# 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 Hubert model.""" import math import unittest import pytest from transformers import SEWDConfig, is_torch_available from transformers.testing_utils import require_torch, require_torchcodec, slow, torch_device from ...test_configuration_common import ConfigTester from ...test_modeling_common import ( ModelTesterMixin, floats_tensor, ids_tensor, random_attention_mask, ) from ...test_pipeline_mixin import PipelineTesterMixin if is_torch_available(): import torch from transformers import ( SEWDForCTC, SEWDForSequenceClassification, SEWDModel, Wav2Vec2FeatureExtractor, Wav2Vec2Processor, ) from transformers.models.hubert.modeling_hubert import _compute_mask_indices class SEWDModelTester: def __init__( self, parent, batch_size=13, seq_length=1024, # speech is longer is_training=False, hidden_size=32, feat_extract_norm="group", feat_extract_dropout=0.0, feat_extract_activation="gelu", conv_dim=(64, 32, 32), conv_stride=(5, 2, 1), conv_kernel=(10, 3, 1), conv_bias=False, num_conv_pos_embeddings=31, num_conv_pos_embedding_groups=2, squeeze_factor=2, max_position_embeddings=512, position_buckets=256, share_att_key=True, relative_attention=True, position_biased_input=False, pos_att_type=("p2c", "c2p"), norm_rel_ebd="layer_norm", num_hidden_layers=2, num_attention_heads=2, hidden_dropout=0.1, 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.squeeze_factor = squeeze_factor self.max_position_embeddings = max_position_embeddings self.position_buckets = position_buckets self.share_att_key = share_att_key self.relative_attention = relative_attention self.position_biased_input = position_biased_input self.pos_att_type = pos_att_type self.norm_rel_ebd = norm_rel_ebd self.num_hidden_layers = num_hidden_layers self.num_attention_heads = num_attention_heads self.hidden_dropout = hidden_dropout 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 // self.squeeze_factor def prepare_config_and_inputs(self): input_values = floats_tensor([self.batch_size, self.seq_length], scale=1.0) attention_mask = random_attention_mask([self.batch_size, self.seq_length]) config = self.get_config() return config, input_values, attention_mask def get_config(self): return SEWDConfig( 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, squeeze_factor=self.squeeze_factor, max_position_embeddings=self.max_position_embeddings, position_buckets=self.position_buckets, share_att_key=self.share_att_key, relative_attention=self.relative_attention, position_biased_input=self.position_biased_input, pos_att_type=self.pos_att_type, norm_rel_ebd=self.norm_rel_ebd, num_hidden_layers=self.num_hidden_layers, num_attention_heads=self.num_attention_heads, hidden_dropout=self.hidden_dropout, 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, ) def create_and_check_model(self, config, input_values, attention_mask): model = SEWDModel(config=config) model.to(torch_device) model.eval() result = model(input_values, attention_mask=attention_mask) self.parent.assertEqual( result.last_hidden_state.shape, (self.batch_size, self.output_seq_length, self.hidden_size) ) def check_ctc_loss(self, config, input_values, *args): model = SEWDForCTC(config=config) model.to(torch_device) # make sure that dropout is disabled model.eval() input_values = input_values[:3] attention_mask = torch.ones(input_values.shape, device=torch_device, dtype=torch.long) input_lengths = [input_values.shape[-1] // i for i in [4, 2, 1]] max_length_labels = model._get_feat_extract_output_lengths(torch.tensor(input_lengths)) labels = ids_tensor((input_values.shape[0], min(max_length_labels) - 1), model.config.vocab_size) # pad input for i in range(len(input_lengths)): input_values[i, input_lengths[i] :] = 0.0 attention_mask[i, input_lengths[i] :] = 0 model.config.ctc_loss_reduction = "sum" sum_loss = model(input_values, attention_mask=attention_mask, labels=labels).loss.item() model.config.ctc_loss_reduction = "mean" mean_loss = model(input_values, attention_mask=attention_mask, labels=labels).loss.item() self.parent.assertTrue(isinstance(sum_loss, float)) self.parent.assertTrue(isinstance(mean_loss, float)) def check_ctc_training(self, config, input_values, *args): config.ctc_zero_infinity = True model = SEWDForCTC(config=config) model.to(torch_device) model.train() # freeze feature encoder model.freeze_feature_encoder() input_values = input_values[:3] input_lengths = [input_values.shape[-1] // i for i in [4, 2, 1]] max_length_labels = model._get_feat_extract_output_lengths(torch.tensor(input_lengths)) labels = ids_tensor((input_values.shape[0], max(max_length_labels) - 2), model.config.vocab_size) # pad input for i in range(len(input_lengths)): input_values[i, input_lengths[i] :] = 