utils instead of torchaudio
feat_extract = self.feature_extraction_class(**self.feat_extract_tester.prepare_feat_extract_dict())
self.assertTrue(hasattr(feat_extract, "window"))
self.assertTrue(hasattr(feat_extract, "mel_filters"))
from transformers.models.speech_to_text.feature_extraction_speech_to_text import is_speech_available
self.assertFalse(is_speech_available())

View File

@@ -0,0 +1,748 @@
# 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 Speech2Text model."""
import copy
import inspect
import os
import tempfile
import unittest
from datasets import load_dataset
from transformers import Speech2TextConfig
from transformers.testing_utils import (
is_torch_available,
require_sentencepiece,
require_tokenizers,
require_torch,
require_torch_fp16,
require_torchaudio,
torch_device,
)
from ...generation.test_utils import GenerationTesterMixin
from ...test_configuration_common import ConfigTester
from ...test_modeling_common import ModelTesterMixin, _config_zero_init, floats_tensor, ids_tensor
from ...test_pipeline_mixin import PipelineTesterMixin
if is_torch_available():
import torch
from transformers import Speech2TextForConditionalGeneration, Speech2TextModel, Speech2TextProcessor
from transformers.models.speech_to_text.modeling_speech_to_text import Speech2TextDecoder, Speech2TextEncoder
def prepare_speech_to_text_inputs_dict(
config,
input_features,
decoder_input_ids,
attention_mask=None,
decoder_attention_mask=None,
head_mask=None,
decoder_head_mask=None,
cross_attn_head_mask=None,
):
if attention_mask is None:
attention_mask = input_features.ne(0)
if decoder_attention_mask is None:
decoder_attention_mask = decoder_input_ids.ne(config.pad_token_id)
if head_mask is None:
head_mask = torch.ones(config.encoder_layers, config.encoder_attention_heads, device=torch_device)
if decoder_head_mask is None:
decoder_head_mask = torch.ones(config.decoder_layers, config.decoder_attention_heads, device=torch_device)
if cross_attn_head_mask is None:
cross_attn_head_mask = torch.ones(config.decoder_layers, config.decoder_attention_heads, device=torch_device)
return {
# "input_ids": input_features,
"input_features": input_features,
"decoder_input_ids": decoder_input_ids,
"attention_mask": attention_mask,
"decoder_attention_mask": attention_mask,
"head_mask": head_mask,
"decoder_head_mask": decoder_head_mask,
"cross_attn_head_mask": cross_attn_head_mask,
}
@require_torch
class Speech2TextModelTester:
def __init__(
self,
parent,
batch_size=13,
seq_length=7,
is_training=True,
use_labels=False,
vocab_size=99,
hidden_size=16,
num_hidden_layers=2,
num_attention_heads=4,
intermediate_size=4,
num_conv_layers=2,
conv_kernel_sizes=(5, 5),
conv_channels=32,
input_feat_per_channel=24,
input_channels=1,
hidden_act="relu",
hidden_dropout_prob=0.1,
attention_probs_dropout_prob=0.1,
max_position_embeddings=20,
max_source_positions=20,
max_target_positions=20,
eos_token_id=2,
pad_token_id=1,
bos_token_id=0,
):
self.parent = parent
self.batch_size = batch_size
self.seq_length = seq_length
self.is_training = is_training
self.use_labels = use_labels
self.vocab_size = vocab_size
self.hidden_size = hidden_size
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
self.intermediate_size = intermediate_size
self.num_conv_layers = num_conv_layers
self.conv_kernel_sizes = conv_kernel_sizes
self.conv_channels = conv_channels
self.input_feat_per_channel = input_feat_per_channel
self.input_channels = input_channels
self.hidden_act = hidden_act
self.hidden_dropout_prob = hidden_dropout_prob
self.attention_probs_dropout_prob = attention_probs_dropout_prob
self.max_position_embeddings = max_position_embeddings
self.max_source_positions = max_source_positions
self.max_target_positions = max_target_positions
self.eos_token_id = eos_token_id
self.pad_token_id = pad_token_id
self.bos_token_id = bos_token_id
def prepare_config_and_inputs(self):
input_features = floats_tensor(
[self.batch_size, self.seq_length, self.input_feat_per_channel], self.vocab_size
)
attention_mask = torch.ones([self.batch_size, self.seq_length], dtype=torch.long, device=torch_device)
decoder_input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size).clamp(2)
config = self.get_config()
inputs_dict = prepare_speech_to_text_inputs_dict(
config,
input_features=input_features,
decoder_input_ids=decoder_input_ids,
attention_mask=attention_mask,
)
return config, inputs_dict
def get_config(self):
return Speech2TextConfig(
vocab_size=self.vocab_size,
d_model=self.hidden_size,
encoder_layers=self.num_hidden_layers,
decoder_layers=self.num_hidden_layers,
encoder_attention_heads=self.num_attention_heads,
decoder_attention_heads=self.num_attention_heads,
