Add CLIPSeg (#20066)

* Add first draft

* Update conversion script

* Improve conversion script

* Improve conversion script some more

* Add conditional embeddings

* Add initial decoder

* Fix activation function of decoder

* Make decoder outputs match original implementation

* Make decoder outputs match original implementation

* Add more copied from statements

* Improve model outputs

* Fix auto tokenizer file

* Fix more tests

* Add test

* Improve README and docs, improve conditional embeddings

* Fix more tests

* Remove print statements

* Remove initial embeddings

* Improve conversion script

* Add interpolation of position embeddings

* Finish addition of interpolation of position embeddings

* Add support for refined checkpoint

* Fix refined checkpoint

* Remove unused parameter

* Improve conversion script

* Add support for training

* Fix conversion script

* Add CLIPSegFeatureExtractor

* Fix processor

* Fix CLIPSegProcessor

* Fix conversion script

* Fix most tests

* Fix equivalence test

* Fix README

* Add model to doc tests

* Use better variable name

* Convert other checkpoint as well

* Update config, add link to paper

* Add docs

* Update organization

* Replace base_model_prefix with clip

* Fix base_model_prefix

* Fix checkpoint of config

* Fix config checkpoint

* Remove file

* Use logits for output

* Fix tests
Co-authored-by: Niels Rogge <nielsrogge@Nielss-MacBook-Pro.local>

src/transformers/models/clipseg/processing_clipseg.py

+# coding=utf-8
+# Copyright 2022 The HuggingFace Inc. team.
+#
+# 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.
+"""
+Image/Text processor class for CLIPSeg
+"""
+from ...processing_utils import ProcessorMixin
+from ...tokenization_utils_base import BatchEncoding
+
+
+class CLIPSegProcessor(ProcessorMixin):
+    r"""
+    Constructs a CLIPSeg processor which wraps a CLIPSeg feature extractor and a CLIP tokenizer into a single
+    processor.
+
+    [`CLIPSegProcessor`] offers all the functionalities of [`ViTFeatureExtractor`] and [`CLIPTokenizerFast`]. See the
+    [`~CLIPSegProcessor.__call__`] and [`~CLIPSegProcessor.decode`] for more information.
+
+    Args:
+        feature_extractor ([`ViTFeatureExtractor`]):
+            The feature extractor is a required input.
+        tokenizer ([`CLIPTokenizerFast`]):
+            The tokenizer is a required input.
+    """
+    feature_extractor_class = "ViTFeatureExtractor"
+    tokenizer_class = ("CLIPTokenizer", "CLIPTokenizerFast")
+
+    def __init__(self, feature_extractor, tokenizer):
+        super().__init__(feature_extractor, tokenizer)
+        self.current_processor = self.feature_extractor
+
+    def __call__(self, text=None, images=None, return_tensors=None, **kwargs):
+        """
+        Main method to prepare for the model one or several sequences(s) and image(s). This method forwards the `text`
+        and `kwargs` arguments to CLIPTokenizerFast's [`~CLIPTokenizerFast.__call__`] if `text` is not `None` to encode
+        the text. To prepare the image(s), this method forwards the `images` and `kwrags` arguments to
+        ViTFeatureExtractor's [`~ViTFeatureExtractor.__call__`] if `images` is not `None`. Please refer to the
+        doctsring of the above two methods for more information.
+
+        Args:
+            text (`str`, `List[str]`, `List[List[str]]`):
+                The sequence or batch of sequences to be encoded. Each sequence can be a string or a list of strings
+                (pretokenized string). If the sequences are provided as list of strings (pretokenized), you must set
+                `is_split_into_words=True` (to lift the ambiguity with a batch of sequences).
+            images (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`, `List[PIL.Image.Image]`, `List[np.ndarray]`, `List[torch.Tensor]`):
+                The image or batch of images to be prepared. Each image can be a PIL image, NumPy array or PyTorch
+                tensor. In case of a NumPy array/PyTorch tensor, each image should be of shape (C, H, W), where C is a
+                number of channels, H and W are image height and width.
+
+            return_tensors (`str` or [`~utils.TensorType`], *optional*):
+                If set, will return tensors of a particular framework. Acceptable values are:
+
+                - `'tf'`: Return TensorFlow `tf.constant` objects.
+                - `'pt'`: Return PyTorch `torch.Tensor` objects.
+                - `'np'`: Return NumPy `np.ndarray` objects.
+                - `'jax'`: Return JAX `jnp.ndarray` objects.
+
+        Returns:
+            [`BatchEncoding`]: A [`BatchEncoding`] with the following fields:
+
+            - **input_ids** -- List of token ids to be fed to a model. Returned when `text` is not `None`.
+            - **attention_mask** -- List of indices specifying which tokens should be attended to by the model (when
+              `return_attention_mask=True` or if *"attention_mask"* is in `self.model_input_names` and if `text` is not
+              `None`).
+            - **pixel_values** -- Pixel values to be fed to a model. Returned when `images` is not `None`.
+        """
+
+        if text is None and images is None:
+            raise ValueError("You have to specify either text or images. Both cannot be none.")
+
+        if text is not None:
+            encoding = self.tokenizer(text, return_tensors=return_tensors, **kwargs)
+
+        if images is not None:
+            image_features = self.feature_extractor(images, return_tensors=return_tensors, **kwargs)
+
+        if text is not None and images is not None:
+            encoding["pixel_values"] = image_features.pixel_values
+            return encoding
+        elif text is not None:
+            return encoding
+        else:
+            return BatchEncoding(data=dict(**image_features), tensor_type=return_tensors)
+
+    def batch_decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to CLIPTokenizerFast's [`~PreTrainedTokenizer.batch_decode`]. Please
+        refer to the docstring of this method for more information.
+        """
+        return self.tokenizer.batch_decode(*args, **kwargs)
+
+    def decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to CLIPTokenizerFast's [`~PreTrainedTokenizer.decode`]. Please refer to
+        the docstring of this method for more information.
+        """
+        return self.tokenizer.decode(*args, **kwargs)

