init

transformers/tests/models/imagegpt/test_image_processing_imagegpt.py

@@ -0,0 +1,376 @@
+# Copyright 2021 HuggingFace Inc.
+#
+# 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 json
+import os
+import tempfile
+import unittest
+
+import numpy as np
+import pytest
+import requests
+from datasets import load_dataset
+from packaging import version
+
+from transformers import AutoImageProcessor
+from transformers.testing_utils import (
+    check_json_file_has_correct_format,
+    require_torch,
+    require_torch_accelerator,
+    require_vision,
+    slow,
+    torch_device,
+)
+from transformers.utils import is_torch_available, is_torchvision_available, is_vision_available
+
+from ...test_image_processing_common import ImageProcessingTestMixin, prepare_image_inputs
+
+
+if is_torch_available():
+    import torch
+
+if is_vision_available():
+    from PIL import Image
+
+    from transformers import ImageGPTImageProcessor
+
+    if is_torchvision_available():
+        from transformers import ImageGPTImageProcessorFast
+
+
+class ImageGPTImageProcessingTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=7,
+        num_channels=3,
+        image_size=18,
+        min_resolution=30,
+        max_resolution=400,
+        do_resize=True,
+        size=None,
+        do_normalize=True,
+    ):
+        size = size if size is not None else {"height": 18, "width": 18}
+        self.parent = parent
+        self.batch_size = batch_size
+        self.num_channels = num_channels
+        self.image_size = image_size
+        self.min_resolution = min_resolution
+        self.max_resolution = max_resolution
+        self.do_resize = do_resize
+        self.size = size
+        self.do_normalize = do_normalize
+
+    def prepare_image_processor_dict(self):
+        return {
+            # here we create 2 clusters for the sake of simplicity
+            "clusters": np.asarray(
+                [
+                    [0.8866443634033203, 0.6618829369544983, 0.3891746401786804],
+                    [-0.6042559146881104, -0.02295008860528469, 0.5423797369003296],
+                ]
+            ),
+            "do_resize": self.do_resize,
+            "size": self.size,
+            "do_normalize": self.do_normalize,
+        }
+
+    def expected_output_image_shape(self, images):
+        return (self.size["height"] * self.size["width"],)
+
+    def prepare_image_inputs(self, equal_resolution=False, numpify=False, torchify=False):
+        return prepare_image_inputs(
+            batch_size=self.batch_size,
+            num_channels=self.num_channels,
+            min_resolution=self.min_resolution,
+            max_resolution=self.max_resolution,
+            equal_resolution=equal_resolution,
+            numpify=numpify,
+            torchify=torchify,
+        )
+
+
+@require_torch
+@require_vision
+class ImageGPTImageProcessingTest(ImageProcessingTestMixin, unittest.TestCase):
+    image_processing_class = ImageGPTImageProcessor if is_vision_available() else None
+    fast_image_processing_class = ImageGPTImageProcessorFast if is_torchvision_available() else None
+
+    def setUp(self):
+        super().setUp()
+        self.image_processor_tester = ImageGPTImageProcessingTester(self)
+
+    @property
+    def image_processor_dict(self):
+        return self.image_processor_tester.prepare_image_processor_dict()
+
+    def test_image_processor_properties(self):
+        for image_processing_class in self.image_processor_list:
+            image_processing = image_processing_class(**self.image_processor_dict)
+            self.assertTrue(hasattr(image_processing, "clusters"))
+            self.assertTrue(hasattr(image_processing, "do_resize"))
+            self.assertTrue(hasattr(image_processing, "size"))
+            self.assertTrue(hasattr(image_processing, "do_normalize"))
+
+    def test_image_processor_from_dict_with_kwargs(self):
+        for image_processing_class in self.image_processor_list:
+            image_processor = image_processing_class.from_dict(self.image_processor_dict)
+            self.assertEqual(image_processor.size, {"height": 18, "width": 18})
+
+            image_processor = image_processing_class.from_dict(self.image_processor_dict, size=42)
+            self.assertEqual(image_processor.size, {"height": 42, "width": 42})
+
+    def test_image_processor_to_json_string(self):
+        for image_processing_class in self.image_processor_list:
+            image_processor = image_processing_class(**self.image_processor_dict)
+            obj = json.loads(image_processor.to_json_string())
+            for key, value in self.image_processor_dict.items():
+                if key == "clusters":
+                    self.assertTrue(np.array_equal(value, obj[key]))
