def test_encoder_decoder_save_load_from_encoder_decoder(self):
config = self.get_encoder_decoder_config_small()
# create two random ViT/GPT2 models for vit-gpt2 & initialize weights (+cross_attention weights)
encoder = TFViTModel(config.encoder)
encoder(encoder.dummy_inputs)
decoder = TFGPT2LMHeadModel(config.decoder)
decoder(decoder.dummy_inputs)
encoder_decoder_orig = TFVisionEncoderDecoderModel(encoder=encoder,
decoder=decoder)
pixel_values = floats_tensor([
13,
encoder.config.num_channels,
encoder.config.image_size,
encoder.config.image_size,
])
decoder_input_ids = ids_tensor([13, 1], decoder.config.vocab_size)
logits_orig = encoder_decoder_orig(
pixel_values=pixel_values,
decoder_input_ids=decoder_input_ids).logits
with tempfile.TemporaryDirectory() as tmp_dirname:
encoder_path = os.path.join(tmp_dirname, "encoder")
decoder_path = os.path.join(tmp_dirname, "decoder")
encoder.save_pretrained(encoder_path)
decoder.save_pretrained(decoder_path)
encoder_decoder = TFVisionEncoderDecoderModel.from_encoder_decoder_pretrained(
encoder_path, decoder_path)
logits_1 = encoder_decoder(pixel_values=pixel_values,
decoder_input_ids=decoder_input_ids).logits
self.assertTrue(
logits_orig.numpy().sum() - logits_1.numpy().sum() < 1e-3)
max_diff = np.max(np.abs(logits_1.numpy() - logits_orig.numpy()))
self.assertAlmostEqual(max_diff, 0.0, places=4)
with tempfile.TemporaryDirectory() as tmp_dirname:
encoder_decoder.save_pretrained(tmp_dirname)
encoder_decoder = TFVisionEncoderDecoderModel.from_pretrained(
tmp_dirname)
logits_2 = encoder_decoder(pixel_values=pixel_values,
decoder_input_ids=decoder_input_ids).logits
max_diff = np.max(np.abs(logits_2.numpy() - logits_orig.numpy()))
self.assertAlmostEqual(max_diff, 0.0, places=4)
def create_and_check_model(self, config, pixel_values, labels):
model = TFViTModel(config=config)
result = model(pixel_values, training=False)
# expected sequence length = num_patches + 1 (we add 1 for the [CLS] token)
image_size = to_2tuple(self.image_size)
patch_size = to_2tuple(self.patch_size)
num_patches = (image_size[1] // patch_size[1]) * (image_size[0] //
patch_size[0])
self.parent.assertEqual(
result.last_hidden_state.shape,
(self.batch_size, num_patches + 1, self.hidden_size))
# Test with an image with different size than the one specified in config.
image_size = self.image_size // 2
pixel_values = pixel_values[:, :, :image_size, :image_size]
result = model(pixel_values,
interpolate_pos_encoding=True,
training=False)
# expected sequence length = num_patches + 1 (we add 1 for the [CLS] token)
image_size = to_2tuple(image_size)
num_patches = (image_size[1] // patch_size[1]) * (image_size[0] //
patch_size[0])
self.parent.assertEqual(
result.last_hidden_state.shape,
(self.batch_size, num_patches + 1, self.hidden_size))
def create_and_check_model(self, config, pixel_values, labels):
model = TFViTModel(config=config)
result = model(pixel_values, training=False)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
# Test with an image with different size than the one specified in config.
image_size = self.image_size // 2
pixel_values = pixel_values[:, :, :image_size, :image_size]
result = model(pixel_values, interpolate_pos_encoding=True, training=False)
seq_length = (image_size // self.patch_size) ** 2 + 1
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, seq_length, self.hidden_size))
def test_model_from_pretrained(self):
model = TFViTModel.from_pretrained("google/vit-base-patch16-224")
self.assertIsNotNone(model)
def get_encoder_decoder_models(self):
encoder_model = TFViTModel.from_pretrained("google/vit-base-patch16-224-in21k", name="encoder")
decoder_model = TFGPT2LMHeadModel.from_pretrained("../gpt2", config=self.get_decoder_config(), name="decoder")
return {"encoder": encoder_model, "decoder": decoder_model}
def get_encoder_decoder_model(self, config, decoder_config):
encoder_model = TFViTModel(config, name="encoder")
decoder_model = TFGPT2LMHeadModel(decoder_config, name="decoder")
return encoder_model, decoder_model
