def check_hidden_states_output(inputs_dict, config, model_class):
model = model_class(config)
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)
# ViTMAE has a different seq_length
image_size = to_2tuple(self.model_tester.image_size)
patch_size = to_2tuple(self.model_tester.patch_size)
num_patches = (image_size[1] // patch_size[1]) * (image_size[0] //
patch_size[0])
seq_length = int(
math.ceil((1 - config.mask_ratio) * (num_patches + 1)))
self.assertListEqual(
list(hidden_states[0].shape[-2:]),
[seq_length, self.model_tester.hidden_size],
)
def create_and_check_for_pretraining(self, config, pixel_values, labels):
model = TFViTMAEForPreTraining(config)
result = model(pixel_values, training=False)
# expected sequence length = num_patches
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])
expected_seq_len = num_patches
expected_num_channels = self.patch_size**2 * self.num_channels
self.parent.assertEqual(result.logits.shape, (self.batch_size, expected_seq_len, expected_num_channels))
def create_and_check_model(self, config, pixel_values, labels):
model = TFViTMAEModel(config=config)
result = model(pixel_values, training=False)
# expected sequence length = (num_patches + 1) * (1 - config.mask_ratio), rounded above
# (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])
expected_seq_len = int(math.ceil((1 - config.mask_ratio) * (num_patches + 1)))
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, expected_seq_len, self.hidden_size))
def test_attention_outputs(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
config.return_dict = True
# in ViTMAE, the seq_len equals (number of patches + 1) * (1 - mask_ratio), rounded above
image_size = to_2tuple(self.model_tester.image_size)
patch_size = to_2tuple(self.model_tester.patch_size)
num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
seq_len = int(math.ceil((1 - config.mask_ratio) * (num_patches + 1)))
encoder_seq_length = getattr(self.model_tester, "encoder_seq_length", seq_len)
encoder_key_length = getattr(self.model_tester, "key_length", encoder_seq_length)
chunk_length = getattr(self.model_tester, "chunk_length", None)
if chunk_length is not None and hasattr(self.model_tester, "num_hashes"):
encoder_seq_length = encoder_seq_length * self.model_tester.num_hashes
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(config)
outputs = model(**self._prepare_for_class(inputs_dict, model_class), training=False)
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)
outputs = model(**self._prepare_for_class(inputs_dict, model_class), training=False)
attentions = outputs.encoder_attentions if config.is_encoder_decoder else outputs.attentions
self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)
if chunk_length is not None:
self.assertListEqual(
list(attentions[0].shape[-4:]),
[self.model_tester.num_attention_heads, encoder_seq_length, chunk_length, encoder_key_length],
)
else:
self.assertListEqual(
list(attentions[0].shape[-3:]),
[self.model_tester.num_attention_heads, encoder_seq_length, encoder_key_length],
)
out_len = len(outputs)
# Check attention is always last and order is fine
inputs_dict["output_attentions"] = True
inputs_dict["output_hidden_states"] = True
model = model_class(config)
outputs = model(**self._prepare_for_class(inputs_dict, model_class), training=False)
if hasattr(self.model_tester, "num_hidden_states_types"):
added_hidden_states = self.model_tester.num_hidden_states_types
elif self.is_encoder_decoder:
added_hidden_states = 2
else:
added_hidden_states = 1
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)
if chunk_length is not None:
self.assertListEqual(
list(self_attentions[0].shape[-4:]),
[self.model_tester.num_attention_heads, encoder_seq_length, chunk_length, encoder_key_length],
)
else:
self.assertListEqual(
list(self_attentions[0].shape[-3:]),
[self.model_tester.num_attention_heads, encoder_seq_length, encoder_key_length],
)