Efficientformer (#20459)

- Adds EfficientFormer V1 to transformers
- PR co-authored by @novice03  and @Bearnardd 
Co-authored-by: novice <pranavpulijala@gmail.com>
Co-authored-by: novice <44259234+novice03@users.noreply.github.com>

src/transformers/utils/dummy_pt_objects.py

@@ -2149,6 +2149,37 @@ class DPTPreTrainedModel(metaclass=DummyObject):
         requires_backends(self, ["torch"])
 
 
+EFFICIENTFORMER_PRETRAINED_MODEL_ARCHIVE_LIST = None
+
+
+class EfficientFormerForImageClassification(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class EfficientFormerForImageClassificationWithTeacher(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class EfficientFormerModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class EfficientFormerPreTrainedModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
 ELECTRA_PRETRAINED_MODEL_ARCHIVE_LIST = None
 
 

src/transformers/utils/dummy_vision_objects.py

@@ -171,6 +171,13 @@ class DPTImageProcessor(metaclass=DummyObject):
         requires_backends(self, ["vision"])
 
 
+class EfficientFormerImageProcessor(metaclass=DummyObject):
+    _backends = ["vision"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["vision"])
+
+
 class FlavaFeatureExtractor(metaclass=DummyObject):
     _backends = ["vision"]
 

tests/models/efficientformer/test_image_processing_efficientformer.py

+# coding=utf-8
+# 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 unittest
+
+import numpy as np
+
+from transformers.testing_utils import require_torch, require_vision
+from transformers.utils import is_torch_available, is_vision_available
+
+from ...test_feature_extraction_common import FeatureExtractionSavingTestMixin, prepare_image_inputs
+
+
+if is_torch_available():
+    import torch
+
+if is_vision_available():
+    from PIL import Image
+
+    from transformers import ViTFeatureExtractor
+
+
+class EfficientFormerImageProcessorTester(unittest.TestCase):
+    def __init__(
+        self,
+        parent,
+        batch_size=13,
+        num_channels=3,
+        image_size=224,
+        min_resolution=30,
+        max_resolution=400,
+        do_resize=True,
+        size=None,
+        do_normalize=True,
+        image_mean=[0.5, 0.5, 0.5],
+        image_std=[0.5, 0.5, 0.5],
+    ):
+        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
+        self.image_mean = image_mean
+        self.image_std = image_std
+
+    def prepare_feat_extract_dict(self):
+        return {
+            "image_mean": self.image_mean,
+            "image_std": self.image_std,
+            "do_normalize": self.do_normalize,
+            "do_resize": self.do_resize,
+            "size": self.size,
+        }
+
+
+@require_torch
+@require_vision
+class EfficientFormerImageProcessorTest(FeatureExtractionSavingTestMixin, unittest.TestCase):
+
+    feature_extraction_class = ViTFeatureExtractor if is_vision_available() else None
+
+    def setUp(self):
+        self.feature_extract_tester = EfficientFormerImageProcessorTester(self)
+
+    @property
+    def feat_extract_dict(self):
+        return self.feature_extract_tester.prepare_feat_extract_dict()
+
+    def test_feat_extract_properties(self):
+        feature_extractor = self.feature_extraction_class(**self.feat_extract_dict)
+        self.assertTrue(hasattr(feature_extractor, "image_mean"))
+        self.assertTrue(hasattr(feature_extractor, "image_std"))
+        self.assertTrue(hasattr(feature_extractor, "do_normalize"))
+        self.assertTrue(hasattr(feature_extractor, "do_resize"))
+        self.assertTrue(hasattr(feature_extractor, "size"))
+
+    def test_batch_feature(self):
+        pass
+
+    def test_call_pil(self):
+        # Initialize feature_extractor