src/transformers/processing_utils.py

@@ -56,7 +56,7 @@ class ProcessorMixin(PushToHubMixin):
         # Sanitize args and kwargs
         for key in kwargs:
             if key not in self.attributes:
-                raise TypeError(f"Unexepcted keyword argument {key}.")
+                raise TypeError(f"Unexpected keyword argument {key}.")
         for arg, attribute_name in zip(args, self.attributes):
             if attribute_name in kwargs:
                 raise TypeError(f"Got multiple values for argument {attribute_name}.")

src/transformers/utils/dummy_pt_objects.py

@@ -1207,6 +1207,44 @@ class CLIPVisionModel(metaclass=DummyObject):
         requires_backends(self, ["torch"])
 
 
+CLIPSEG_PRETRAINED_MODEL_ARCHIVE_LIST = None
+
+
+class CLIPSegForImageSegmentation(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class CLIPSegModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class CLIPSegPreTrainedModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class CLIPSegTextModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class CLIPSegVisionModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
 CODEGEN_PRETRAINED_MODEL_ARCHIVE_LIST = None
 
 

tests/models/clipseg/test_modeling_clipseg.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 CLIPSeg model. """
+
+
+import inspect
+import os
+import tempfile
+import unittest
+
+import numpy as np
+
+import requests
+import transformers
+from transformers import MODEL_MAPPING, CLIPSegConfig, CLIPSegProcessor, CLIPSegTextConfig, CLIPSegVisionConfig
+from transformers.models.auto import get_values
+from transformers.testing_utils import (
+    is_flax_available,
+    is_pt_flax_cross_test,
+    require_torch,
+    require_vision,
+    slow,
+    torch_device,
+)
+from transformers.utils import is_torch_available, is_vision_available
+
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_common import (
+    ModelTesterMixin,
+    _config_zero_init,
+    floats_tensor,
+    ids_tensor,
+    random_attention_mask,
+)
+
+
+if is_torch_available():
+    import torch
+    from torch import nn
+
+    from transformers import CLIPSegForImageSegmentation, CLIPSegModel, CLIPSegTextModel, CLIPSegVisionModel
+    from transformers.models.clipseg.modeling_clipseg import CLIPSEG_PRETRAINED_MODEL_ARCHIVE_LIST
+
+
+if is_vision_available():
+    from PIL import Image
+
+
+if is_flax_available():
+    import jax.numpy as jnp
+    from transformers.modeling_flax_pytorch_utils import (
+        convert_pytorch_state_dict_to_flax,
+        load_flax_weights_in_pytorch_model,
+    )
+
+
+class CLIPSegVisionModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=12,
+        image_size=30,
+        patch_size=2,
+        num_channels=3,
+        is_training=True,
+        hidden_size=32,
+        num_hidden_layers=5,
+        num_attention_heads=4,
+        intermediate_size=37,
+        dropout=0.1,
+        attention_dropout=0.1,
+        initializer_range=0.02,
+        scope=None,
+    ):
+        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.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.intermediate_size = intermediate_size
+        self.dropout = dropout
+        self.attention_dropout = attention_dropout
+        self.initializer_range = initializer_range
+        self.scope = scope
+
+        # in ViT, the seq length equals the number of patches + 1 (we add 1 for the [CLS] token)
+        num_patches = (image_size // patch_size) ** 2
+        self.seq_length = num_patches + 1
+
+    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 CLIPSegVisionConfig(
+            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,
+            dropout=self.dropout,
+            attention_dropout=self.attention_dropout,
+            initializer_range=self.initializer_range,
+        )
+
+    def create_and_check_model(self, config, pixel_values):
+        model = CLIPSegVisionModel(config=config)
+        model.to(torch_device)
+        model.eval()
+        with torch.no_grad():
+            result = model(pixel_values)
+        # expected sequence length = num_patches + 1 (we add 1 for the [CLS] token)
+        image_size = (self.image_size, self.image_size)
+        patch_size = (self.patch_size, self.patch_size)
+        num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