encoder_ffn_dim=self.intermediate_size,
decoder_ffn_dim=self.intermediate_size,
num_conv_layers=self.num_conv_layers,
conv_kernel_sizes=self.conv_kernel_sizes,
conv_channels=self.conv_channels,
input_feat_per_channel=self.input_feat_per_channel,
input_channels=self.input_channels,
dropout=self.hidden_dropout_prob,
attention_dropout=self.attention_probs_dropout_prob,
max_position_embeddings=self.max_position_embeddings,
max_source_positions=self.max_source_positions,
max_target_positions=self.max_target_positions,
eos_token_id=self.eos_token_id,
bos_token_id=self.bos_token_id,
pad_token_id=self.pad_token_id,
)
def prepare_config_and_inputs_for_common(self):
config, inputs_dict = self.prepare_config_and_inputs()
return config, inputs_dict
def get_subsampled_output_lengths(self, input_lengths):
"""
Computes the output length of the convolutional layers
"""
for i in range(self.num_conv_layers):
input_lengths = (input_lengths - 1) // 2 + 1
return input_lengths
def create_and_check_model_forward(self, config, inputs_dict):
model = Speech2TextModel(config=config).to(torch_device).eval()
input_features = inputs_dict["input_features"]
decoder_input_ids = inputs_dict["decoder_input_ids"]
# first forward pass
last_hidden_state = model(input_features, decoder_input_ids=decoder_input_ids).last_hidden_state
self.parent.assertTrue(last_hidden_state.shape, (13, 7, 16))
def create_and_check_decoder_model_past_large_inputs(self, config, inputs_dict):
model = Speech2TextModel(config=config).get_decoder().to(torch_device).eval()
input_ids = inputs_dict["decoder_input_ids"]
attention_mask = inputs_dict["decoder_attention_mask"]
# first forward pass
outputs = model(input_ids, attention_mask=attention_mask, use_cache=True)
output, past_key_values = outputs.to_tuple()
# create hypothetical multiple next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size).clamp(2)
next_attn_mask = ids_tensor((self.batch_size, 3), 2)
# append to next input_ids and
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
next_attention_mask = torch.cat([attention_mask, next_attn_mask], dim=-1)
output_from_no_past = model(next_input_ids, attention_mask=next_attention_mask)["last_hidden_state"]
output_from_past = model(next_tokens, attention_mask=next_attention_mask, past_key_values=past_key_values)[
"last_hidden_state"
]
# select random slice
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
# test that outputs are equal for slice
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-2))
def check_encoder_decoder_model_standalone(self, config, inputs_dict):
model = Speech2TextModel(config=config).to(torch_device).eval()
outputs = model(**inputs_dict)
encoder_last_hidden_state = outputs.encoder_last_hidden_state
last_hidden_state = outputs.last_hidden_state
with tempfile.TemporaryDirectory() as tmpdirname:
encoder = model.get_encoder()
encoder.save_pretrained(tmpdirname)
encoder = Speech2TextEncoder.from_pretrained(tmpdirname).to(torch_device)
encoder_last_hidden_state_2 = encoder(
inputs_dict["input_features"], attention_mask=inputs_dict["attention_mask"]
)[0]
self.parent.assertTrue((encoder_last_hidden_state_2 - encoder_last_hidden_state).abs().max().item() < 1e-3)
with tempfile.TemporaryDirectory() as tmpdirname:
decoder = model.get_decoder()
decoder.save_pretrained(tmpdirname)
decoder = Speech2TextDecoder.from_pretrained(tmpdirname).to(torch_device)
encoder_attention_mask = encoder._get_feature_vector_attention_mask(
encoder_last_hidden_state.shape[1], inputs_dict["attention_mask"]
)
last_hidden_state_2 = decoder(
input_ids=inputs_dict["decoder_input_ids"],
attention_mask=inputs_dict["decoder_attention_mask"],
encoder_hidden_states=encoder_last_hidden_state,
encoder_attention_mask=encoder_attention_mask,
)[0]
self.parent.assertTrue((last_hidden_state_2 - last_hidden_state).abs().max().item() < 1e-3)
@require_torch
class Speech2TextModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin, unittest.TestCase):
all_model_classes = (Speech2TextModel, Speech2TextForConditionalGeneration) if is_torch_available() else ()
pipeline_model_mapping = (
{"automatic-speech-recognition": Speech2TextForConditionalGeneration, "feature-extraction": Speech2TextModel}
if is_torch_available()
else {}
)
is_encoder_decoder = True
fx_compatible = True
test_pruning = False
test_missing_keys = False
def setUp(self):
self.model_tester = Speech2TextModelTester(self)
self.config_tester = ConfigTester(self, config_class=Speech2TextConfig)
self.maxDiff = 3000
def test_config(self):
self.config_tester.run_common_tests()
def test_save_load_strict(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs()
for model_class in self.all_model_classes:
model = model_class(config)
