test_modeling_tf_vit.py

# coding=utf-8
# 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 TensorFlow ViT model. """


import inspect
import os
import tempfile
import unittest

from transformers import ViTConfig
from transformers.file_utils import cached_property, is_tf_available, is_vision_available
from transformers.testing_utils import require_tf, require_vision, slow, tooslow

from .test_configuration_common import ConfigTester
from .test_modeling_tf_common import TFModelTesterMixin, floats_tensor, ids_tensor


if is_tf_available():
    import tensorflow as tf

    from transformers import TFViTForImageClassification, TFViTModel
    from transformers.models.vit.modeling_tf_vit import to_2tuple


if is_vision_available():
    from PIL import Image

    from transformers import ViTFeatureExtractor


class TFViTModelTester:
    def __init__(
        self,
        parent,
        batch_size=13,
        image_size=30,
        patch_size=2,
        num_channels=3,
        is_training=True,
        use_labels=True,
        hidden_size=32,
        num_hidden_layers=5,
        num_attention_heads=4,
        intermediate_size=37,
        hidden_act="gelu",
        hidden_dropout_prob=0.1,
        attention_probs_dropout_prob=0.1,
        type_sequence_label_size=10,
        initializer_range=0.02,
        num_labels=3,
        scope=None,
    ):
        self.parent = parent
        self.batch_size = batch_size
        self.image_size = image_size
        self.patch_size = patch_size
        self.num_channels = num_channels
        self.is_training = is_training
        self.use_labels = use_labels
        self.hidden_size = hidden_size
        self.num_hidden_layers = num_hidden_layers
        self.num_attention_heads = num_attention_heads
        self.intermediate_size = intermediate_size
        self.hidden_act = hidden_act
        self.hidden_dropout_prob = hidden_dropout_prob
        self.attention_probs_dropout_prob = attention_probs_dropout_prob
        self.type_sequence_label_size = type_sequence_label_size
        self.initializer_range = initializer_range
        self.scope = scope

    def prepare_config_and_inputs(self):
        pixel_values = floats_tensor([self.batch_size, self.num_channels, self.image_size, self.image_size])

        labels = None
        if self.use_labels:
            labels = ids_tensor([self.batch_size], self.type_sequence_label_size)

        config = self.get_config()

        return config, pixel_values, labels

    def get_config(self):
        return ViTConfig(
            image_size=self.image_size,
            patch_size=self.patch_size,
            num_channels=self.num_channels,
            hidden_size=self.hidden_size,
            num_hidden_layers=self.num_hidden_layers,
            num_attention_heads=self.num_attention_heads,
            intermediate_size=self.intermediate_size,
            hidden_act=self.hidden_act,
            hidden_dropout_prob=self.hidden_dropout_prob,
            attention_probs_dropout_prob=self.attention_probs_dropout_prob,
            is_decoder=False,
            initializer_range=self.initializer_range,
        )

    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_for_image_classification(self, config, pixel_values, labels):
        config.num_labels = self.type_sequence_label_size
        model = TFViTForImageClassification(config)
        result = model(pixel_values, labels=labels, training=False)
        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.type_sequence_label_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)
        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.type_sequence_label_size))

    def prepare_config_and_inputs_for_common(self):
        config_and_inputs = self.prepare_config_and_inputs()
        config, pixel_values, labels = config_and_inputs
        inputs_dict = {"pixel_values": pixel_values}
        return config, inputs_dict


@require_tf
class TFViTModelTest(TFModelTesterMixin, unittest.TestCase):
    """
    Here we also overwrite some of the tests of test_modeling_tf_common.py, as ViT does not use input_ids, inputs_embeds,
    attention_mask and seq_length.
    """

    all_model_classes = (TFViTModel, TFViTForImageClassification) if is_tf_available() else ()

    test_resize_embeddings = False
    test_head_masking = False
    test_onnx = False

    def setUp(self):
        self.model_tester = TFViTModelTester(self)
        self.config_tester = ConfigTester(self, config_class=ViTConfig, has_text_modality=False, hidden_size=37)

    def test_config(self):
        self.config_tester.run_common_tests()

    def test_inputs_embeds(self):
        # ViT does not use inputs_embeds
        pass

    def test_graph_mode_with_inputs_embeds(self):
        # ViT does not use inputs_embeds
        pass

    def test_model_common_attributes(self):
        config, _ = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            model = model_class(config)
            self.assertIsInstance(model.get_input_embeddings(), (tf.keras.layers.Layer))
            x = model.get_output_embeddings()
            self.assertTrue(x is None or isinstance(x, tf.keras.layers.Layer))

    def test_forward_signature(self):
        config, _ = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            model = model_class(config)
            signature = inspect.signature(model.call)
            # signature.parameters is an OrderedDict => so arg_names order is deterministic
            arg_names = [*signature.parameters.keys()]

            expected_arg_names = ["pixel_values"]
            self.assertListEqual(arg_names[:1], expected_arg_names)

    def test_model(self):
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
        self.model_tester.create_and_check_model(*config_and_inputs)