+        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, num_patches + 1, self.hidden_size))
+        self.parent.assertEqual(result.pooler_output.shape, (self.batch_size, self.hidden_size))
+
+    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_torch
+class CLIPSegVisionModelTest(ModelTesterMixin, unittest.TestCase):
+    """
+    Here we also overwrite some of the tests of test_modeling_common.py, as CLIPSeg does not use input_ids, inputs_embeds,
+    attention_mask and seq_length.
+    """
+
+    all_model_classes = (CLIPSegVisionModel,) 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 = CLIPSegVisionModelTester(self)
+        self.config_tester = ConfigTester(
+            self, config_class=CLIPSegVisionConfig, has_text_modality=False, hidden_size=37
+        )
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    @unittest.skip(reason="CLIPSeg does not use inputs_embeds")
+    def test_inputs_embeds(self):
+        pass
+
+    def test_model_common_attributes(self):
+        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            model = model_class(config)
+            self.assertIsInstance(model.get_input_embeddings(), (nn.Module))
+            x = model.get_output_embeddings()
+            self.assertTrue(x is None or isinstance(x, nn.Linear))
+
+    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_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_model(*config_and_inputs)
+
+    def test_training(self):
+        pass
+
+    def test_training_gradient_checkpointing(self):
+        pass
+
+    @unittest.skip(reason="CLIPSegVisionModel has no base class and is not available in MODEL_MAPPING")
+    def test_save_load_fast_init_from_base(self):
+        pass
+
+    @unittest.skip(reason="CLIPSegVisionModel has no base class and is not available in MODEL_MAPPING")
+    def test_save_load_fast_init_to_base(self):
+        pass
+
+    @slow
+    def test_model_from_pretrained(self):
+        for model_name in CLIPSEG_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+            model = CLIPSegVisionModel.from_pretrained(model_name)
+            self.assertIsNotNone(model)
+
+
+class CLIPSegTextModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=12,
+        seq_length=7,
+        is_training=True,
+        use_input_mask=True,
+        use_labels=True,
+        vocab_size=99,
+        hidden_size=32,
+        num_hidden_layers=5,
+        num_attention_heads=4,
+        intermediate_size=37,
+        dropout=0.1,
+        attention_dropout=0.1,
+        max_position_embeddings=512,
+        initializer_range=0.02,
+        scope=None,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.seq_length = seq_length
+        self.is_training = is_training
+        self.use_input_mask = use_input_mask
+        self.use_labels = use_labels
+        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.dropout = dropout
+        self.attention_dropout = attention_dropout
+        self.max_position_embeddings = max_position_embeddings
+        self.initializer_range = initializer_range
+        self.scope = scope
+
+    def prepare_config_and_inputs(self):
+        input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
+
+        input_mask = None
+        if self.use_input_mask:
+            input_mask = random_attention_mask([self.batch_size, self.seq_length])
+
+        if input_mask is not None:
+            batch_size, seq_length = input_mask.shape
+            rnd_start_indices = np.random.randint(1, seq_length - 1, size=(batch_size,))
+            for batch_idx, start_index in enumerate(rnd_start_indices):
+                input_mask[batch_idx, :start_index] = 1
+                input_mask[batch_idx, start_index:] = 0
+
+        config = self.get_config()
+
+        return config, input_ids, input_mask
+
+    def get_config(self):
+        return CLIPSegTextConfig(
+            vocab_size=self.vocab_size,
+            hidden_size=self.hidden_size,
+            num_hidden_layers=self.num_hidden_layers,
+            num_attention_heads=self.num_attention_heads,
+            intermediate_size=self.intermediate_size,
+            dropout=self.dropout,
+            attention_dropout=self.attention_dropout,
+            max_position_embeddings=self.max_position_embeddings,