+        feature_extractor = self.feature_extraction_class(**self.feat_extract_dict)
+        # create random PIL images
+        image_inputs = prepare_image_inputs(self.feature_extract_tester, equal_resolution=False)
+        for image in image_inputs:
+            self.assertIsInstance(image, Image.Image)
+
+        # Test not batched input
+        encoded_images = feature_extractor(image_inputs[0], return_tensors="pt").pixel_values
+        self.assertEqual(
+            encoded_images.shape,
+            (
+                1,
+                self.feature_extract_tester.num_channels,
+                self.feature_extract_tester.size["height"],
+                self.feature_extract_tester.size["width"],
+            ),
+        )
+
+        # Test batched
+        encoded_images = feature_extractor(image_inputs, return_tensors="pt").pixel_values
+        self.assertEqual(
+            encoded_images.shape,
+            (
+                self.feature_extract_tester.batch_size,
+                self.feature_extract_tester.num_channels,
+                self.feature_extract_tester.size["height"],
+                self.feature_extract_tester.size["width"],
+            ),
+        )
+
+    def test_call_numpy(self):
+        # Initialize feature_extractor
+        feature_extractor = self.feature_extraction_class(**self.feat_extract_dict)
+        # create random numpy tensors
+        image_inputs = prepare_image_inputs(self.feature_extract_tester, equal_resolution=False, numpify=True)
+        for image in image_inputs:
+            self.assertIsInstance(image, np.ndarray)
+
+        # Test not batched input
+        encoded_images = feature_extractor(image_inputs[0], return_tensors="pt").pixel_values
+        self.assertEqual(
+            encoded_images.shape,
+            (
+                1,
+                self.feature_extract_tester.num_channels,
+                self.feature_extract_tester.size["height"],
+                self.feature_extract_tester.size["width"],
+            ),
+        )
+
+        # Test batched
+        encoded_images = feature_extractor(image_inputs, return_tensors="pt").pixel_values
+        self.assertEqual(
+            encoded_images.shape,
+            (
+                self.feature_extract_tester.batch_size,
+                self.feature_extract_tester.num_channels,
+                self.feature_extract_tester.size["height"],
+                self.feature_extract_tester.size["width"],
+            ),
+        )
+
+    def test_call_pytorch(self):
+        # Initialize feature_extractor
+        feature_extractor = self.feature_extraction_class(**self.feat_extract_dict)
+        # create random PyTorch tensors
+        image_inputs = prepare_image_inputs(self.feature_extract_tester, equal_resolution=False, torchify=True)
+        for image in image_inputs:
+            self.assertIsInstance(image, torch.Tensor)
+
+        # Test not batched input
+        encoded_images = feature_extractor(image_inputs[0], return_tensors="pt").pixel_values
+        self.assertEqual(
+            encoded_images.shape,
+            (
+                1,
+                self.feature_extract_tester.num_channels,
+                self.feature_extract_tester.size["height"],
+                self.feature_extract_tester.size["width"],
+            ),
+        )
+
+        # Test batched
+        encoded_images = feature_extractor(image_inputs, return_tensors="pt").pixel_values
+        self.assertEqual(
+            encoded_images.shape,
+            (
+                self.feature_extract_tester.batch_size,
+                self.feature_extract_tester.num_channels,
+                self.feature_extract_tester.size["height"],
+                self.feature_extract_tester.size["width"],
+            ),
+        )

