Resnet flax (#21472)

* [WIP] flax resnet

* added pretrained flax models, results reproducible

* Added pretrained flax models, results reproducible

* working on tests

* no real code change, just some comments

* [flax] adding support for batch norm layers

* fixing bugs related to pt+flax integration

* removing loss from modeling flax output class

* fixing classifier tests

* fixing comments, model output

* cleaning comments

* review changes

* review changes

* Apply suggestions from code review
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* renaming Flax to PyTorch

---------
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

docs/source/de/index.mdx

@@ -285,7 +285,7 @@ Flax), PyTorch, und/oder TensorFlow haben.
 |          Reformer           |       ✅       |       ✅       |       ✅        |         ❌         |      ❌      |
 |           RegNet            |       ❌       |       ❌       |       ✅        |         ✅         |      ❌      |
 |           RemBERT           |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
-|           ResNet            |       ❌       |       ❌       |       ✅        |         ✅         |      ❌      |
+|           ResNet            |       ❌       |       ❌       |       ✅        |         ✅         |      ✅      |
 |          RetriBERT          |       ✅       |       ✅       |       ✅        |         ❌         |      ❌      |
 |           RoBERTa           |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
 |          RoFormer           |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |

docs/source/en/index.mdx

@@ -377,7 +377,7 @@ Flax), PyTorch, and/or TensorFlow.
 |           Reformer            |       ✅       |       ✅       |       ✅        |         ❌         |      ❌      |
 |            RegNet             |       ❌       |       ❌       |       ✅        |         ✅         |      ❌      |
 |            RemBERT            |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
-|            ResNet             |       ❌       |       ❌       |       ✅        |         ✅         |      ❌      |
+|            ResNet             |       ❌       |       ❌       |       ✅        |         ✅         |      ✅      |
 |           RetriBERT           |       ✅       |       ✅       |       ✅        |         ❌         |      ❌      |
 |            RoBERTa            |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
 |     RoBERTa-PreLayerNorm      |       ❌       |       ❌       |       ✅        |         ✅         |      ✅      |

docs/source/en/model_doc/resnet.mdx

@@ -71,3 +71,13 @@ If you're interested in submitting a resource to be included here, please feel f
 
 [[autodoc]] TFResNetForImageClassification
     - call
+
+## FlaxResNetModel
+
+[[autodoc]] FlaxResNetModel
+    - __call__
+
+## FlaxResNetForImageClassification
+
+[[autodoc]] FlaxResNetForImageClassification
+    - __call__

docs/source/es/index.mdx

@@ -237,7 +237,7 @@ Flax), PyTorch y/o TensorFlow.
 |          Reformer           |       ✅       |       ✅       |       ✅        |         ❌         |      ❌      |
 |           RegNet            |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |           RemBERT           |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
-|           ResNet            |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+|           ResNet            |       ❌       |       ❌       |       ✅        |         ❌         |      ✅      |
 |          RetriBERT          |       ✅       |       ✅       |       ✅        |         ❌         |      ❌      |
 |           RoBERTa           |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
 |          RoFormer           |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |

docs/source/it/index.mdx

@@ -254,7 +254,7 @@ tokenizer (chiamato "slow"). Un tokenizer "fast" supportato dalla libreria 🤗
 |          Reformer           |       ✅       |       ✅       |       ✅        |         ❌         |      ❌      |
 |           RegNet            |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |           RemBERT           |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
-|           ResNet            |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+|           ResNet            |       ❌       |       ❌       |       ✅        |         ✅         |      ✅      |
 |          RetriBERT          |       ✅       |       ✅       |       ✅        |         ❌         |      ❌      |
 |           RoBERTa           |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
 |          RoFormer           |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |

docs/source/ja/index.mdx

@@ -339,7 +339,7 @@ specific language governing permissions and limitations under the License.
 |           Reformer            |       ✅       |       ✅       |       ✅        |         ❌         |      ❌      |
 |            RegNet             |       ❌       |       ❌       |       ✅        |         ✅         |      ❌      |
 |            RemBERT            |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
-|            ResNet             |       ❌       |       ❌       |       ✅        |         ✅         |      ❌      |
+|            ResNet             |       ❌       |       ❌       |       ✅        |         ✅         |      ✅      |
 |           RetriBERT           |       ✅       |       ✅       |       ✅        |         ❌         |      ❌      |
 |            RoBERTa            |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
 |     RoBERTa-PreLayerNorm      |       ❌       |       ❌       |       ✅        |         ✅         |      ✅      |