with tempfile.TemporaryDirectory() as tmpdirname:
model.save_pretrained(tmpdirname)
model2, info = model_class.from_pretrained(tmpdirname, output_loading_info=True)
self.assertEqual(info["missing_keys"], [])
def test_model_forward(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_model_forward(*config_and_inputs)
def test_decoder_model_past_with_large_inputs(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_decoder_model_past_large_inputs(*config_and_inputs)
def test_encoder_decoder_model_standalone(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs_for_common()
self.model_tester.check_encoder_decoder_model_standalone(*config_and_inputs)
@unittest.skip(reason="Not implemented currently")
def test_inputs_embeds(self):
pass
@unittest.skip(reason="Training is not supported yet")
def test_training(self):
pass
@unittest.skip
def test_training_gradient_checkpointing(self):
pass
@unittest.skip(
reason="This architecture seem to not compute gradients properly when using GC, check: https://github.com/huggingface/transformers/pull/27124"
)
def test_training_gradient_checkpointing_use_reentrant(self):
pass
@unittest.skip(
reason="This architecture seem to not compute gradients properly when using GC, check: https://github.com/huggingface/transformers/pull/27124"
)
def test_training_gradient_checkpointing_use_reentrant_false(self):
pass
@require_torch_fp16
def test_generate_fp16(self):
config, input_dict = self.model_tester.prepare_config_and_inputs()
input_features = input_dict["input_features"]
attention_mask = input_dict["attention_mask"]
model = Speech2TextForConditionalGeneration(config).eval().to(torch_device)
input_features = input_features.half()
model.half()
model.generate(input_features, attention_mask=attention_mask)
model.generate(input_features, num_beams=4, do_sample=True, early_stopping=False, num_return_sequences=3)
def test_forward_signature(self):
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
model = model_class(config)
signature = inspect.signature(model.forward)
# signature.parameters is an OrderedDict => so arg_names order is deterministic
arg_names = [*signature.parameters.keys()]
expected_arg_names = [
"input_features",
"attention_mask",
"decoder_input_ids",
"decoder_attention_mask",
]
expected_arg_names.extend(
["head_mask", "decoder_head_mask", "cross_attn_head_mask", "encoder_outputs"]
if "head_mask" and "decoder_head_mask" and "cross_attn_head_mask" in arg_names
else ["encoder_outputs"]
)
self.assertListEqual(arg_names[: len(expected_arg_names)], expected_arg_names)
def test_hidden_states_output(self):
def check_hidden_states_output(inputs_dict, config, model_class):
model = model_class(config)
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(**self._prepare_for_class(inputs_dict, model_class))
hidden_states = outputs.encoder_hidden_states if config.is_encoder_decoder else outputs.hidden_states
expected_num_layers = getattr(
self.model_tester, "expected_num_hidden_layers", self.model_tester.num_hidden_layers + 1
)
self.assertEqual(len(hidden_states), expected_num_layers)
if hasattr(self.model_tester, "encoder_seq_length"):
seq_length = self.model_tester.encoder_seq_length
else:
seq_length = self.model_tester.seq_length
subsampled_seq_length = model._get_feat_extract_output_lengths(seq_length)
self.assertListEqual(
list(hidden_states[0].shape[-2:]),
[subsampled_seq_length, self.model_tester.hidden_size],
)
if config.is_encoder_decoder:
hidden_states = outputs.decoder_hidden_states
self.assertIsInstance(hidden_states, (list, tuple))
self.assertEqual(len(hidden_states), expected_num_layers)
seq_len = getattr(self.model_tester, "seq_length", None)
decoder_seq_length = getattr(self.model_tester, "decoder_seq_length", seq_len)
self.assertListEqual(
list(hidden_states[0].shape[-2:]),
[decoder_seq_length, self.model_tester.hidden_size],
)
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
inputs_dict["output_hidden_states"] = True
check_hidden_states_output(inputs_dict, config, model_class)
# check that output_hidden_states also work using config
del inputs_dict["output_hidden_states"]
config.output_hidden_states = True
check_hidden_states_output(inputs_dict, config, model_class)
def test_attention_outputs(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
config.return_dict = True
seq_len = getattr(self.model_tester, "seq_length", None)
decoder_seq_length = getattr(self.model_tester, "decoder_seq_length", seq_len)
encoder_seq_length = getattr(self.model_tester, "encoder_seq_length", seq_len)
decoder_key_length = getattr(self.model_tester, "decoder_key_length", decoder_seq_length)
encoder_key_length = getattr(self.model_tester, "key_length", encoder_seq_length)
