test_modeling_tf_resnet.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 ResNet model. """


import inspect
import unittest

import numpy as np

from transformers import ResNetConfig
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


if is_tf_available():
    import tensorflow as tf

    from transformers import TFResNetForImageClassification, TFResNetModel
    from transformers.models.resnet.modeling_tf_resnet import TF_RESNET_PRETRAINED_MODEL_ARCHIVE_LIST


if is_vision_available():
    from PIL import Image

    from transformers import AutoFeatureExtractor


class TFResNetModelTester:
    def __init__(
        self,
        parent,
        batch_size=3,
        image_size=32,
        num_channels=3,
        embeddings_size=10,
        hidden_sizes=[10, 20, 30, 40],
        depths=[1, 1, 2, 1],
        is_training=True,
        use_labels=True,
        hidden_act="relu",
        num_labels=3,
        scope=None,
    ):
        self.parent = parent
        self.batch_size = batch_size
        self.image_size = image_size
        self.num_channels = num_channels
        self.embeddings_size = embeddings_size
        self.hidden_sizes = hidden_sizes
        self.depths = depths
        self.is_training = is_training
        self.use_labels = use_labels
        self.hidden_act = hidden_act
        self.num_labels = num_labels
        self.scope = scope
        self.num_stages = len(hidden_sizes)

    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.num_labels)

        config = self.get_config()

        return config, pixel_values, labels

    def get_config(self):
        return ResNetConfig(
            num_channels=self.num_channels,
            embeddings_size=self.embeddings_size,
            hidden_sizes=self.hidden_sizes,
            depths=self.depths,
            hidden_act=self.hidden_act,
            num_labels=self.num_labels,
            image_size=self.image_size,
        )

    def create_and_check_model(self, config, pixel_values, labels):
        model = TFResNetModel(config=config)
        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.num_labels
        model = TFResNetForImageClassification(config)
        result = model(pixel_values, labels=labels)
        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_labels))

    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 TFResNetModelTest(TFModelTesterMixin, unittest.TestCase):
    """
    Here we also overwrite some of the tests of test_modeling_common.py, as ResNet does not use input_ids, inputs_embeds,
    attention_mask and seq_length.
    """

    all_model_classes = (TFResNetModel, TFResNetForImageClassification) if is_tf_available() else ()

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

    def setUp(self):
        self.model_tester = TFResNetModelTester(self)
        self.config_tester = ConfigTester(self, config_class=ResNetConfig, has_text_modality=False)

    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="ResNet does not use inputs_embeds")
    def test_inputs_embeds(self):
        pass

    @unittest.skip(reason="ResNet 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)

    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)

            # ResNet'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()
        layers_type = ["basic", "bottleneck"]
        for model_class in self.all_model_classes:
            for layer_type in layers_type:
                config.layer_type = layer_type
                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):
        for model_name in TF_RESNET_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
            model = TFResNetModel.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_tf
@require_vision
class TFResNetModelIntegrationTest(unittest.TestCase):
    @cached_property
    def default_feature_extractor(self):
        return (
            AutoFeatureExtractor.from_pretrained(TF_RESNET_PRETRAINED_MODEL_ARCHIVE_LIST[0])
            if is_vision_available()
            else None
        )

    @slow
    def test_inference_image_classification_head(self):
        model = TFResNetForImageClassification.from_pretrained(TF_RESNET_PRETRAINED_MODEL_ARCHIVE_LIST[0])

        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([-11.1069, -9.7877, -8.3777])

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