docs/source/ko/index.mdx

@@ -308,7 +308,7 @@ specific language governing permissions and limitations under the License.
 |          Reformer           |       ✅       |       ✅       |       ✅        |         ❌         |      ❌      |
 |           RegNet            |       ❌       |       ❌       |       ✅        |         ✅         |      ❌      |
 |           RemBERT           |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
-|           ResNet            |       ❌       |       ❌       |       ✅        |         ✅         |      ❌      |
+|           ResNet            |       ❌       |       ❌       |       ✅        |         ✅         |      ✅      |
 |          RetriBERT          |       ✅       |       ✅       |       ✅        |         ❌         |      ❌      |
 |           RoBERTa           |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
 |           RoCBert           |       ✅       |       ❌       |       ✅        |         ❌         |      ❌      |

docs/source/pt/index.mdx

@@ -252,7 +252,7 @@ disso, são diferenciados pelo suporte em diferentes frameworks: JAX (por meio d
 |          Reformer           |       ✅       |       ✅       |       ✅        |         ❌         |      ❌      |
 |           RegNet            |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |           RemBERT           |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
-|           ResNet            |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+|           ResNet            |       ❌       |       ❌       |       ✅        |         ❌         |      ✅      |
 |          RetriBERT          |       ✅       |       ✅       |       ✅        |         ❌         |      ❌      |
 |           RoBERTa           |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
 |          RoFormer           |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |

src/transformers/__init__.py

@@ -3637,6 +3637,9 @@ else:
             "FlaxPegasusPreTrainedModel",
         ]
     )
+    _import_structure["models.resnet"].extend(
+        ["FlaxResNetForImageClassification", "FlaxResNetModel", "FlaxResNetPreTrainedModel"]
+    )
     _import_structure["models.roberta"].extend(
         [
             "FlaxRobertaForCausalLM",
@@ -6692,6 +6695,7 @@ if TYPE_CHECKING:
         from .models.mt5 import FlaxMT5EncoderModel, FlaxMT5ForConditionalGeneration, FlaxMT5Model
         from .models.opt import FlaxOPTForCausalLM, FlaxOPTModel, FlaxOPTPreTrainedModel
         from .models.pegasus import FlaxPegasusForConditionalGeneration, FlaxPegasusModel, FlaxPegasusPreTrainedModel
+        from .models.resnet import FlaxResNetForImageClassification, FlaxResNetModel, FlaxResNetPreTrainedModel
         from .models.roberta import (
             FlaxRobertaForCausalLM,
             FlaxRobertaForMaskedLM,

src/transformers/modeling_flax_outputs.py

@@ -45,6 +45,64 @@ class FlaxBaseModelOutput(ModelOutput):
     attentions: Optional[Tuple[jnp.ndarray]] = None
 