+            initializer_range=self.initializer_range,
+        )
+
+    def create_and_check_model(self, config, input_ids, input_mask):
+        model = CLIPSegTextModel(config=config)
+        model.to(torch_device)
+        model.eval()
+        with torch.no_grad():
+            result = model(input_ids, attention_mask=input_mask)
+            result = model(input_ids)
+        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
+        self.parent.assertEqual(result.pooler_output.shape, (self.batch_size, self.hidden_size))
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        config, input_ids, input_mask = config_and_inputs
+        inputs_dict = {"input_ids": input_ids, "attention_mask": input_mask}
+        return config, inputs_dict
+
+
+@require_torch
+class CLIPSegTextModelTest(ModelTesterMixin, unittest.TestCase):
+
+    all_model_classes = (CLIPSegTextModel,) if is_torch_available() else ()
+    fx_compatible = False
+    test_pruning = False
+    test_head_masking = False
+
+    def setUp(self):
+        self.model_tester = CLIPSegTextModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=CLIPSegTextConfig, hidden_size=37)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    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_training(self):
+        pass
+
+    def test_training_gradient_checkpointing(self):
+        pass
+
+    @unittest.skip(reason="CLIPSeg does not use inputs_embeds")
+    def test_inputs_embeds(self):
+        pass
+
+    @unittest.skip(reason="CLIPSegTextModel has no base class and is not available in MODEL_MAPPING")
+    def test_save_load_fast_init_from_base(self):
+        pass
+
+    @unittest.skip(reason="CLIPSegTextModel has no base class and is not available in MODEL_MAPPING")
+    def test_save_load_fast_init_to_base(self):
+        pass
+
+    @slow
+    def test_model_from_pretrained(self):
+        for model_name in CLIPSEG_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+            model = CLIPSegTextModel.from_pretrained(model_name)
+            self.assertIsNotNone(model)
+
+
+class CLIPSegModelTester:
+    def __init__(self, parent, is_training=True):
+        self.parent = parent
+        self.text_model_tester = CLIPSegTextModelTester(parent)
+        self.vision_model_tester = CLIPSegVisionModelTester(parent)
+        self.is_training = is_training
+
+    def prepare_config_and_inputs(self):
+        text_config, input_ids, attention_mask = self.text_model_tester.prepare_config_and_inputs()
+        vision_config, pixel_values = self.vision_model_tester.prepare_config_and_inputs()
+
+        config = self.get_config()
+
+        return config, input_ids, attention_mask, pixel_values
+
+    def get_config(self):
+        return CLIPSegConfig.from_text_vision_configs(
+            self.text_model_tester.get_config(),
+            self.vision_model_tester.get_config(),
+            projection_dim=64,
+            reduce_dim=32,
+            extract_layers=[1, 2, 3],
+        )
+
+    def create_and_check_model(self, config, input_ids, attention_mask, pixel_values):
+        model = CLIPSegModel(config).to(torch_device).eval()
+        with torch.no_grad():
+            result = model(input_ids, pixel_values, attention_mask)
+        self.parent.assertEqual(
+            result.logits_per_image.shape, (self.vision_model_tester.batch_size, self.text_model_tester.batch_size)
+        )
+        self.parent.assertEqual(
+            result.logits_per_text.shape, (self.text_model_tester.batch_size, self.vision_model_tester.batch_size)
+        )
+
+    def create_and_check_model_for_image_segmentation(self, config, input_ids, attention_maks, pixel_values):
+        model = CLIPSegForImageSegmentation(config).to(torch_device).eval()
+        with torch.no_grad():
+            result = model(input_ids, pixel_values)
+        self.parent.assertEqual(
+            result.logits.shape,
+            (
+                self.vision_model_tester.batch_size,
+                self.vision_model_tester.image_size,
+                self.vision_model_tester.image_size,
+            ),
+        )
+        self.parent.assertEqual(
+            result.conditional_embeddings.shape, (self.text_model_tester.batch_size, config.projection_dim)
+        )
