def create_and_check_model(self, config, pixel_values, labels):
model = SwinModel(config=config)
model.to(torch_device)
model.eval()
result = model(pixel_values)
num_features = int(config.embed_dim * 2**(len(config.depths) - 1))
self.parent.assertEqual(result.last_hidden_state.shape,
(self.batch_size, num_features))
def create_and_check_model(self, config, pixel_values, labels):
model = SwinModel(config=config)
model.to(torch_device)
model.eval()
result = model(pixel_values)
expected_seq_len = ((config.image_size // config.patch_size) ** 2) // (4 ** (len(config.depths) - 1))
expected_dim = int(config.embed_dim * 2 ** (len(config.depths) - 1))
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, expected_seq_len, expected_dim))
def create_and_check_model(self, config, pixel_values, labels):
model = SwinModel(config=config)
model.to(torch_device)
model.eval()
result = model(pixel_values)
# since the model we're testing only consists of a single layer, expected_seq_len = number of patches
expected_seq_len = (config.image_size // config.patch_size) ** 2
expected_dim = int(config.embed_dim * 2 ** (len(config.depths) - 1))
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, expected_seq_len, expected_dim))
def test_model_from_pretrained(self):
for model_name in SWIN_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
model = SwinModel.from_pretrained(model_name)
self.assertIsNotNone(model)
def get_encoder_decoder_model(self, config, decoder_config):
encoder_model = SwinModel(config).eval()
decoder_model = BartForCausalLM(decoder_config).eval()
return encoder_model, decoder_model