 
+@flax.struct.dataclass
+class FlaxBaseModelOutputWithNoAttention(ModelOutput):
+    """
+    Base class for model's outputs, with potential hidden states.
+
+    Args:
+        last_hidden_state (`jnp.ndarray` of shape `(batch_size, num_channels, height, width)`):
+            Sequence of hidden-states at the output of the last layer of the model.
+        hidden_states (`tuple(jnp.ndarray)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `jnp.ndarray` (one for the output of the embeddings, if the model has an embedding layer, + one
+            for the output of each layer) of shape `(batch_size, num_channels, height, width)`. Hidden-states of the
+            model at the output of each layer plus the optional initial embedding outputs.
+    """
+
+    last_hidden_state: jnp.ndarray = None
+    hidden_states: Optional[Tuple[jnp.ndarray]] = None
+
+
+@flax.struct.dataclass
+class FlaxBaseModelOutputWithPoolingAndNoAttention(ModelOutput):
+    """
+    Base class for model's outputs that also contains a pooling of the last hidden states.
+
+    Args:
+        last_hidden_state (`jnp.ndarray` of shape `(batch_size, num_channels, height, width)`):
+            Sequence of hidden-states at the output of the last layer of the model.
+        pooler_output (`jnp.ndarray` of shape `(batch_size, hidden_size)`):
+            Last layer hidden-state after a pooling operation on the spatial dimensions.
+        hidden_states (`tuple(jnp.ndarray)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `jnp.ndarray` (one for the output of the embeddings, if the model has an embedding layer, + one
+            for the output of each layer) of shape `(batch_size, num_channels, height, width)`. Hidden-states of the
+            model at the output of each layer plus the optional initial embedding outputs.
+    """
+
+    last_hidden_state: jnp.ndarray = None
+    pooler_output: jnp.ndarray = None
+    hidden_states: Optional[Tuple[jnp.ndarray]] = None
+
+
+@flax.struct.dataclass
+class FlaxImageClassifierOutputWithNoAttention(ModelOutput):
+    """
+    Base class for outputs of image classification models.
+
+    Args:
+        logits (`jnp.ndarray` of shape `(batch_size, config.num_labels)`):
+            Classification (or regression if config.num_labels==1) scores (before SoftMax).
+        hidden_states (`tuple(jnp.ndarray)`, *optional*, returned when `output_hidden_states=True` is passed or when
+        `config.output_hidden_states=True`):
+            Tuple of `jnp.ndarray` (one for the output of the embeddings, if the model has an embedding layer, + one
+            for the output of each stage) of shape `(batch_size, num_channels, height, width)`. Hidden-states (also
+            called feature maps) of the model at the output of each stage.
+    """
+
+    logits: jnp.ndarray = None
+    hidden_states: Optional[Tuple[jnp.ndarray]] = None
+
+
 @flax.struct.dataclass
 class FlaxBaseModelOutputWithPast(ModelOutput):
     """

src/transformers/models/auto/modeling_flax_auto.py

@@ -48,6 +48,7 @@ FLAX_MODEL_MAPPING_NAMES = OrderedDict(
         ("mt5", "FlaxMT5Model"),
         ("opt", "FlaxOPTModel"),
         ("pegasus", "FlaxPegasusModel"),
+        ("resnet", "FlaxResNetModel"),
         ("roberta", "FlaxRobertaModel"),
         ("roberta-prelayernorm", "FlaxRobertaPreLayerNormModel"),
         ("roformer", "FlaxRoFormerModel"),
@@ -119,6 +120,7 @@ FLAX_MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
     [
         # Model for Image-classsification
         ("beit", "FlaxBeitForImageClassification"),
+        ("resnet", "FlaxResNetForImageClassification"),
         ("vit", "FlaxViTForImageClassification"),
     ]
 )

src/transformers/models/resnet/__init__.py

@@ -13,7 +13,13 @@
 # limitations under the License.
 from typing import TYPE_CHECKING
 
-from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_tf_available, is_torch_available
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_flax_available,
+    is_tf_available,
+    is_torch_available,
+)
 
 
 _import_structure = {
@@ -47,6 +53,17 @@ else:
         "TFResNetPreTrainedModel",
     ]
 
+try:
+    if not is_flax_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_flax_resnet"] = [
+        "FlaxResNetForImageClassification",
+        "FlaxResNetModel",
+        "FlaxResNetPreTrainedModel",
+    ]
 
 if TYPE_CHECKING:
     from .configuration_resnet import RESNET_PRETRAINED_CONFIG_ARCHIVE_MAP, ResNetConfig, ResNetOnnxConfig
@@ -78,6 +95,14 @@ if TYPE_CHECKING:
             TFResNetPreTrainedModel,
         )
 
+    try:
+        if not is_flax_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_flax_resnet import FlaxResNetForImageClassification, FlaxResNetModel, FlaxResNetPreTrainedModel
+
 
 else:
     import sys

src/transformers/utils/dummy_flax_objects.py

@@ -881,6 +881,27 @@ class FlaxPegasusPreTrainedModel(metaclass=DummyObject):
         requires_backends(self, ["flax"])
 