+                else:
+                    self.assertEqual(obj[key], value)
+
+    def test_image_processor_to_json_file(self):
+        for image_processing_class in self.image_processor_list:
+            image_processor_first = image_processing_class(**self.image_processor_dict)
+
+            with tempfile.TemporaryDirectory() as tmpdirname:
+                json_file_path = os.path.join(tmpdirname, "image_processor.json")
+                image_processor_first.to_json_file(json_file_path)
+                image_processor_second = image_processing_class.from_json_file(json_file_path).to_dict()
+
+            image_processor_first = image_processor_first.to_dict()
+            for key, value in image_processor_first.items():
+                if key == "clusters":
+                    self.assertTrue(np.array_equal(value, image_processor_second[key]))
+                else:
+                    self.assertEqual(image_processor_first[key], value)
+
+    def test_image_processor_from_and_save_pretrained(self):
+        for image_processing_class in self.image_processor_list:
+            image_processor_first = image_processing_class(**self.image_processor_dict)
+
+            with tempfile.TemporaryDirectory() as tmpdirname:
+                image_processor_first.save_pretrained(tmpdirname)
+                image_processor_second = image_processing_class.from_pretrained(tmpdirname).to_dict()
+
+            image_processor_first = image_processor_first.to_dict()
+            for key, value in image_processor_first.items():
+                if key == "clusters":
+                    self.assertTrue(np.array_equal(value, image_processor_second[key]))
+                else:
+                    self.assertEqual(value, value)
+
+    def test_image_processor_save_load_with_autoimageprocessor(self):
+        for image_processing_class in self.image_processor_list:
+            image_processor_first = image_processing_class(**self.image_processor_dict)
+
+            with tempfile.TemporaryDirectory() as tmpdirname:
+                saved_file = image_processor_first.save_pretrained(tmpdirname)[0]
+                check_json_file_has_correct_format(saved_file)
+
+                image_processor_second = AutoImageProcessor.from_pretrained(tmpdirname)
+
+            image_processor_first = image_processor_first.to_dict()
+            image_processor_second = image_processor_second.to_dict()
+
+            for key, value in image_processor_first.items():
+                if key == "clusters":
+                    self.assertTrue(np.array_equal(value, image_processor_second[key]))
+                else:
+                    self.assertEqual(value, value)
+
+    @unittest.skip(reason="ImageGPT requires clusters at initialization")
+    def test_init_without_params(self):
+        pass
+
+    # Override the test from ImageProcessingTestMixin as ImageGPT model takes input_ids as input
+    def test_call_pil(self):
+        for image_processing_class in self.image_processor_list:
+            # Initialize image_processing
+            image_processing = image_processing_class(**self.image_processor_dict)
+            # create random PIL images
+            image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=False)
+            for image in image_inputs:
+                self.assertIsInstance(image, Image.Image)
+
+            # Test not batched input
+            encoded_images = image_processing(image_inputs[0], return_tensors="pt").input_ids
+            expected_output_image_shape = self.image_processor_tester.expected_output_image_shape(encoded_images)
+            self.assertEqual(tuple(encoded_images.shape), (1, *expected_output_image_shape))
+
+            # Test batched
+            encoded_images = image_processing(image_inputs, return_tensors="pt").input_ids
+            self.assertEqual(
+                tuple(encoded_images.shape), (self.image_processor_tester.batch_size, *expected_output_image_shape)
+            )
+
+    # Override the test from ImageProcessingTestMixin as ImageGPT model takes input_ids as input
+    def test_call_numpy(self):
+        for image_processing_class in self.image_processor_list:
+            # Initialize image_processing
+            image_processing = image_processing_class(**self.image_processor_dict)
+            # create random numpy tensors
+            image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=False, numpify=True)
+            for image in image_inputs:
+                self.assertIsInstance(image, np.ndarray)
+
+            # Test not batched input
+            encoded_images = image_processing(image_inputs[0], return_tensors="pt").input_ids
+            expected_output_image_shape = self.image_processor_tester.expected_output_image_shape(encoded_images)
