def test_preprocessing_with_resnet_encoder(self):
self._image_modality_config = MMFTransformerModalityConfig(
type="image",
key="image",
embedding_dim=2048,
position_dim=1,
segment_id=0,
encoder=ImageEncoderFactory.Config(
type=ImageEncoderTypes.resnet152,
params=ResNet152ImageEncoder.Config(pretrained=False),
),
)
modalities_config = [
self._image_modality_config, self._text_modality_config
]
config = MMFTransformer.Config(modalities=modalities_config,
num_labels=2)
mmft = build_model(config)
sample_list = SampleList()
sample_list.image = torch.rand(2, 3, 224, 224)
sample_list.text = torch.randint(0, 512, (2, 128))
transformer_input = mmft.preprocess_sample(sample_list)
input_ids = transformer_input["input_ids"]
self.assertEqual(input_ids["image"].dim(), 3)
self.assertEqual(list(input_ids["image"].size()), [2, 1, 2048])
self.assertEqual(input_ids["text"].dim(), 2)
self.assertEqual(list(input_ids["text"].size()), [2, 128])
position_ids = transformer_input["position_ids"]
test_utils.compare_tensors(position_ids["image"],
torch.tensor([[0], [0]]))
test_utils.compare_tensors(
position_ids["text"],
torch.arange(0, 128).unsqueeze(0).expand((2, 128)))
masks = transformer_input["masks"]
test_utils.compare_tensors(masks["image"], torch.tensor([[1], [1]]))
test_utils.compare_tensors(masks["text"], torch.ones((2, 128)).long())
segment_ids = transformer_input["segment_ids"]
test_utils.compare_tensors(segment_ids["image"],
torch.tensor([[0], [0]]))
test_utils.compare_tensors(segment_ids["text"],
torch.ones((2, 128)).long())
def test_one_dim_feature_preprocessing(self):
modalities_config = [
self._image_modality_config, self._text_modality_config
]
config = MMFTransformer.Config(modalities=modalities_config,
num_labels=2)
mmft = build_model(config)
sample_list = SampleList()
sample_list.image = torch.rand(2, 256)
sample_list.text = torch.randint(0, 512, (2, 128))
transformer_input = mmft.preprocess_sample(sample_list)
input_ids = transformer_input["input_ids"]
self.assertEqual(input_ids["image"].dim(), 3)
self.assertEqual(list(input_ids["image"].size()), [2, 1, 256])
self.assertEqual(input_ids["text"].dim(), 2)
self.assertEqual(list(input_ids["text"].size()), [2, 128])
position_ids = transformer_input["position_ids"]
test_utils.compare_tensors(position_ids["image"],
torch.tensor([[0], [0]]))
test_utils.compare_tensors(
position_ids["text"],
torch.arange(0, 128).unsqueeze(0).expand((2, 128)))
masks = transformer_input["masks"]
masks = mmft._infer_masks(sample_list, input_ids)
test_utils.compare_tensors(masks["image"], torch.tensor([[1], [1]]))
test_utils.compare_tensors(masks["text"], torch.ones((2, 128)).long())
segment_ids = transformer_input["segment_ids"]
test_utils.compare_tensors(segment_ids["image"],
torch.tensor([[0], [0]]))
test_utils.compare_tensors(segment_ids["text"],
torch.ones((2, 128)).long())
mlm_labels = transformer_input["mlm_labels"]
test_utils.compare_tensors(
mlm_labels["combined_labels"],
torch.full((2, 129), dtype=torch.long, fill_value=-1),
)
def test_custom_feature_and_mask_preprocessing(self):
extra_modality = MMFTransformerModalityConfig(
type="my_random_feature",
key="my_random_feature",
embedding_dim=128,
position_dim=4,
segment_id=3,
encoder=EncoderFactory.Config(type="identity"),
)
modalities_config = [
self._image_modality_config,
self._text_modality_config,
extra_modality,
]
config = MMFTransformer.Config(modalities=modalities_config,
num_labels=2)
mmft = build_model(config)
sample_list = SampleList()
sample_list.image = torch.rand(2, 256)