0.0 if max_length_labels[i] < labels.shape[-1]: # it's important that we make sure that target lengths are at least # one shorter than logit lengths to prevent -inf labels[i, max_length_labels[i] - 1 :] = -100 loss = model(input_values, labels=labels).loss self.parent.assertFalse(torch.isinf(loss).item()) loss.backward() def check_seq_classifier_loss(self, config, input_values, *args): model = SEWDForSequenceClassification(config=config) model.to(torch_device) # make sure that dropout is disabled model.eval() input_values = input_values[:3] attention_mask = torch.ones(input_values.shape, device=torch_device, dtype=torch.long) input_lengths = [input_values.shape[-1] // i for i in [4, 2, 1]] labels = ids_tensor((input_values.shape[0], 1), len(model.config.id2label)) # pad input for i in range(len(input_lengths)): input_values[i, input_lengths[i] :] = 0.0 attention_mask[i, input_lengths[i] :] = 0 masked_loss = model(input_values, attention_mask=attention_mask, labels=labels).loss.item() unmasked_loss = model(input_values, labels=labels).loss.item() self.parent.assertTrue(isinstance(masked_loss, float)) self.parent.assertTrue(isinstance(unmasked_loss, float)) self.parent.assertTrue(masked_loss != unmasked_loss) def check_seq_classifier_training(self, config, input_values, *args): config.ctc_zero_infinity = True model = SEWDForSequenceClassification(config=config) model.to(torch_device) model.train() # freeze everything but the classification head model.freeze_base_model() input_values = input_values[:3] input_lengths = [input_values.shape[-1] // i for i in [4, 2, 1]] labels = ids_tensor((input_values.shape[0], 1), len(model.config.id2label)) # pad input for i in range(len(input_lengths)): input_values[i, input_lengths[i] :] = 0.0 loss = model(input_values, labels=labels).loss self.parent.assertFalse(torch.isinf(loss).item()) loss.backward() def check_labels_out_of_vocab(self, config, input_values, *args): model = SEWDForCTC(config) model.to(torch_device) model.train() input_values = input_values[:3] input_lengths = [input_values.shape[-1] // i for i in [4, 2, 1]] max_length_labels = model._get_feat_extract_output_lengths(torch.tensor(input_lengths)) labels = ids_tensor((input_values.shape[0], max(max_length_labels) - 2), model.config.vocab_size + 100) with pytest.raises(ValueError): model(input_values, labels=labels) def prepare_config_and_inputs_for_common(self): config, input_values, attention_mask = self.prepare_config_and_inputs() inputs_dict = {"input_values": input_values, "attention_mask": attention_mask} return config, inputs_dict @require_torch class SEWDModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase): all_model_classes = (SEWDForCTC, SEWDModel, SEWDForSequenceClassification) if is_torch_available() else () pipeline_model_mapping = ( { "audio-classification": SEWDForSequenceClassification, "automatic-speech-recognition": SEWDForCTC, "feature-extraction": SEWDModel, } if is_torch_available() else {} ) def setUp(self): self.model_tester = SEWDModelTester(self) self.config_tester = ConfigTester(self, config_class=SEWDConfig, 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) def test_ctc_loss_inference(self): config_and_inputs = self.model_tester.prepare_config_and_inputs() self.model_tester.check_ctc_loss(*config_and_inputs) def test_ctc_train(self): config_and_inputs = self.model_tester.prepare_config_and_inputs() self.model_tester.check_ctc_training(*config_and_inputs) def test_labels_out_of_vocab(self): config_and_inputs = self.model_tester.prepare_config_and_inputs() self.model_tester.check_labels_out_of_vocab(*config_and_inputs) @unittest.skip(reason="Model has no inputs_embeds") def test_inputs_embeds(self): pass @unittest.skip(reason="Model has input_values instead of input_ids") def test_forward_signature(self): pass @unittest.skip(reason="Model has no tokens embeddings") def test_resize_tokens_embeddings(self): pass @unittest.skip(reason="Model has no inputs_embeds") def test_model_get_set_embeddings(self): pass def test_retain_grad_hidden_states_attentions(self): config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common() config.output_hidden_states = True config.output_attentions = True # no need to test all models as different heads yield the same functionality model_class = self.all_model_classes[0] model = model_class(config) model.to(torch_device) # set layer drop to 0 model.config.layerdrop = 0.0 input_values = inputs_dict["input_values"] input_lengths = torch.tensor( [input_values.shape[1] for _ in range(input_values.shape[0])], dtype=torch.long, device=torch_device ) output_lengths = model._get_feat_extract_output_lengths(input_lengths) labels = ids_tensor((input_values.shape[0], output_lengths[0] - 