transformers/tests/models/convnext/test_modeling_tf_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 TensorFlow ConvNext model. """

import inspect
import unittest
from typing import List, Tuple

from transformers import ConvNextConfig
from transformers.testing_utils import require_tf, require_vision, slow
from transformers.utils import cached_property, is_tf_available, is_vision_available

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


if is_tf_available():
    import tensorflow as tf

    from transformers import TFConvNextForImageClassification, TFConvNextModel


if is_vision_available():
    from PIL import Image

    from transformers import ConvNextFeatureExtractor


class TFConvNextModelTester:
    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 = TFConvNextModel(config=config)
        result = model(pixel_values, training=False)
        # 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 = TFConvNextForImageClassification(config)
        result = model(pixel_values, labels=labels, 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 TFConvNextModelTest(TFModelTesterMixin, PipelineTesterMixin, 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 = (TFConvNextModel, TFConvNextForImageClassification) if is_tf_available() else ()
    pipeline_model_mapping = (
        {"feature-extraction": TFConvNextModel, "image-classification": TFConvNextForImageClassification}
        if is_tf_available()
        else {}
    )

    test_pruning = False
    test_onnx = False
    test_resize_embeddings = False
    test_head_masking = False
    has_attentions = False

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

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

    @unittest.skipIf(
        not is_tf_available() or len(tf.config.list_physical_devices("GPU")) == 0,
        reason="TF does not support backprop for grouped convolutions on CPU.",
    )
    def test_keras_fit(self):
        super().test_keras_fit()

    @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.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)

    @unittest.skipIf(
        not is_tf_available() or len(tf.config.list_physical_devices("GPU")) == 0,
        reason="TF does not support backprop for grouped convolutions on CPU.",
    )
    def test_dataset_conversion(self):
        super().test_dataset_conversion()

    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_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)

    # Since ConvNext does not have any attention we need to rewrite this test.
    def test_model_outputs_equivalence(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        def check_equivalence(model, tuple_inputs, dict_inputs, additional_kwargs={}):
            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 tuple_object is None:
                    return
                else:
                    self.assertTrue(
                        all(tf.equal(tuple_object, dict_object)),
                        msg=(
                            "Tuple and dict output are not equal. Difference:"
                            f" {tf.math.reduce_max(tf.abs(tuple_object - dict_object))}"
                        ),
                    )

                recursive_check(tuple_output, dict_output)

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

            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):
        model = TFConvNextModel.from_pretrained("facebook/convnext-tiny-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 TFConvNextModelIntegrationTest(unittest.TestCase):
    @cached_property
    def default_feature_extractor(self):
        return (
            ConvNextFeatureExtractor.from_pretrained("facebook/convnext-tiny-224") if is_vision_available() else None
        )

    @slow
    def test_inference_image_classification_head(self):
        model = TFConvNextForImageClassification.from_pretrained("facebook/convnext-tiny-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([-0.0260, -0.4739, 0.1911])

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