    # overwrite from common since `encoder_seq_length` and `encoder_key_length` are calculated
    # in a different way than in text models.
    @tooslow
    def test_saved_model_creation_extended(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
        config.output_hidden_states = True
        config.output_attentions = True

        if hasattr(config, "use_cache"):
            config.use_cache = True

        # in ViT, the seq_len equals the number of patches + 1 (we add 1 for the [CLS] token)
        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 = 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)

        for model_class in self.all_model_classes:
            class_inputs_dict = self._prepare_for_class(inputs_dict, model_class)
            model = model_class(config)
            num_out = len(model(class_inputs_dict))

            with tempfile.TemporaryDirectory() as tmpdirname:
                model.save_pretrained(tmpdirname, saved_model=True)
                saved_model_dir = os.path.join(tmpdirname, "saved_model", "1")
                model = tf.keras.models.load_model(saved_model_dir)
                outputs = model(class_inputs_dict)

                if self.is_encoder_decoder:
                    output_hidden_states = outputs["encoder_hidden_states"]
                    output_attentions = outputs["encoder_attentions"]
                else:
                    output_hidden_states = outputs["hidden_states"]
                    output_attentions = outputs["attentions"]

                self.assertEqual(len(outputs), num_out)

                expected_num_layers = getattr(
                    self.model_tester, "expected_num_hidden_layers", self.model_tester.num_hidden_layers + 1
                )

                self.assertEqual(len(output_hidden_states), expected_num_layers)
                self.assertListEqual(
                    list(output_hidden_states[0].shape[-2:]),
                    [seq_len, self.model_tester.hidden_size],
                )

                self.assertEqual(len(output_attentions), self.model_tester.num_hidden_layers)
                self.assertListEqual(
                    list(output_attentions[0].shape[-3:]),
                    [self.model_tester.num_attention_heads, encoder_seq_length, encoder_key_length],
                )

    def test_attention_outputs(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
        config.return_dict = True

        # in ViT, the seq_len equals the number of patches + 1 (we add 1 for the [CLS] token)
        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 = 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)

        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)

            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)
            self.assertListEqual(
                list(self_attentions[0].shape[-3:]),
                [self.model_tester.num_attention_heads, encoder_seq_length, encoder_key_length],
            )

    def test_hidden_states_output(self):
        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)

            # ViT 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 = num_patches + 1

            self.assertListEqual(
                list(hidden_states[0].shape[-2:]),
                [seq_length, self.model_tester.hidden_size],
            )

        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            inputs_dict["output_hidden_states"] = True
            check_hidden_states_output(inputs_dict, config, model_class)

            # check that output_hidden_states also work using config
            del inputs_dict["output_hidden_states"]
            config.output_hidden_states = True

            check_hidden_states_output(inputs_dict, config, model_class)

    def test_for_image_classification(self):
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
        self.model_tester.create_and_check_for_image_classification(*config_and_inputs)

    @slow
    def test_model_from_pretrained(self):

        model = TFViTModel.from_pretrained("google/vit-base-patch16-224")
        self.assertIsNotNone(model)


# We will verify our results on an image of cute cats
def prepare_img():
    image = Image.open("./tests/fixtures/tests_samples/COCO/000000039769.png")
    return image


@require_tf
@require_vision
class TFViTModelIntegrationTest(unittest.TestCase):
    @cached_property
    def default_feature_extractor(self):
        return ViTFeatureExtractor.from_pretrained("google/vit-base-patch16-224") if is_vision_available() else None

    @slow
    def test_inference_image_classification_head(self):
        model = TFViTForImageClassification.from_pretrained("google/vit-base-patch16-224")

        feature_extractor = self.default_feature_extractor
        image = prepare_img()
        inputs = feature_extractor(images=image, return_tensors="tf")

        # forward pass
        outputs = model(**inputs)

        # verify the logits
        expected_shape = tf.TensorShape((1, 1000))
        self.assertEqual(outputs.logits.shape, expected_shape)

        expected_slice = tf.constant([-0.2744, 0.8215, -0.0836])

        tf.debugging.assert_near(outputs.logits[0, :3], expected_slice, atol=1e-4)