tests/models/efficientformer/test_modeling_efficientformer.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 EfficientFormer model. """
+
+
+import inspect
+import unittest
+import warnings
+
+from transformers import EfficientFormerConfig
+from transformers.models.auto import get_values
+from transformers.testing_utils import require_torch, require_vision, slow, torch_device
+from transformers.utils import cached_property, is_torch_available, is_vision_available
+
+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 (
+        MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING,
+        MODEL_MAPPING,
+        EfficientFormerForImageClassification,
+        EfficientFormerForImageClassificationWithTeacher,
+        EfficientFormerModel,
+    )
+    from transformers.models.efficientformer.modeling_efficientformer import (
+        EFFICIENTFORMER_PRETRAINED_MODEL_ARCHIVE_LIST,
+    )
+
+
+if is_vision_available():
+    from PIL import Image
+
+    from transformers import EfficientFormerImageProcessor
+
+
+class EfficientFormerModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size: int = 13,
+        image_size: int = 224,
+        patch_size: int = 2,
+        embed_dim: int = 48,  # last embed dim of stem
+        num_channels: int = 3,
+        is_training: bool = True,
+        use_labels: bool = True,
+        hidden_size: int = 448,
+        num_hidden_layers: int = 7,  # For the l1
+        num_attention_heads: int = 8,
+        intermediate_size: int = 37,
+        hidden_act: str = "gelu",
+        hidden_dropout_prob: float = 0.1,
+        attention_probs_dropout_prob: float = 0.1,
+        type_sequence_label_size: int = 10,
+        initializer_range: float = 0.02,
+        encoder_stride: int = 2,
+        num_attention_outputs: int = 1,  # For l1
+    ):
+        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
+        self.encoder_stride = encoder_stride
+        self.num_attention_outputs = num_attention_outputs
+        self.embed_dim = embed_dim
+        self.seq_length = embed_dim + 1
+
+    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 EfficientFormerConfig(
+            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,
+            encoder_stride=self.encoder_stride,
+        )
+
+    def create_and_check_model(self, config, pixel_values, labels):
+        model = EfficientFormerModel(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(pixel_values)
+        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
+
+    def create_and_check_for_image_classification(self, config, pixel_values, labels):
+        config.num_labels = self.type_sequence_label_size
+        model = EfficientFormerForImageClassification(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))
+
+        # test greyscale images
+        config.num_channels = 1
+        model = EfficientFormerForImageClassification(config)
+        model.to(torch_device)
+        model.eval()
+
+        pixel_values = floats_tensor([self.batch_size, 1, self.image_size, self.image_size])
+        result = model(pixel_values)
+        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 EfficientFormerModelTest(ModelTesterMixin, unittest.TestCase):
+    """
+    Here we also overwrite some of the tests of test_modeling_common.py, as EfficientFormer does not use input_ids, inputs_embeds,
+    attention_mask and seq_length.
+    """
+
+    all_model_classes = (
+        (
+            EfficientFormerModel,
+            EfficientFormerForImageClassificationWithTeacher,
+            EfficientFormerForImageClassification,
+        )
+        if is_torch_available()
+        else ()
+    )
+    fx_compatible = False
+
+    test_pruning = False
+    test_resize_embeddings = False
+    test_head_masking = False
+
+    def setUp(self):
+        self.model_tester = EfficientFormerModelTester(self)
+        self.config_tester = ConfigTester(
+            self, config_class=EfficientFormerConfig, has_text_modality=False, hidden_size=37
+        )
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    @unittest.skip(reason="EfficientFormer does not use inputs_embeds")
+    def test_inputs_embeds(self):
+        pass
+
+    @unittest.skip(reason="EfficientFormer 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_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_layers = getattr(
+                self.model_tester, "expected_num_hidden_layers", self.model_tester.num_hidden_layers + 1
+            )
+            self.assertEqual(len(hidden_states), expected_num_layers)
+
+            if hasattr(self.model_tester, "encoder_seq_length"):
+                seq_length = self.model_tester.encoder_seq_length
+                if hasattr(self.model_tester, "chunk_length") and self.model_tester.chunk_length > 1:
+                    seq_length = seq_length * self.model_tester.chunk_length
+            else:
+                seq_length = self.model_tester.seq_length
+
+            self.assertListEqual(
+                list(hidden_states[-1].shape[-2:]),
+                [seq_length, self.model_tester.hidden_size],
+            )
+
+            if config.is_encoder_decoder:
+                hidden_states = outputs.decoder_hidden_states
+
+                self.assertIsInstance(hidden_states, (list, tuple))
+                self.assertEqual(len(hidden_states), expected_num_layers)
+                seq_len = getattr(self.model_tester, "seq_length", None)
+                decoder_seq_length = getattr(self.model_tester, "decoder_seq_length", seq_len)
+
+                self.assertListEqual(
+                    list(hidden_states[-1].shape[-2:]),
+                    [decoder_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)
+
+    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__ == "EfficientFormerForImageClassificationWithTeacher":
+                del inputs_dict["labels"]
+
+        return inputs_dict
+
+    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="EfficientFormer does not implement masked image modeling yet")
+    def test_for_masked_image_modeling(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_masked_image_modeling(*config_and_inputs)
+
+    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)
+
+    # special case for EfficientFormerForImageClassificationWithTeacher model
+    def test_training(self):
+        if not self.model_tester.is_training:
+            return
+
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        config.return_dict = True
+
+        for model_class in self.all_model_classes:
+            # EfficientFormerForImageClassificationWithTeacher supports inference-only
+            if (
+                model_class in get_values(MODEL_MAPPING)