for model_class in self.all_model_classes:
inputs_dict["output_attentions"] = True
inputs_dict["output_hidden_states"] = False
config.return_dict = True
model = model_class._from_config(config, attn_implementation="eager")
config = model.config
model.to(torch_device)
model.eval()
subsampled_encoder_seq_length = model._get_feat_extract_output_lengths(encoder_seq_length)
subsampled_encoder_key_length = model._get_feat_extract_output_lengths(encoder_key_length)
with torch.no_grad():
outputs = model(**self._prepare_for_class(inputs_dict, model_class))
attentions = outputs.encoder_attentions if config.is_encoder_decoder else outputs.attentions
self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)
# check that output_attentions also work using config
del inputs_dict["output_attentions"]
config.output_attentions = True
model = model_class(config)
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(**self._prepare_for_class(inputs_dict, model_class))
attentions = outputs.encoder_attentions if config.is_encoder_decoder else outputs.attentions
self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)
self.assertListEqual(
list(attentions[0].shape[-3:]),
[self.model_tester.num_attention_heads, subsampled_encoder_seq_length, subsampled_encoder_key_length],
)
out_len = len(outputs)
correct_outlen = 5
# loss is at first position
if "labels" in inputs_dict:
correct_outlen += 1 # loss is added to beginning
if "past_key_values" in outputs:
correct_outlen += 1 # past_key_values have been returned
self.assertEqual(out_len, correct_outlen)
# decoder attentions
decoder_attentions = outputs.decoder_attentions
self.assertIsInstance(decoder_attentions, (list, tuple))
self.assertEqual(len(decoder_attentions), self.model_tester.num_hidden_layers)
self.assertListEqual(
list(decoder_attentions[0].shape[-3:]),
[self.model_tester.num_attention_heads, decoder_seq_length, decoder_key_length],
)
# cross attentions
cross_attentions = outputs.cross_attentions
self.assertIsInstance(cross_attentions, (list, tuple))
self.assertEqual(len(cross_attentions), self.model_tester.num_hidden_layers)
self.assertListEqual(
list(cross_attentions[0].shape[-3:]),
[
self.model_tester.num_attention_heads,
decoder_seq_length,
subsampled_encoder_key_length,
],
)
# Check attention is always last and order is fine
inputs_dict["output_attentions"] = True
inputs_dict["output_hidden_states"] = True
model = model_class(config)
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(**self._prepare_for_class(inputs_dict, model_class))
added_hidden_states = 2
self.assertEqual(out_len + added_hidden_states, len(outputs))
self_attentions = outputs.encoder_attentions if config.is_encoder_decoder else outputs.attentions
self.assertEqual(len(self_attentions), self.model_tester.num_hidden_layers)
self.assertListEqual(
list(self_attentions[0].shape[-3:]),
[self.model_tester.num_attention_heads, subsampled_encoder_seq_length, subsampled_encoder_key_length],
)
def test_resize_tokens_embeddings(self):
(
original_config,
inputs_dict,
) = self.model_tester.prepare_config_and_inputs_for_common()
if not self.test_resize_embeddings:
self.skipTest(reason="test_resize_embeddings is set to False")
for model_class in self.all_model_classes:
config = copy.deepcopy(original_config)
model = model_class(config)
model.to(torch_device)
if self.model_tester.is_training is False:
model.eval()
model_vocab_size = config.vocab_size
# Retrieve the embeddings and clone theme
model_embed = model.resize_token_embeddings(model_vocab_size)
cloned_embeddings = model_embed.weight.clone()
# Check that resizing the token embeddings with a larger vocab size increases the model's vocab size
model_embed = model.resize_token_embeddings(model_vocab_size + 10)
self.assertEqual(model.config.vocab_size, model_vocab_size + 10)
# Check that it actually resizes the embeddings matrix
self.assertEqual(model_embed.weight.shape[0], cloned_embeddings.shape[0] + 10)
# Check that the model can still do a forward pass successfully (every parameter should be resized)
model(**self._prepare_for_class(inputs_dict, model_class))
# Check that resizing the token embeddings with a smaller vocab size decreases the model's vocab size
model_embed = model.resize_token_embeddings(model_vocab_size - 15)
self.assertEqual(model.config.vocab_size, model_vocab_size - 15)
# Check that it actually resizes the embeddings matrix
self.assertEqual(model_embed.weight.shape[0], cloned_embeddings.shape[0] - 15)
# make sure that decoder_input_ids are resized
if "decoder_input_ids" in inputs_dict:
inputs_dict["decoder_input_ids"].clamp_(max=model_vocab_size - 15 - 1)
model(**self._prepare_for_class(inputs_dict, model_class))