2), self.model_tester.vocab_size) inputs_dict["attention_mask"] = torch.ones_like(inputs_dict["attention_mask"]) inputs_dict["labels"] = labels outputs = model(**inputs_dict) output = outputs[0] # Encoder-/Decoder-only models hidden_states = outputs.hidden_states[0] attentions = outputs.attentions[0] hidden_states.retain_grad() attentions.retain_grad() output.flatten()[0].backward(retain_graph=True) self.assertIsNotNone(hidden_states.grad) self.assertIsNotNone(attentions.grad) # overwrite from test_modeling_common def _mock_init_weights(self, module): if hasattr(module, "weight") and module.weight is not None: module.weight.fill_(3) if hasattr(module, "weight_g") and module.weight_g is not None: module.weight_g.data.fill_(3) if hasattr(module, "weight_v") and module.weight_v is not None: module.weight_v.data.fill_(3) if hasattr(module, "bias") and module.bias is not None: module.bias.fill_(3) if hasattr(module, "masked_spec_embed") and module.masked_spec_embed is not None: module.masked_spec_embed.data.fill_(3) @unittest.skip(reason="Feed forward chunking is not implemented") def test_feed_forward_chunking(self): pass @slow def test_model_from_pretrained(self): model = SEWDModel.from_pretrained("asapp/sew-d-tiny-100k") self.assertIsNotNone(model) @require_torch class SEWDUtilsTest(unittest.TestCase): def test_compute_mask_indices(self): batch_size = 4 sequence_length = 60 mask_prob = 0.5 mask_length = 1 mask = _compute_mask_indices((batch_size, sequence_length), mask_prob, mask_length) mask = torch.from_numpy(mask).to(torch_device) self.assertListEqual(mask.sum(axis=-1).tolist(), [mask_prob * sequence_length for _ in range(batch_size)]) def test_compute_mask_indices_overlap(self): batch_size = 4 sequence_length = 80 mask_prob = 0.5 mask_length = 4 mask = _compute_mask_indices((batch_size, sequence_length), mask_prob, mask_length) mask = torch.from_numpy(mask).to(torch_device) # because of overlap mask don't have to add up exactly to `mask_prob * sequence_length`, but have to be smaller or equal for batch_sum in mask.sum(axis=-1): self.assertTrue(int(batch_sum) <= mask_prob * sequence_length) @require_torch @require_torchcodec @slow class SEWDModelIntegrationTest(unittest.TestCase): 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").filter( lambda x: x["id"] in [f"1272-141231-000{i}" for i in range(num_samples)] )[:num_samples]["audio"] return [x["array"] for x in speech_samples] def test_inference_pretrained_batched(self): model = SEWDModel.from_pretrained("asapp/sew-d-tiny-100k").to(torch_device) processor = Wav2Vec2FeatureExtractor.from_pretrained("asapp/sew-d-tiny-100k") input_speech = self._load_datasamples(2) inputs = processor(input_speech, return_tensors="pt", padding=True) input_values = inputs.input_values.to(torch_device) with torch.no_grad(): outputs = model(input_values).last_hidden_state # expected outputs taken from the original SEW-D implementation expected_outputs_first = torch.tensor( [ [ [-0.1619, 0.6995, 0.4062, -0.1014], [-0.1364, 0.5960, 0.0952, -0.0873], [-0.1572, 0.5718, 0.4228, -0.0864], [-0.1325, 0.6823, 0.1387, -0.0871], ], [ [-0.1296, 0.4008, 0.4952, -0.1450], [-0.1152, 0.3693, 0.3037, -0.1290], [-0.1194, 0.6074, 0.3531, -0.1466], [-0.1113, 0.3135, 0.2224, -0.1338], ], ], device=torch_device, ) expected_outputs_last = torch.tensor( [ [ [-0.1577, 0.5108, 0.8553, 0.2550], [-0.1530, 0.3580, 0.6143, 0.2672], [-0.1535, 0.4954, 0.8503, 0.1387], [-0.1572, 0.3363, 0.6217, 0.1490], ], [ [-0.1338, 0.5459, 0.9607, -0.1133], [-0.1502, 0.3738, 0.7313, -0.0986], [-0.0953, 0.4708, 1.0821, -0.0944], [-0.1474, 0.3598, 0.7248, -0.0748], ], ], device=torch_device, ) expected_output_sum = 54201.0469 torch.testing.assert_close(outputs[:, :4, :4], expected_outputs_first, rtol=1e-3, atol=1e-3) torch.testing.assert_close(outputs[:, -4:, -4:], expected_outputs_last, rtol=1e-3, atol=1e-3) self.assertTrue(abs(outputs.sum() - expected_output_sum) < 1) def test_inference_ctc_batched(self): model = SEWDForCTC.from_pretrained("asapp/sew-d-tiny-100k-ft-ls100h").to(torch_device) processor = Wav2Vec2Processor.from_pretrained("asapp/sew-d-tiny-100k-ft-ls100h", do_lower_case=True) input_speech = self._load_datasamples(2) inputs = processor(input_speech, return_tensors="pt", padding=True) input_values = inputs.input_values.to(torch_device) with torch.no_grad(): logits = model(input_values).logits predicted_ids = torch.argmax(logits, dim=-1) predicted_trans = processor.batch_decode(predicted_ids) EXPECTED_TRANSCRIPTIONS = [ "a man said to the universe sir i exist", "swet covered breon's body trickling into the titlowing closs that was the only garmened he war", ] self.assertListEqual(predicted_trans, EXPECTED_TRANSCRIPTIONS)