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        config, input_ids, attention_mask, pixel_values = config_and_inputs
+        inputs_dict = {
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+            "pixel_values": pixel_values,
+        }
+        return config, inputs_dict
+
+
+@require_torch
+class CLIPSegModelTest(ModelTesterMixin, unittest.TestCase):
+    all_model_classes = (CLIPSegModel, CLIPSegForImageSegmentation) if is_torch_available() else ()
+    fx_compatible = False
+    test_head_masking = False
+    test_pruning = False
+    test_resize_embeddings = False
+    test_attention_outputs = False
+
+    def _prepare_for_class(self, inputs_dict, model_class, return_labels=False):
+        # CLIPSegForImageSegmentation requires special treatment
+        if return_labels:
+            if model_class.__name__ == "CLIPSegForImageSegmentation":
+                batch_size, _, height, width = inputs_dict["pixel_values"].shape
+                inputs_dict["labels"] = torch.zeros(
+                    [batch_size, height, width], device=torch_device, dtype=torch.float
+                )
+
+        return inputs_dict
+
+    def setUp(self):
+        self.model_tester = CLIPSegModelTester(self)
+
+    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_model_for_image_segmentation(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_model_for_image_segmentation(*config_and_inputs)
+
+    @unittest.skip(reason="Hidden_states is tested in individual model tests")
+    def test_hidden_states_output(self):
+        pass
+
+    @unittest.skip(reason="Inputs_embeds is tested in individual model tests")
+    def test_inputs_embeds(self):
+        pass
+
+    @unittest.skip(reason="Retain_grad is tested in individual model tests")
+    def test_retain_grad_hidden_states_attentions(self):
+        pass
+
+    @unittest.skip(reason="CLIPSegModel does not have input/output embeddings")
+    def test_model_common_attributes(self):
+        pass
+
+    # override as the some parameters require custom initialization
+    def test_initialization(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        configs_no_init = _config_zero_init(config)
+        for model_class in self.all_model_classes:
+            model = model_class(config=configs_no_init)
+            for name, param in model.named_parameters():
+                if param.requires_grad:
+                    # check if `logit_scale` is initilized as per the original implementation
+                    if "logit_scale" in name:
+                        self.assertAlmostEqual(
+                            param.data.item(),
+                            np.log(1 / 0.07),
+                            delta=1e-3,
+                            msg=f"Parameter {name} of model {model_class} seems not properly initialized",
+                        )
+                    elif "film" in name or "transposed_conv" in name or "reduce" in name:
+                        # those parameters use PyTorch' default nn.Linear initialization scheme
+                        pass
+                    else:
+                        self.assertIn(
+                            ((param.data.mean() * 1e9).round() / 1e9).item(),
+                            [0.0, 1.0],
+                            msg=f"Parameter {name} of model {model_class} seems not properly initialized",
+                        )
+
+    def _create_and_check_torchscript(self, config, inputs_dict):
+        if not self.test_torchscript:
+            return
+
+        configs_no_init = _config_zero_init(config)  # To be sure we have no Nan
+        configs_no_init.torchscript = True
+        configs_no_init.return_dict = False
+        for model_class in self.all_model_classes:
+            model = model_class(config=configs_no_init)
+            model.to(torch_device)
+            model.eval()
+
+            try:
+                input_ids = inputs_dict["input_ids"]
+                pixel_values = inputs_dict["pixel_values"]  # CLIPSeg needs pixel_values
+                traced_model = torch.jit.trace(model, (input_ids, pixel_values))
+            except RuntimeError:
+                self.fail("Couldn't trace module.")
+
+            with tempfile.TemporaryDirectory() as tmp_dir_name:
