test_modeling_flax_vit.py

# Copyright 2021 The HuggingFace 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.

import inspect
import unittest

import numpy as np

from transformers import ViTConfig, is_flax_available
from transformers.testing_utils import require_flax, slow

from ..test_configuration_common import ConfigTester
from ..test_modeling_flax_common import FlaxModelTesterMixin, floats_tensor


if is_flax_available():

    import jax
    from transformers.models.vit.modeling_flax_vit import FlaxViTForImageClassification, FlaxViTModel


class FlaxViTModelTester(unittest.TestCase):
    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,
    ):
        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

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

        config = 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,
        )

        return config, pixel_values

    def create_and_check_model(self, config, pixel_values, labels):

        model = FlaxViTModel(config=config)
        result = model(pixel_values)
        # expected sequence length = num_patches + 1 (we add 1 for the [CLS] token)
        image_size = (self.image_size, self.image_size)
        patch_size = (self.patch_size, 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))

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


@require_flax
class FlaxViTModelTest(FlaxModelTesterMixin, unittest.TestCase):

    all_model_classes = (FlaxViTModel, FlaxViTForImageClassification) if is_flax_available() else ()

    def setUp(self) -> None:
        self.model_tester = FlaxViTModelTester(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()

    # We need to override this test because in ViT, the seq_len equals the number of patches + 1
    # we compute that here
    def test_attention_outputs(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
        config.return_dict = True

        num_patches = (config.image_size // config.patch_size) ** 2
        seq_length = num_patches + 1

        for model_class in self.all_model_classes:
            inputs_dict["output_attentions"] = True
            inputs_dict["output_hidden_states"] = False
            model = model_class(config)
            outputs = model(**self._prepare_for_class(inputs_dict, model_class))
            attentions = 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))
            attentions = 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, seq_length, seq_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))

            added_hidden_states = 1
            self.assertEqual(out_len + added_hidden_states, len(outputs))

            self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)
            self.assertListEqual(
                list(attentions[0].shape[-3:]),
                [self.model_tester.num_attention_heads, seq_length, seq_length],
            )

    # We neeed to override this test because ViT's forward signature is different than text models.
    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)

    # We neeed to override this test because ViT expects pixel_values instead of input_ids
    def test_jit_compilation(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            with self.subTest(model_class.__name__):
                prepared_inputs_dict = self._prepare_for_class(inputs_dict, model_class)
                model = model_class(config)

                @jax.jit
                def model_jitted(pixel_values, **kwargs):
                    return model(pixel_values=pixel_values, **kwargs)

                with self.subTest("JIT Enabled"):
                    jitted_outputs = model_jitted(**prepared_inputs_dict).to_tuple()

                with self.subTest("JIT Disabled"):
                    with jax.disable_jit():
                        outputs = model_jitted(**prepared_inputs_dict).to_tuple()

                self.assertEqual(len(outputs), len(jitted_outputs))
                for jitted_output, output in zip(jitted_outputs, outputs):
                    self.assertEqual(jitted_output.shape, output.shape)

    # We need to override this test because in ViT, the seq_len equals the number of patches + 1
    # we compute that here
    def test_hidden_states_output(self):
        def check_hidden_states_output(inputs_dict, config, model_class):
            model = model_class(config)
            num_patches = (config.image_size // config.patch_size) ** 2
            seq_length = num_patches + 1  # we add 1 for the [CLS] token

            outputs = model(**self._prepare_for_class(inputs_dict, model_class))
            hidden_states = outputs.hidden_states

            self.assertEqual(len(hidden_states), self.model_tester.num_hidden_layers + 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)

    @slow
    def test_model_from_pretrained(self):
        for model_class_name in self.all_model_classes:
            model = model_class_name.from_pretrained("google/vit-base-patch16-224")
            outputs = model(np.ones((1, 3, 224, 224)))
            self.assertIsNotNone(outputs)