+                or model_class.__name__ == "EfficientFormerForImageClassificationWithTeacher"
+            ):
+                continue
+            model = model_class(config)
+            model.to(torch_device)
+            model.train()
+            inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
+            loss = model(**inputs).loss
+            loss.backward()
+
+    def test_problem_types(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        problem_types = [
+            {"title": "multi_label_classification", "num_labels": 2, "dtype": torch.float},
+            {"title": "single_label_classification", "num_labels": 1, "dtype": torch.long},
+            {"title": "regression", "num_labels": 1, "dtype": torch.float},
+        ]
+
+        for model_class in self.all_model_classes:
+            if (
+                model_class
+                not in [
+                    *get_values(MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING),
+                ]
+                or model_class.__name__ == "EfficientFormerForImageClassificationWithTeacher"
+            ):
+                continue
+
+            for problem_type in problem_types:
+                with self.subTest(msg=f"Testing {model_class} with {problem_type['title']}"):
+
+                    config.problem_type = problem_type["title"]
+                    config.num_labels = problem_type["num_labels"]
+
+                    model = model_class(config)
+                    model.to(torch_device)
+                    model.train()
+
+                    inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
+
+                    if problem_type["num_labels"] > 1:
+                        inputs["labels"] = inputs["labels"].unsqueeze(1).repeat(1, problem_type["num_labels"])
+
+                    inputs["labels"] = inputs["labels"].to(problem_type["dtype"])
+
+                    # This tests that we do not trigger the warning form PyTorch "Using a target size that is different
+                    # to the input size. This will likely lead to incorrect results due to broadcasting. Please ensure
+                    # they have the same size." which is a symptom something in wrong for the regression problem.
+                    # See https://github.com/huggingface/transformers/issues/11780
+                    with warnings.catch_warnings(record=True) as warning_list:
+                        loss = model(**inputs).loss
+                    for w in warning_list:
+                        if "Using a target size that is different to the input size" in str(w.message):
+                            raise ValueError(
+                                f"Something is going wrong in the regression problem: intercepted {w.message}"
+                            )
+
+                    loss.backward()
+
+    @slow
+    def test_model_from_pretrained(self):
+        for model_name in EFFICIENTFORMER_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+            model = EfficientFormerModel.from_pretrained(model_name)
+            self.assertIsNotNone(model)
+
+    def test_attention_outputs(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        config.return_dict = True
+
+        seq_len = getattr(self.model_tester, "seq_length", None)
+        encoder_seq_length = getattr(self.model_tester, "encoder_seq_length", seq_len)
+        encoder_key_length = getattr(self.model_tester, "key_length", encoder_seq_length)
+        chunk_length = getattr(self.model_tester, "chunk_length", None)
+        if chunk_length is not None and hasattr(self.model_tester, "num_hashes"):
+            encoder_seq_length = encoder_seq_length * self.model_tester.num_hashes
+
+        for model_class in self.all_model_classes:
+            inputs_dict["output_attentions"] = True
+            inputs_dict["output_hidden_states"] = False
+            config.return_dict = True
+            model = model_class(config)
+            model.to(torch_device)
+            model.eval()
+            with torch.no_grad():
+                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
+            attentions = outputs.encoder_attentions if config.is_encoder_decoder else outputs.attentions
+            self.assertEqual(len(attentions), self.model_tester.num_attention_outputs)
+
+            # check that output_attentions also work using config
+            del inputs_dict["output_attentions"]
+            config.output_attentions = True
+            model = model_class(config)
+            model.to(torch_device)
+            model.eval()
+            with torch.no_grad():
+                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
+            attentions = outputs.encoder_attentions if config.is_encoder_decoder else outputs.attentions
+            self.assertEqual(len(attentions), self.model_tester.num_attention_outputs)
+
+            if chunk_length is not None:
+                self.assertListEqual(
+                    list(attentions[0].shape[-4:]),
+                    [self.model_tester.num_attention_heads, encoder_seq_length, chunk_length, encoder_key_length],
+                )
+            else:
+                self.assertListEqual(
+                    list(attentions[0].shape[-3:]),
+                    [self.model_tester.num_attention_heads, encoder_seq_length, encoder_key_length],
+                )
+
+
+# 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 EfficientFormerModelIntegrationTest(unittest.TestCase):
+    @cached_property
+    def default_feature_extractor(self):
+        return (
+            EfficientFormerImageProcessor.from_pretrained("snap-research/efficientformer-l1-300")
+            if is_vision_available()
+            else None
+        )
+
+    @slow
+    def test_inference_image_classification_head(self):
+        model = EfficientFormerForImageClassification.from_pretrained("snap-research/efficientformer-l1-300").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 = (1, 1000)
+        self.assertEqual(outputs.logits.shape, expected_shape)
+
+        expected_slice = torch.tensor([-0.0555, 0.4825, -0.0852]).to(torch_device)
+        self.assertTrue(torch.allclose(outputs.logits[0][:3], expected_slice, atol=1e-4))
+
+    @slow
+    def test_inference_image_classification_head_with_teacher(self):
+        model = EfficientFormerForImageClassificationWithTeacher.from_pretrained(
+            "snap-research/efficientformer-l1-300"
+        ).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 = (1, 1000)
+        self.assertEqual(outputs.logits.shape, expected_shape)
+
+        expected_slice = torch.tensor([-0.1312, 0.4353, -1.0499]).to(torch_device)
+        self.assertTrue(torch.allclose(outputs.logits[0][:3], expected_slice, atol=1e-4))