 
+class FlaxResNetForImageClassification(metaclass=DummyObject):
+    _backends = ["flax"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["flax"])
+
+
+class FlaxResNetModel(metaclass=DummyObject):
+    _backends = ["flax"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["flax"])
+
+
+class FlaxResNetPreTrainedModel(metaclass=DummyObject):
+    _backends = ["flax"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["flax"])
+
+
 class FlaxRobertaForCausalLM(metaclass=DummyObject):
     _backends = ["flax"]
 

tests/models/resnet/test_modeling_flax_resnet.py

+# Copyright 2023 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 transformers import ResNetConfig, is_flax_available
+from transformers.testing_utils import require_flax, slow
+from transformers.utils import cached_property, is_vision_available
+
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_flax_common import FlaxModelTesterMixin, floats_tensor
+
+
+if is_flax_available():
+    import jax
+    import jax.numpy as jnp
+
+    from transformers.models.resnet.modeling_flax_resnet import FlaxResNetForImageClassification, FlaxResNetModel
+
+if is_vision_available():
+    from PIL import Image
+
+    from transformers import AutoFeatureExtractor
+
+
+class FlaxResNetModelTester(unittest.TestCase):
+    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])
+
+        config = self.get_config()
+
+        return config, pixel_values
+
+    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):
+        model = FlaxResNetModel(config=config)
+        result = model(pixel_values)
+
+        # Output shape (b, c, h, w)
+        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):
+        config.num_labels = self.num_labels
+        model = FlaxResNetForImageClassification(config=config)
+        result = model(pixel_values)
+        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 = config_and_inputs
+        inputs_dict = {"pixel_values": pixel_values}
+        return config, inputs_dict
+
+
+@require_flax
+class FlaxResNetModelTest(FlaxModelTesterMixin, unittest.TestCase):
+    all_model_classes = (FlaxResNetModel, FlaxResNetForImageClassification) if is_flax_available() else ()
+
+    is_encoder_decoder = False
+    test_head_masking = False
+    has_attentions = False
+
+    def setUp(self) -> None:
+        self.model_tester = FlaxResNetModelTester(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
+
+    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_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)
+
+    @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_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)
+
+    @unittest.skip(reason="ResNet does not use feedforward chunking")
+    def test_feed_forward_chunking(self):
+        pass
+
+    def test_jit_compilation(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            with self.subTest(model_class.__name__):
+                prepared_inputs_dict = self._prepare_for_class(inputs_dict, model_class)
+                model = model_class(config)
+
+                @jax.jit
+                def model_jitted(pixel_values, **kwargs):
+                    return model(pixel_values=pixel_values, **kwargs)
+
+                with self.subTest("JIT Enabled"):
+                    jitted_outputs = model_jitted(**prepared_inputs_dict).to_tuple()
+
+                with self.subTest("JIT Disabled"):
+                    with jax.disable_jit():
+                        outputs = model_jitted(**prepared_inputs_dict).to_tuple()
+
+                self.assertEqual(len(outputs), len(jitted_outputs))
+                for jitted_output, output in zip(jitted_outputs, outputs):
+                    self.assertEqual(jitted_output.shape, output.shape)
+
+
+# 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_flax
+class FlaxResNetModelIntegrationTest(unittest.TestCase):
+    @cached_property
+    def default_feature_extractor(self):
+        return AutoFeatureExtractor.from_pretrained("microsoft/resnet-50") if is_vision_available() else None
+
+    @slow
+    def test_inference_image_classification_head(self):
+        model = FlaxResNetForImageClassification.from_pretrained("microsoft/resnet-50")
+
+        feature_extractor = self.default_feature_extractor
+        image = prepare_img()
+        inputs = feature_extractor(images=image, return_tensors="np")
+
+        outputs = model(**inputs)
+
+        # verify the logits
+        expected_shape = (1, 1000)
+        self.assertEqual(outputs.logits.shape, expected_shape)
+
+        expected_slice = jnp.array([-11.1069, -9.7877, -8.3777])
+
+        self.assertTrue(jnp.allclose(outputs.logits[0, :3], expected_slice, atol=1e-4))