# Check that adding and removing tokens has not modified the first part of the embedding matrix.
models_equal = True
for p1, p2 in zip(cloned_embeddings, model_embed.weight):
if p1.data.ne(p2.data).sum() > 0:
models_equal = False
self.assertTrue(models_equal)
def test_resize_embeddings_untied(self):
(
original_config,
inputs_dict,
) = self.model_tester.prepare_config_and_inputs_for_common()
if not self.test_resize_embeddings:
self.skipTest(reason="test_resize_embeddings is set to False")
original_config.tie_word_embeddings = False
# if model cannot untied embeddings -> leave test
if original_config.tie_word_embeddings:
self.skipTest(reason="Model cannot untie embeddings")
for model_class in self.all_model_classes:
config = copy.deepcopy(original_config)
model = model_class(config).to(torch_device)
model.eval()
# if no output embeddings -> leave test
if model.get_output_embeddings() is None:
continue
# Check that resizing the token embeddings with a larger vocab size increases the model's vocab size
model_vocab_size = config.vocab_size
model.resize_token_embeddings(model_vocab_size + 10)
self.assertEqual(model.config.vocab_size, model_vocab_size + 10)
output_embeds = model.get_output_embeddings()
self.assertEqual(output_embeds.weight.shape[0], model_vocab_size + 10)
# Check bias if present
if output_embeds.bias is not None:
self.assertEqual(output_embeds.bias.shape[0], model_vocab_size + 10)
# Check that the model can still do a forward pass successfully (every parameter should be resized)
model(**self._prepare_for_class(inputs_dict, model_class))
# Check that resizing the token embeddings with a smaller vocab size decreases the model's vocab size
model.resize_token_embeddings(model_vocab_size - 15)
self.assertEqual(model.config.vocab_size, model_vocab_size - 15)
# Check that it actually resizes the embeddings matrix
output_embeds = model.get_output_embeddings()
self.assertEqual(output_embeds.weight.shape[0], model_vocab_size - 15)
# Check bias if present
if output_embeds.bias is not None:
self.assertEqual(output_embeds.bias.shape[0], model_vocab_size - 15)
# Check that the model can still do a forward pass successfully (every parameter should be resized)
if "decoder_input_ids" in inputs_dict:
inputs_dict["decoder_input_ids"].clamp_(max=model_vocab_size - 15 - 1)
# Check that the model can still do a forward pass successfully (every parameter should be resized)
model(**self._prepare_for_class(inputs_dict, model_class))
@unittest.skip
def test_generate_without_input_ids(self):
pass
def _create_and_check_torchscript(self, config, inputs_dict):
if not self.test_torchscript:
self.skipTest(reason="test_torchscript is set to False")
configs_no_init = _config_zero_init(config) # To be sure we have no Nan
configs_no_init.torchscript = True
for model_class in self.all_model_classes:
model = model_class(config=configs_no_init)
model.to(torch_device)
model.eval()
inputs = self._prepare_for_class(inputs_dict, model_class)
try:
model.config.use_cache = False # FSTM still requires this hack -> FSTM should probably be refactored similar to BART afterward
input_features = inputs["input_features"]
attention_mask = inputs["attention_mask"]
decoder_input_ids = inputs["decoder_input_ids"]
decoder_attention_mask = inputs["decoder_attention_mask"]
traced_model = torch.jit.trace(
model, (input_features, attention_mask, decoder_input_ids, decoder_attention_mask)
)
except RuntimeError:
self.fail("Couldn't trace module.")
with tempfile.TemporaryDirectory() as tmp_dir_name:
pt_file_name = os.path.join(tmp_dir_name, "traced_model.pt")
try:
torch.jit.save(traced_model, pt_file_name)
except Exception:
self.fail("Couldn't save module.")
try:
loaded_model = torch.jit.load(pt_file_name)
except Exception:
self.fail("Couldn't load module.")
model.to(torch_device)
model.eval()
loaded_model.to(torch_device)
loaded_model.eval()
model_state_dict = model.state_dict()
loaded_model_state_dict = loaded_model.state_dict()
non_persistent_buffers = {}
for key in loaded_model_state_dict:
if key not in model_state_dict:
non_persistent_buffers[key] = loaded_model_state_dict[key]
loaded_model_state_dict = {
key: value for key, value in loaded_model_state_dict.items() if key not in non_persistent_buffers
}
self.assertEqual(set(model_state_dict.keys()), set(loaded_model_state_dict.keys()))
model_buffers = list(model.buffers())
for non_persistent_buffer in non_persistent_buffers.values():
found_buffer = False
for i, model_buffer in enumerate(model_buffers):
if torch.equal(non_persistent_buffer, model_buffer):
found_buffer = True
break
self.assertTrue(found_buffer)
model_buffers.pop(i)
models_equal = True
for layer_name, p1 in model_state_dict.items():
p2 = loaded_model_state_dict[layer_name]
if p1.data.ne(p2.data).sum() > 0:
models_equal = False
self.assertTrue(models_equal)
@unittest.skip(reason="Test failing, @RocketNight is looking into it")
def test_tf_from_pt_safetensors(self):
pass
@require_torch
@require_torchaudio
@require_sentencepiece
@require_tokenizers
class Speech2TextModelIntegrationTests(unittest.TestCase):
@classmethod
def setUpClass(cls):
model_name = "facebook/s2t-small-librispeech-asr"
cls.model = Speech2TextForConditionalGeneration.from_pretrained(model_name, device_map="auto")
cls.processor = Speech2TextProcessor.from_pretrained(model_name)
# loads 4 samples
ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
speech_samples = ds.sort("id").select(range(4))[:4]["audio"]
cls.dataset = [x["array"] for x in speech_samples]
def test_generation_librispeech(self):
input_speech = [self.dataset[0]]
input_features = self.processor(input_speech, return_tensors="pt").input_features.to(torch_device)
generated_ids = self.model.generate(input_features)
generated_transcript = self.processor.batch_decode(generated_ids, skip_special_tokens=True)
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):
input_speech = self.dataset
inputs = self.processor(input_speech, return_tensors="pt", padding=True)
input_features = inputs.input_features.to(torch_device)
attention_mask = inputs.attention_mask.to(torch_device)
generated_ids = self.model.generate(input_features, attention_mask=attention_mask)
generated_transcripts = self.processor.batch_decode(generated_ids, skip_special_tokens=True)
EXPECTED_TRANSCRIPTIONS = [
"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)

View File