+            self.assertEqual(tuple(encoded_images.shape), (1, *expected_output_image_shape))
+
+            # Test batched
+            encoded_images = image_processing(image_inputs, return_tensors="pt").input_ids
+            self.assertEqual(
+                tuple(encoded_images.shape), (self.image_processor_tester.batch_size, *expected_output_image_shape)
+            )
+
+    @unittest.skip(reason="ImageGPT assumes clusters for 3 channels")
+    def test_call_numpy_4_channels(self):
+        pass
+
+    # Override the test from ImageProcessingTestMixin as ImageGPT model takes input_ids as input
+    def test_call_pytorch(self):
+        for image_processing_class in self.image_processor_list:
+            # Initialize image_processing
+            image_processing = image_processing_class(**self.image_processor_dict)
+            # create random PyTorch tensors
+            image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=False, torchify=True)
+            expected_output_image_shape = self.image_processor_tester.expected_output_image_shape(image_inputs)
+
+            for image in image_inputs:
+                self.assertIsInstance(image, torch.Tensor)
+
+            # Test not batched input
+            encoded_images = image_processing(image_inputs[0], return_tensors="pt").input_ids
+            self.assertEqual(tuple(encoded_images.shape), (1, *expected_output_image_shape))
+
+            # Test batched
+            encoded_images = image_processing(image_inputs, return_tensors="pt").input_ids
+            self.assertEqual(
+                tuple(encoded_images.shape),
+                (self.image_processor_tester.batch_size, *expected_output_image_shape),
+            )
+
+    # For quantization-based processors, use absolute tolerance only to avoid infinity issues
+    @require_vision
+    @require_torch
+    def test_slow_fast_equivalence(self):
+        if not self.test_slow_image_processor or not self.test_fast_image_processor:
+            self.skipTest(reason="Skipping slow/fast equivalence test")
+
+        if self.image_processing_class is None or self.fast_image_processing_class is None:
+            self.skipTest(reason="Skipping slow/fast equivalence test as one of the image processors is not defined")
+
+        dummy_image = Image.open(
+            requests.get("http://images.cocodataset.org/val2017/000000039769.jpg", stream=True).raw
+        )
+        image_processor_slow = self.image_processing_class(**self.image_processor_dict)
+        image_processor_fast = self.fast_image_processing_class(**self.image_processor_dict)
+
+        encoding_slow = image_processor_slow(dummy_image, return_tensors="pt")
+        encoding_fast = image_processor_fast(dummy_image, return_tensors="pt")
+        self._assert_slow_fast_tensors_equivalence(
+            encoding_slow.input_ids.float(), encoding_fast.input_ids.float(), atol=1.0, rtol=0
+        )
+
+    @require_vision
+    @require_torch
+    def test_slow_fast_equivalence_batched(self):
+        if not self.test_slow_image_processor or not self.test_fast_image_processor:
+            self.skipTest(reason="Skipping slow/fast equivalence test")
+
+        if self.image_processing_class is None or self.fast_image_processing_class is None:
+            self.skipTest(reason="Skipping slow/fast equivalence test as one of the image processors is not defined")
+
+        if hasattr(self.image_processor_tester, "do_center_crop") and self.image_processor_tester.do_center_crop:
+            self.skipTest(
+                reason="Skipping as do_center_crop is True and center_crop functions are not equivalent for fast and slow processors"
+            )
+
+        dummy_images = self.image_processor_tester.prepare_image_inputs(equal_resolution=False, torchify=True)
+        image_processor_slow = self.image_processing_class(**self.image_processor_dict)
+        image_processor_fast = self.fast_image_processing_class(**self.image_processor_dict)
+
+        encoding_slow = image_processor_slow(dummy_images, return_tensors="pt")
+        encoding_fast = image_processor_fast(dummy_images, return_tensors="pt")
+
+        self._assert_slow_fast_tensors_equivalence(
+            encoding_slow.input_ids.float(), encoding_fast.input_ids.float(), atol=1.0, rtol=0
+        )
+
+    @slow
+    @require_torch_accelerator
+    @require_vision
+    @pytest.mark.torch_compile_test
+    def test_can_compile_fast_image_processor(self):
+        if self.fast_image_processing_class is None:
+            self.skipTest("Skipping compilation test as fast image processor is not defined")
+        if version.parse(torch.__version__) < version.parse("2.3"):
+            self.skipTest(reason="This test requires torch >= 2.3 to run.")