sample_list.text = torch.randint(0, 512, (2, 128))
sample_list.text_mask = torch.ones(2, 128)
sample_list.text_mask[:, 70:] = 0
sample_list.my_random_feature = torch.rand(2, 4, 128)
sample_list.my_random_feature_mask = torch.ones(2, 4)
sample_list.my_random_feature_mask[:, 3:] = 0
transformer_input = mmft.preprocess_sample(sample_list)
input_ids = transformer_input["input_ids"]
self.assertEqual(input_ids["image"].dim(), 3)
self.assertEqual(list(input_ids["image"].size()), [2, 1, 256])
self.assertEqual(input_ids["text"].dim(), 2)
self.assertEqual(list(input_ids["text"].size()), [2, 128])
self.assertEqual(input_ids["my_random_feature"].dim(), 3)
self.assertEqual(list(input_ids["my_random_feature"].size()),
[2, 4, 128])
position_ids = transformer_input["position_ids"]
test_utils.compare_tensors(position_ids["image"],
torch.tensor([[0], [0]]))
test_utils.compare_tensors(
position_ids["text"],
torch.arange(0, 128).unsqueeze(0).expand((2, 128)))
test_utils.compare_tensors(
position_ids["my_random_feature"],
torch.arange(0, 4).unsqueeze(0).expand((2, 4)),
)
masks = transformer_input["masks"]
test_utils.compare_tensors(masks["image"], torch.tensor([[1], [1]]))
self.assertEqual(masks["text"].sum().item(), 140)
self.assertEqual(masks["my_random_feature"].sum().item(), 6)
segment_ids = transformer_input["segment_ids"]
test_utils.compare_tensors(segment_ids["image"],
torch.tensor([[0], [0]]))
test_utils.compare_tensors(segment_ids["text"],
torch.ones((2, 128)).long())
test_utils.compare_tensors(
segment_ids["my_random_feature"],
torch.full((2, 4), dtype=torch.long, fill_value=3).long(),
)
def test_preprocessing_with_mvit_encoder(self):
encoder_config = OmegaConf.create({
"name":
"pytorchvideo",
"model_name":
"mvit_base_32x3",
"random_init":
True,
"drop_last_n_layers":
0,
"pooler_name":
"cls",
"spatial_size":
224,
"temporal_size":
8,
"head":
None,
"embed_dim_mul": [[1, 2.0], [3, 2.0], [14, 2.0]],
"atten_head_mul": [[1, 2.0], [3, 2.0], [14, 2.0]],
"pool_q_stride_size": [[1, 1, 2, 2], [3, 1, 2, 2], [14, 1, 2, 2]],
"pool_kv_stride_adaptive": [1, 8, 8],
"pool_kvq_kernel": [3, 3, 3],
})
self._image_modality_config = MMFTransformerModalityConfig(
type="image",
key="image",
embedding_dim=768,
position_dim=1,
segment_id=0,
encoder=encoder_config,
)
modalities_config = [
self._image_modality_config, self._text_modality_config
]
config = MMFTransformer.Config(modalities=modalities_config,
num_labels=2)
mmft = build_model(config)
sample_list = SampleList()
sample_list.image = torch.rand((2, 3, 8, 224, 224))
sample_list.text = torch.randint(0, 512, (2, 128))
transformer_input = mmft.preprocess_sample(sample_list)
input_ids = transformer_input["input_ids"]
self.assertEqual(input_ids["image"].dim(), 3)
self.assertEqual(list(input_ids["image"].size()), [2, 1, 768])
self.assertEqual(input_ids["text"].dim(), 2)
self.assertEqual(list(input_ids["text"].size()), [2, 128])
position_ids = transformer_input["position_ids"]
test_utils.compare_tensors(position_ids["image"],
torch.tensor([[0], [0]]))
test_utils.compare_tensors(
position_ids["text"],
torch.arange(0, 128).unsqueeze(0).expand((2, 128)))
masks = transformer_input["masks"]
test_utils.compare_tensors(masks["image"], torch.tensor([[1], [1]]))
test_utils.compare_tensors(masks["text"], torch.ones((2, 128)).long())
segment_ids = transformer_input["segment_ids"]
test_utils.compare_tensors(segment_ids["image"],
torch.tensor([[0], [0]]))
test_utils.compare_tensors(segment_ids["text"],
torch.ones((2, 128)).long())