@@ -0,0 +1,147 @@
# 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.
import shutil
import tempfile
import unittest
from pathlib import Path
from shutil import copyfile
from transformers import Speech2TextFeatureExtractor, Speech2TextProcessor, Speech2TextTokenizer
from transformers.models.speech_to_text.tokenization_speech_to_text import VOCAB_FILES_NAMES, save_json
from transformers.testing_utils import get_tests_dir, require_sentencepiece, require_torch, require_torchaudio
from transformers.utils import FEATURE_EXTRACTOR_NAME
from .test_feature_extraction_speech_to_text import floats_list
SAMPLE_SP = get_tests_dir("fixtures/test_sentencepiece.model")
@require_torch
@require_torchaudio
@require_sentencepiece
class Speech2TextProcessorTest(unittest.TestCase):
@classmethod
def setUpClass(cls):
cls.tmpdirname = tempfile.mkdtemp()
vocab = ["<s>", "<pad>", "</s>", "<unk>", "▁This", "▁is", "▁a", "▁t", "est"]
vocab_tokens = dict(zip(vocab, range(len(vocab))))
save_dir = Path(cls.tmpdirname)
save_json(vocab_tokens, save_dir / VOCAB_FILES_NAMES["vocab_file"])
if not (save_dir / VOCAB_FILES_NAMES["spm_file"]).exists():
copyfile(SAMPLE_SP, save_dir / VOCAB_FILES_NAMES["spm_file"])
tokenizer = Speech2TextTokenizer.from_pretrained(cls.tmpdirname)
tokenizer.save_pretrained(cls.tmpdirname)
feature_extractor_map = {
"feature_size": 24,
"num_mel_bins": 24,
"padding_value": 0.0,
"sampling_rate": 16000,
"return_attention_mask": False,
"do_normalize": True,
}
save_json(feature_extractor_map, save_dir / FEATURE_EXTRACTOR_NAME)
def get_tokenizer(self, **kwargs):
return Speech2TextTokenizer.from_pretrained(self.tmpdirname, **kwargs)
def get_feature_extractor(self, **kwargs):
return Speech2TextFeatureExtractor.from_pretrained(self.tmpdirname, **kwargs)
@classmethod
def tearDownClass(cls):
shutil.rmtree(cls.tmpdirname, ignore_errors=True)
def test_save_load_pretrained_default(self):
tokenizer = self.get_tokenizer()
feature_extractor = self.get_feature_extractor()
processor = Speech2TextProcessor(tokenizer=tokenizer, feature_extractor=feature_extractor)
processor.save_pretrained(self.tmpdirname)
processor = Speech2TextProcessor.from_pretrained(self.tmpdirname)
self.assertEqual(processor.tokenizer.get_vocab(), tokenizer.get_vocab())
self.assertIsInstance(processor.tokenizer, Speech2TextTokenizer)
self.assertEqual(processor.feature_extractor.to_json_string(), feature_extractor.to_json_string())
self.assertIsInstance(processor.feature_extractor, Speech2TextFeatureExtractor)
def test_save_load_pretrained_additional_features(self):
with tempfile.TemporaryDirectory() as tmpdir:
processor = Speech2TextProcessor(
tokenizer=self.get_tokenizer(), feature_extractor=self.get_feature_extractor()
)
processor.save_pretrained(tmpdir)
tokenizer_add_kwargs = Speech2TextTokenizer.from_pretrained(tmpdir, bos_token="(BOS)", eos_token="(EOS)")
feature_extractor_add_kwargs = Speech2TextFeatureExtractor.from_pretrained(
tmpdir, do_normalize=False, padding_value=1.0
)
processor = Speech2TextProcessor.from_pretrained(
tmpdir, 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, Speech2TextTokenizer)
self.assertEqual(processor.feature_extractor.to_json_string(), feature_extractor_add_kwargs.to_json_string())
self.assertIsInstance(processor.feature_extractor, Speech2TextFeatureExtractor)
def test_feature_extractor(self):
feature_extractor = self.get_feature_extractor()
tokenizer = self.get_tokenizer()
processor = Speech2TextProcessor(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:
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 = Speech2TextProcessor(tokenizer=tokenizer, feature_extractor=feature_extractor)
input_str = "This is a test string"
encoded_processor = processor(text=input_str)
encoded_tok = tokenizer(input_str)
for key in encoded_tok:
self.assertListEqual(encoded_tok[key], encoded_processor[key])
def test_tokenizer_decode(self):
feature_extractor = self.get_feature_extractor()
tokenizer = self.get_tokenizer()
processor = Speech2TextProcessor(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)

View File