+
+        torch.compiler.reset()
+        input_image = torch.randint(0, 255, (3, 224, 224), dtype=torch.uint8)
+        image_processor = self.fast_image_processing_class(**self.image_processor_dict)
+        output_eager = image_processor(input_image, device=torch_device, return_tensors="pt")
+
+        image_processor = torch.compile(image_processor, mode="reduce-overhead")
+        output_compiled = image_processor(input_image, device=torch_device, return_tensors="pt")
+        self._assert_slow_fast_tensors_equivalence(
+            output_eager.input_ids.float(), output_compiled.input_ids.float(), atol=1.0, rtol=0
+        )
+
+
+def prepare_images():
+    # we use revision="refs/pr/1" until the PR is merged
+    # https://hf.co/datasets/hf-internal-testing/fixtures_image_utils/discussions/1
+    dataset = load_dataset("hf-internal-testing/fixtures_image_utils", split="test", revision="refs/pr/1")
+
+    image1 = dataset[4]["image"]
+    image2 = dataset[5]["image"]
+
+    images = [image1, image2]
+
+    return images
+
+
+@require_vision
+@require_torch
+class ImageGPTImageProcessorIntegrationTest(unittest.TestCase):
+    @slow
+    def test_image(self):
+        image_processing = ImageGPTImageProcessor.from_pretrained("openai/imagegpt-small")
+
+        images = prepare_images()
+
+        # test non-batched
+        encoding = image_processing(images[0], return_tensors="pt")
+
+        self.assertIsInstance(encoding.input_ids, torch.LongTensor)
+        self.assertEqual(encoding.input_ids.shape, (1, 1024))
+
+        expected_slice = [306, 191, 191]
+        self.assertEqual(encoding.input_ids[0, :3].tolist(), expected_slice)
+
+        # test batched
+        encoding = image_processing(images, return_tensors="pt")
+
+        self.assertIsInstance(encoding.input_ids, torch.LongTensor)
+        self.assertEqual(encoding.input_ids.shape, (2, 1024))
+
+        expected_slice = [303, 13, 13]
+        self.assertEqual(encoding.input_ids[1, -3:].tolist(), expected_slice)

transformers/tests/models/imagegpt/test_modeling_imagegpt.py

@@ -0,0 +1,357 @@
+# 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
+from functools import cached_property
+
+from transformers import ImageGPTConfig
+from transformers.testing_utils import require_torch, require_vision, run_test_using_subprocess, slow, torch_device
+from transformers.utils import is_torch_available, is_vision_available
+
+from ...generation.test_utils import GenerationTesterMixin
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_common import ModelTesterMixin, ids_tensor, random_attention_mask
+from ...test_pipeline_mixin import PipelineTesterMixin
+
+
+if is_torch_available():
+    import torch
+
+    from transformers import (
+        ImageGPTForCausalImageModeling,
+        ImageGPTForImageClassification,
+        ImageGPTModel,
+    )
+
+if is_vision_available():
+    from PIL import Image
+
+    from transformers import ImageGPTImageProcessor
+
+
+class ImageGPTModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=14,
+        seq_length=7,
+        is_training=True,
+        use_token_type_ids=True,
+        use_input_mask=True,
+        use_labels=True,
+        use_mc_token_ids=True,
+        vocab_size=99,
+        hidden_size=32,
+        num_hidden_layers=2,
+        num_attention_heads=4,
+        intermediate_size=37,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        max_position_embeddings=512,
+        type_vocab_size=16,
+        type_sequence_label_size=2,
+        initializer_range=0.02,
+        num_labels=3,
+        num_choices=4,
+        scope=None,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.seq_length = seq_length
+        self.is_training = is_training
+        self.use_token_type_ids = use_token_type_ids
+        self.use_input_mask = use_input_mask
+        self.use_labels = use_labels
+        self.use_mc_token_ids = use_mc_token_ids
+        self.vocab_size = vocab_size
+        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.max_position_embeddings = max_position_embeddings
+        self.type_vocab_size = type_vocab_size
+        self.type_sequence_label_size = type_sequence_label_size
+        self.initializer_range = initializer_range
+        self.num_labels = num_labels
+        self.num_choices = num_choices
+        self.scope = None
+
+    def get_large_model_config(self):
+        return ImageGPTConfig.from_pretrained("imagegpt")
+
+    def prepare_config_and_inputs(
+        self, gradient_checkpointing=False, scale_attn_by_inverse_layer_idx=False, reorder_and_upcast_attn=False
+    ):
+        input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size - 1)
