test_modeling_convnext.py

# coding=utf-8
# Copyright 2022 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 PyTorch ConvNext model. """


import inspect
import unittest
from typing import Dict, List, Tuple

from transformers import ConvNextConfig
from transformers.file_utils import cached_property, is_torch_available, is_vision_available
from transformers.testing_utils import require_torch, require_vision, slow, torch_device

from .test_configuration_common import ConfigTester
from .test_modeling_common import ModelTesterMixin, floats_tensor, ids_tensor


if is_torch_available():
    import torch

    from transformers import ConvNextForImageClassification, ConvNextModel
    from transformers.models.convnext.modeling_convnext import CONVNEXT_PRETRAINED_MODEL_ARCHIVE_LIST


if is_vision_available():
    from PIL import Image

    from transformers import AutoFeatureExtractor


class ConvNextModelTester:
    def __init__(
        self,
        parent,
        batch_size=13,
        image_size=32,
        num_channels=3,
        num_stages=4,
        hidden_sizes=[10, 20, 30, 40],
        depths=[2, 2, 3, 2],
        is_training=True,
        use_labels=True,
        intermediate_size=37,
        hidden_act="gelu",
        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.num_channels = num_channels
        self.num_stages = num_stages
        self.hidden_sizes = hidden_sizes
        self.depths = depths
        self.is_training = is_training
        self.use_labels = use_labels
        self.intermediate_size = intermediate_size
        self.hidden_act = hidden_act
        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 ConvNextConfig(
            num_channels=self.num_channels,
            hidden_sizes=self.hidden_sizes,
            depths=self.depths,
            num_stages=self.num_stages,
            hidden_act=self.hidden_act,
            is_decoder=False,
            initializer_range=self.initializer_range,
        )

    def create_and_check_model(self, config, pixel_values, labels):
        model = ConvNextModel(config=config)
        model.to(torch_device)
        model.eval()
        result = model(pixel_values)
        # expected last hidden states: B, C, H // 32, W // 32
        self.parent.assertEqual(
            result.last_hidden_state.shape,
            (self.batch_size, self.hidden_sizes[-1], self.image_size // 32, self.image_size // 32),
        )

    def create_and_check_for_image_classification(self, config, pixel_values, labels):
        config.num_labels = self.type_sequence_label_size
        model = ConvNextForImageClassification(config)
        model.to(torch_device)
        model.eval()
        result = model(pixel_values, labels=labels)
        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_torch
class ConvNextModelTest(ModelTesterMixin, unittest.TestCase):
    """
    Here we also overwrite some of the tests of test_modeling_common.py, as ConvNext does not use input_ids, inputs_embeds,
    attention_mask and seq_length.
    """

    all_model_classes = (
        (
            ConvNextModel,
            ConvNextForImageClassification,
        )
        if is_torch_available()
        else ()
    )

    test_pruning = False
    test_torchscript = False
    test_resize_embeddings = False
    test_head_masking = False

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

    def test_config(self):
        self.create_and_test_config_common_properties()
        self.config_tester.create_and_test_config_to_json_string()
        self.config_tester.create_and_test_config_to_json_file()
        self.config_tester.create_and_test_config_from_and_save_pretrained()
        self.config_tester.create_and_test_config_with_num_labels()
        self.config_tester.check_config_can_be_init_without_params()
        self.config_tester.check_config_arguments_init()

    def create_and_test_config_common_properties(self):
        return

    @unittest.skip(reason="ConvNext does not use inputs_embeds")
    def test_inputs_embeds(self):
        pass

    @unittest.skip(reason="ConvNext does not support input and output embeddings")
    def test_model_common_attributes(self):
        pass

    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.forward)
            # 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)

    @unittest.skip(reason="Model doesn't have attention layers")
    def test_attention_outputs(self):
        pass

    def test_hidden_states_output(self):
        def check_hidden_states_output(inputs_dict, config, model_class):
            model = model_class(config)
            model.to(torch_device)
            model.eval()

            with torch.no_grad():
                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_stages = self.model_tester.num_stages
            self.assertEqual(len(hidden_states), expected_num_stages + 1)