@@ -0,0 +1,153 @@
# 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.
import unittest
from pathlib import Path
from shutil import copyfile
from transformers import SPIECE_UNDERLINE, is_sentencepiece_available
from transformers.models.speech_to_text import Speech2TextTokenizer
from transformers.models.speech_to_text.tokenization_speech_to_text import VOCAB_FILES_NAMES, save_json
from transformers.testing_utils import get_tests_dir, require_sentencepiece, require_tokenizers, slow
from ...test_tokenization_common import TokenizerTesterMixin
SAMPLE_VOCAB = get_tests_dir("fixtures/test_sentencepiece.model")
if is_sentencepiece_available():
import sentencepiece as sp
FR_CODE = 5
ES_CODE = 10
@require_sentencepiece
@require_tokenizers
class SpeechToTextTokenizerTest(TokenizerTesterMixin, unittest.TestCase):
from_pretrained_id = "facebook/s2t-small-librispeech-asr"
tokenizer_class = Speech2TextTokenizer
test_rust_tokenizer = False
test_sentencepiece = True
@classmethod
def setUpClass(cls):
super().setUpClass()
spm_model = sp.SentencePieceProcessor()
spm_model.Load(SAMPLE_VOCAB)
vocab = ["<s>", "<pad>", "</s>", "<unk>"]
vocab += [spm_model.IdToPiece(id_) for id_ in range(len(spm_model))]
vocab_tokens = dict(zip(vocab, range(len(vocab))))
save_dir = Path(cls.tmpdirname)
save_json(vocab_tokens, save_dir / VOCAB_FILES_NAMES["vocab_file"])
if not (save_dir / VOCAB_FILES_NAMES["spm_file"]).exists():
copyfile(SAMPLE_VOCAB, save_dir / VOCAB_FILES_NAMES["spm_file"])
tokenizer = Speech2TextTokenizer.from_pretrained(cls.tmpdirname)
tokenizer.save_pretrained(cls.tmpdirname)
def test_convert_token_and_id(self):
"""Test ``_convert_token_to_id`` and ``_convert_id_to_token``."""
token = "<pad>"
token_id = 1
self.assertEqual(self.get_tokenizer()._convert_token_to_id(token), token_id)
self.assertEqual(self.get_tokenizer()._convert_id_to_token(token_id), token)
def test_get_vocab(self):
vocab_keys = list(self.get_tokenizer().get_vocab().keys())
self.assertEqual(vocab_keys[0], "<s>")
self.assertEqual(vocab_keys[1], "<pad>")
self.assertEqual(vocab_keys[-1], "j")
self.assertEqual(len(vocab_keys), 1_001)
def test_vocab_size(self):
self.assertEqual(self.get_tokenizer().vocab_size, 1_001)
def test_full_tokenizer(self):
tokenizer = Speech2TextTokenizer.from_pretrained(self.tmpdirname)
tokens = tokenizer.tokenize("This is a test")
self.assertListEqual(tokens, ["▁This", "▁is", "▁a", "▁t", "est"])
self.assertListEqual(
tokenizer.convert_tokens_to_ids(tokens),
[289, 50, 14, 174, 386],
)
tokens = tokenizer.tokenize("I was born in 92000, and this is falsé.")