+
+        input_mask = None
+        if self.use_input_mask:
+            input_mask = random_attention_mask([self.batch_size, self.seq_length])
+
+        token_type_ids = None
+        if self.use_token_type_ids:
+            token_type_ids = ids_tensor([self.batch_size, self.seq_length], self.type_vocab_size)
+
+        mc_token_ids = None
+        if self.use_mc_token_ids:
+            mc_token_ids = ids_tensor([self.batch_size, self.num_choices], self.seq_length)
+
+        sequence_labels = None
+        token_labels = None
+        choice_labels = None
+        if self.use_labels:
+            sequence_labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
+            token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels)
+            choice_labels = ids_tensor([self.batch_size], self.num_choices)
+
+        config = self.get_config(
+            gradient_checkpointing=gradient_checkpointing,
+            scale_attn_by_inverse_layer_idx=scale_attn_by_inverse_layer_idx,
+            reorder_and_upcast_attn=reorder_and_upcast_attn,
+        )
+
+        head_mask = ids_tensor([self.num_hidden_layers, self.num_attention_heads], 2)
+
+        return (
+            config,
+            input_ids,
+            input_mask,
+            head_mask,
+            token_type_ids,
+            mc_token_ids,
+            sequence_labels,
+            token_labels,
+            choice_labels,
+        )
+
+    def get_config(
+        self, gradient_checkpointing=False, scale_attn_by_inverse_layer_idx=False, reorder_and_upcast_attn=False
+    ):
+        return ImageGPTConfig(
+            vocab_size=self.vocab_size,
+            n_embd=self.hidden_size,
+            n_layer=self.num_hidden_layers,
+            n_head=self.num_attention_heads,
+            n_inner=self.intermediate_size,
+            activation_function=self.hidden_act,
+            resid_pdrop=self.hidden_dropout_prob,
+            attn_pdrop=self.attention_probs_dropout_prob,
+            n_positions=self.max_position_embeddings,
+            type_vocab_size=self.type_vocab_size,
+            initializer_range=self.initializer_range,
+            use_cache=True,
+            gradient_checkpointing=gradient_checkpointing,
+            scale_attn_by_inverse_layer_idx=scale_attn_by_inverse_layer_idx,
+            reorder_and_upcast_attn=reorder_and_upcast_attn,
+        )
+
+    def get_pipeline_config(self):
+        config = self.get_config()
+        config.vocab_size = 513
+        config.max_position_embeddings = 1024
+        return config
+
+    def create_and_check_imagegpt_model(self, config, input_ids, input_mask, head_mask, token_type_ids, *args):
+        model = ImageGPTModel(config=config)
+        model.to(torch_device)
+        model.eval()
+
+        result = model(input_ids, token_type_ids=token_type_ids, head_mask=head_mask)
+        result = model(input_ids, token_type_ids=token_type_ids)
+        result = model(input_ids)
+
+        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
+        self.parent.assertEqual(len(result.past_key_values), config.n_layer)
+
+    def create_and_check_lm_head_model(self, config, input_ids, input_mask, head_mask, token_type_ids, *args):
+        model = ImageGPTForCausalImageModeling(config)
+        model.to(torch_device)
+        model.eval()
+
+        labels = ids_tensor([self.batch_size, self.seq_length], self.vocab_size - 1)
+        result = model(input_ids, token_type_ids=token_type_ids, labels=labels)
+        self.parent.assertEqual(result.loss.shape, ())
+        # ImageGPTForCausalImageModeling doesn't have tied input- and output embeddings
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size - 1))
+
+    def create_and_check_imagegpt_for_image_classification(
+        self, config, input_ids, input_mask, head_mask, token_type_ids, mc_token_ids, sequence_labels, *args
+    ):
+        config.num_labels = self.num_labels
+        model = ImageGPTForImageClassification(config)
+        model.to(torch_device)
+        model.eval()
+        result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=sequence_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,
+            input_ids,
+            input_mask,
+            head_mask,
+            token_type_ids,
+            mc_token_ids,
+            sequence_labels,
+            token_labels,
+            choice_labels,
+        ) = config_and_inputs
+
+        inputs_dict = {
+            "input_ids": input_ids,
+            "token_type_ids": token_type_ids,
+            "head_mask": head_mask,
+        }
+
+        return config, inputs_dict