            # ConvNext's feature maps are of shape (batch_size, num_channels, height, width)
            self.assertListEqual(
                list(hidden_states[0].shape[-2:]),
                [self.model_tester.image_size // 4, self.model_tester.image_size // 4],
            )

        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_retain_grad_hidden_states_attentions(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
        config.output_hidden_states = True
        config.output_attentions = True

        # no need to test all models as different heads yield the same functionality
        model_class = self.all_model_classes[0]
        model = model_class(config)
        model.to(torch_device)

        inputs = self._prepare_for_class(inputs_dict, model_class)
        outputs = model(**inputs)
        output = outputs[0]

        hidden_states = outputs.hidden_states[0]
        hidden_states.retain_grad()

        output.flatten()[0].backward(retain_graph=True)

        self.assertIsNotNone(hidden_states.grad)

    def test_model_outputs_equivalence(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        def set_nan_tensor_to_zero(t):
            t[t != t] = 0
            return t

        def check_equivalence(model, tuple_inputs, dict_inputs, additional_kwargs={}):
            with torch.no_grad():
                tuple_output = model(**tuple_inputs, return_dict=False, **additional_kwargs)
                dict_output = model(**dict_inputs, return_dict=True, **additional_kwargs).to_tuple()

                def recursive_check(tuple_object, dict_object):
                    if isinstance(tuple_object, (List, Tuple)):
                        for tuple_iterable_value, dict_iterable_value in zip(tuple_object, dict_object):
                            recursive_check(tuple_iterable_value, dict_iterable_value)
                    elif isinstance(tuple_object, Dict):
                        for tuple_iterable_value, dict_iterable_value in zip(
                            tuple_object.values(), dict_object.values()
                        ):
                            recursive_check(tuple_iterable_value, dict_iterable_value)
                    elif tuple_object is None:
                        return
                    else:
                        self.assertTrue(
                            torch.allclose(
                                set_nan_tensor_to_zero(tuple_object), set_nan_tensor_to_zero(dict_object), atol=1e-5
                            ),
                            msg=f"Tuple and dict output are not equal. Difference: {torch.max(torch.abs(tuple_object - dict_object))}. Tuple has `nan`: {torch.isnan(tuple_object).any()} and `inf`: {torch.isinf(tuple_object)}. Dict has `nan`: {torch.isnan(dict_object).any()} and `inf`: {torch.isinf(dict_object)}.",
                        )

                recursive_check(tuple_output, dict_output)

        for model_class in self.all_model_classes:
            model = model_class(config)
            model.to(torch_device)
            model.eval()

            tuple_inputs = self._prepare_for_class(inputs_dict, model_class)
            dict_inputs = self._prepare_for_class(inputs_dict, model_class)
            check_equivalence(model, tuple_inputs, dict_inputs)

            tuple_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
            dict_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
            check_equivalence(model, tuple_inputs, dict_inputs)

            tuple_inputs = self._prepare_for_class(inputs_dict, model_class)
            dict_inputs = self._prepare_for_class(inputs_dict, model_class)
            check_equivalence(model, tuple_inputs, dict_inputs, {"output_hidden_states": True})

            tuple_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
            dict_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
            check_equivalence(model, tuple_inputs, dict_inputs, {"output_hidden_states": True})

    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):
        for model_name in CONVNEXT_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
            model = ConvNextModel.from_pretrained(model_name)
            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_torch
@require_vision
class ConvNextModelIntegrationTest(unittest.TestCase):
    @cached_property
    def default_feature_extractor(self):
        return AutoFeatureExtractor.from_pretrained("facebook/convnext-tiny-224") if is_vision_available() else None

    @slow
    def test_inference_image_classification_head(self):
        model = ConvNextForImageClassification.from_pretrained("facebook/convnext-tiny-224").to(torch_device)

        feature_extractor = self.default_feature_extractor
        image = prepare_img()
        inputs = feature_extractor(images=image, return_tensors="pt").to(torch_device)

        # forward pass
        with torch.no_grad():
            outputs = model(**inputs)

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

        expected_slice = torch.tensor([-0.0260, -0.4739, 0.1911]).to(torch_device)

        self.assertTrue(torch.allclose(outputs.logits[0, :3], expected_slice, atol=1e-4))