self.assertListEqual(tokens,[SPIECE_UNDERLINE + "I", SPIECE_UNDERLINE + "was", SPIECE_UNDERLINE + "b", "or", "n", SPIECE_UNDERLINE + "in", SPIECE_UNDERLINE + "", "9", "2", "0", "0", "0", ",", SPIECE_UNDERLINE + "and", SPIECE_UNDERLINE + "this", SPIECE_UNDERLINE + "is", SPIECE_UNDERLINE + "f", "al", "s", "é", "."]) # fmt: skip
ids = tokenizer.convert_tokens_to_ids(tokens)
self.assertListEqual(ids, [12, 25, 88, 59, 28, 23, 11, 4, 606, 351, 351, 351, 7, 16, 70, 50, 76, 84, 10, 4, 8])
back_tokens = tokenizer.convert_ids_to_tokens(ids)
self.assertListEqual(back_tokens,[SPIECE_UNDERLINE + "I", SPIECE_UNDERLINE + "was", SPIECE_UNDERLINE + "b", "or", "n", SPIECE_UNDERLINE + "in", SPIECE_UNDERLINE + "", "<unk>", "2", "0", "0", "0", ",", SPIECE_UNDERLINE + "and", SPIECE_UNDERLINE + "this", SPIECE_UNDERLINE + "is", SPIECE_UNDERLINE + "f", "al", "s", "<unk>", "."]) # fmt: skip
@slow
def test_tokenizer_integration(self):
expected_encoding = {'input_ids': [[3791, 797, 31, 11, 64, 797, 31, 2429, 433, 12, 1176, 12, 20, 786, 915, 142, 2413, 240, 37, 3238, 797, 31, 11, 35, 93, 915, 142, 2413, 240, 37, 5540, 567, 1276, 93, 37, 610, 40, 62, 455, 657, 1042, 123, 780, 177, 37, 309, 241, 1298, 514, 20, 292, 2737, 114, 2469, 241, 85, 64, 302, 548, 528, 423, 4, 509, 406, 423, 37, 601, 4, 777, 302, 548, 528, 423, 284, 4, 3388, 511, 459, 4, 3555, 40, 321, 302, 705, 4, 3388, 511, 583, 326, 5, 5, 5, 62, 3310, 560, 177, 2680, 217, 1508, 32, 31, 853, 418, 64, 583, 511, 1605, 62, 35, 93, 560, 177, 2680, 217, 1508, 1521, 64, 583, 511, 519, 62, 20, 1515, 764, 20, 149, 261, 5625, 7972, 20, 5540, 567, 1276, 93, 3925, 1675, 11, 15, 802, 7972, 576, 217, 1508, 11, 35, 93, 1253, 2441, 15, 289, 652, 31, 416, 321, 3842, 115, 40, 911, 8, 476, 619, 4, 380, 142, 423, 335, 240, 35, 93, 264, 8, 11, 335, 569, 420, 163, 5, 2], [260, 548, 528, 423, 20, 451, 20, 2681, 1153, 3434, 20, 5540, 37, 567, 126, 1253, 2441, 3376, 449, 210, 431, 1563, 177, 767, 5540, 11, 1203, 472, 11, 2953, 685, 285, 364, 706, 1153, 20, 6799, 20, 2869, 20, 4464, 126, 40, 2429, 20, 1040, 866, 2664, 418, 20, 318, 20, 1726, 186, 20, 265, 522, 35, 93, 2191, 4634, 20, 1040, 12, 6799, 15, 228, 2356, 142, 31, 11, 5, 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [2575, 2666, 684, 1582, 1176, 12, 627, 149, 619, 20, 4902, 563, 11, 20, 149, 261, 3420, 2356, 174, 142, 4714, 131, 5, 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]], 'attention_mask': [[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]]} # fmt: skip
self.tokenizer_integration_test_util(
expected_encoding=expected_encoding,
model_name="facebook/s2t-small-mustc-en-de-st",
revision="a14f04cf0776c02f62a8cb800cf7909e15ea23ad",
)
@require_sentencepiece
class SpeechToTextTokenizerMultilinguialTest(unittest.TestCase):
checkpoint_name = "valhalla/s2t_mustc_multilinguial_medium"
french_text = "C'est trop cool"
spanish_text = "Esto es genial"
@classmethod
def setUpClass(cls):
cls.tokenizer: Speech2TextTokenizer = Speech2TextTokenizer.from_pretrained(cls.checkpoint_name)
return cls
def check_language_codes(self):
self.assertEqual(self.tokenizer.lang_code_to_id["pt"], 4)
self.assertEqual(self.tokenizer.lang_code_to_id["ru"], 6)
self.assertEqual(self.tokenizer.lang_code_to_id["it"], 9)
self.assertEqual(self.tokenizer.lang_code_to_id["de"], 11)
def test_vocab_size(self):
self.assertEqual(self.tokenizer.vocab_size, 10_000)
def test_tokenizer_decode_ignores_language_codes(self):
self.assertIn(ES_CODE, self.tokenizer.all_special_ids)
generated_ids = [ES_CODE, 4, 1601, 47, 7647, 2]
result = self.tokenizer.decode(generated_ids, skip_special_tokens=True)
expected_spanish = self.tokenizer.decode(generated_ids[1:], skip_special_tokens=True)
self.assertEqual(result, expected_spanish)
self.assertNotIn(self.tokenizer.eos_token, result)
def test_tokenizer_adds_special_tokens(self):
self.tokenizer.tgt_lang = "fr"
encoded = self.tokenizer(self.french_text).input_ids
self.assertEqual(encoded[0], FR_CODE)
self.assertEqual(encoded[-1], self.tokenizer.eos_token_id)
def test_tgt_lang_setter(self):
self.tokenizer.tgt_lang = "fr"
self.assertListEqual(self.tokenizer.prefix_tokens, [FR_CODE])
self.tokenizer.tgt_lang = "es"
self.assertListEqual(self.tokenizer.prefix_tokens, [ES_CODE])