+
+
+@require_torch
+class ImageGPTModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin, unittest.TestCase):
+    all_model_classes = (
+        (ImageGPTForCausalImageModeling, ImageGPTForImageClassification, ImageGPTModel) if is_torch_available() else ()
+    )
+    pipeline_model_mapping = (
+        {"image-feature-extraction": ImageGPTModel, "image-classification": ImageGPTForImageClassification}
+        if is_torch_available()
+        else {}
+    )
+    test_missing_keys = False
+    test_torch_exportable = True
+
+    # as ImageGPTForImageClassification isn't included in any auto mapping, we add labels here
+    def _prepare_for_class(self, inputs_dict, model_class, return_labels=False):
+        inputs_dict = super()._prepare_for_class(inputs_dict, model_class, return_labels=return_labels)
+
+        if return_labels:
+            if model_class.__name__ == "ImageGPTForImageClassification":
+                inputs_dict["labels"] = torch.zeros(
+                    self.model_tester.batch_size, dtype=torch.long, device=torch_device
+                )
+
+        return inputs_dict
+
+    # we overwrite the _check_scores method of GenerationTesterMixin, as ImageGPTForCausalImageModeling doesn't have tied input- and output embeddings
+    def _check_scores(self, batch_size, scores, generated_length, config):
+        expected_shape = (batch_size, config.vocab_size - 1)
+        self.assertIsInstance(scores, tuple)
+        self.assertEqual(len(scores), generated_length)
+        self.assertListEqual([iter_scores.shape for iter_scores in scores], [expected_shape] * len(scores))
+
+    @run_test_using_subprocess
+    def test_beam_search_generate_dict_outputs_use_cache(self):
+        super().test_beam_search_generate_dict_outputs_use_cache()
+
+    def setUp(self):
+        self.model_tester = ImageGPTModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=ImageGPTConfig, n_embd=37)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_imagegpt_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_imagegpt_model(*config_and_inputs)
+
+    def test_imagegpt_causal_lm(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_lm_head_model(*config_and_inputs)
+
+    def test_imagegpt_image_classification(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_imagegpt_for_image_classification(*config_and_inputs)
+
+    @unittest.skip(
+        reason="This architecture seem to not compute gradients properly when using GC, check: https://github.com/huggingface/transformers/pull/27124"
+    )
+    def test_training_gradient_checkpointing(self):
+        pass
+
+    @unittest.skip(
+        reason="This architecture seem to not compute gradients properly when using GC, check: https://github.com/huggingface/transformers/pull/27124"
+    )
+    def test_training_gradient_checkpointing_use_reentrant(self):
+        pass
+
+    @unittest.skip(
+        reason="This architecture seem to not compute gradients properly when using GC, check: https://github.com/huggingface/transformers/pull/27124"
+    )
+    def test_training_gradient_checkpointing_use_reentrant_false(self):
+        pass
+
+    @slow
+    def test_model_from_pretrained(self):
+        model_name = "openai/imagegpt-small"
+        model = ImageGPTModel.from_pretrained(model_name)
+        self.assertIsNotNone(model)
+
+    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 = ["input_ids"]
+            self.assertListEqual(arg_names[:1], expected_arg_names)
+
+    @unittest.skip(reason="Model inputs don't fit test pattern")  # and it's not used enough to be worth fixing :)
+    def test_past_key_values_format(self):
+        pass
+
+
+# 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 ImageGPTModelIntegrationTest(unittest.TestCase):
+    @cached_property
+    def default_image_processor(self):
+        return ImageGPTImageProcessor.from_pretrained("openai/imagegpt-small") if is_vision_available() else None
+
+    @slow
+    def test_inference_causal_lm_head(self):
+        model = ImageGPTForCausalImageModeling.from_pretrained("openai/imagegpt-small").to(torch_device)
+
+        image_processor = self.default_image_processor
+        image = prepare_img()
+        inputs = image_processor(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, 1024, 512))
+        self.assertEqual(outputs.logits.shape, expected_shape)
+
+        expected_slice = torch.tensor(
+            [[2.3445, 2.6889, 2.7313], [1.0530, 1.2416, 0.5699], [0.2205, 0.7749, 0.3953]]
+        ).to(torch_device)
+
+        torch.testing.assert_close(outputs.logits[0, :3, :3], expected_slice, rtol=1e-4, atol=1e-4)