+                pt_file_name = os.path.join(tmp_dir_name, "traced_model.pt")
+
+                try:
+                    torch.jit.save(traced_model, pt_file_name)
+                except Exception:
+                    self.fail("Couldn't save module.")
+
+                try:
+                    loaded_model = torch.jit.load(pt_file_name)
+                except Exception:
+                    self.fail("Couldn't load module.")
+
+            model.to(torch_device)
+            model.eval()
+
+            loaded_model.to(torch_device)
+            loaded_model.eval()
+
+            model_state_dict = model.state_dict()
+            loaded_model_state_dict = loaded_model.state_dict()
+
+            self.assertEqual(set(model_state_dict.keys()), set(loaded_model_state_dict.keys()))
+
+            models_equal = True
+            for layer_name, p1 in model_state_dict.items():
+                p2 = loaded_model_state_dict[layer_name]
+                if p1.data.ne(p2.data).sum() > 0:
+                    models_equal = False
+
+            self.assertTrue(models_equal)
+
+    def test_load_vision_text_config(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        # Save CLIPSegConfig and check if we can load CLIPSegVisionConfig from it
+        with tempfile.TemporaryDirectory() as tmp_dir_name:
+            config.save_pretrained(tmp_dir_name)
+            vision_config = CLIPSegVisionConfig.from_pretrained(tmp_dir_name)
+            self.assertDictEqual(config.vision_config.to_dict(), vision_config.to_dict())
+
+        # Save CLIPSegConfig and check if we can load CLIPSegTextConfig from it
+        with tempfile.TemporaryDirectory() as tmp_dir_name:
+            config.save_pretrained(tmp_dir_name)
+            text_config = CLIPSegTextConfig.from_pretrained(tmp_dir_name)
+            self.assertDictEqual(config.text_config.to_dict(), text_config.to_dict())
+
+    # overwrite from common since FlaxCLIPSegModel returns nested output
+    # which is not supported in the common test
+    @is_pt_flax_cross_test
+    def test_equivalence_pt_to_flax(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__):
+
+                # load PyTorch class
+                pt_model = model_class(config).eval()
+                # Flax models don't use the `use_cache` option and cache is not returned as a default.
+                # So we disable `use_cache` here for PyTorch model.
+                pt_model.config.use_cache = False
+
+                fx_model_class_name = "Flax" + model_class.__name__
+
+                if not hasattr(transformers, fx_model_class_name):
+                    return
+
+                fx_model_class = getattr(transformers, fx_model_class_name)
+
+                # load Flax class
+                fx_model = fx_model_class(config, dtype=jnp.float32)
+                # make sure only flax inputs are forward that actually exist in function args
+                fx_input_keys = inspect.signature(fx_model.__call__).parameters.keys()
+
+                # prepare inputs
+                pt_inputs = self._prepare_for_class(inputs_dict, model_class)
+
+                # remove function args that don't exist in Flax
+                pt_inputs = {k: v for k, v in pt_inputs.items() if k in fx_input_keys}
+
+                fx_state = convert_pytorch_state_dict_to_flax(pt_model.state_dict(), fx_model)
+                fx_model.params = fx_state
+
+                with torch.no_grad():
+                    pt_outputs = pt_model(**pt_inputs).to_tuple()
+
+                # convert inputs to Flax
+                fx_inputs = {k: np.array(v) for k, v in pt_inputs.items() if torch.is_tensor(v)}
+                fx_outputs = fx_model(**fx_inputs).to_tuple()
+                self.assertEqual(len(fx_outputs), len(pt_outputs), "Output lengths differ between Flax and PyTorch")
+                for fx_output, pt_output in zip(fx_outputs[:4], pt_outputs[:4]):
+                    self.assert_almost_equals(fx_output, pt_output.numpy(), 4e-2)
+
+                with tempfile.TemporaryDirectory() as tmpdirname:
+                    pt_model.save_pretrained(tmpdirname)
+                    fx_model_loaded = fx_model_class.from_pretrained(tmpdirname, from_pt=True)
+
+                fx_outputs_loaded = fx_model_loaded(**fx_inputs).to_tuple()
+                self.assertEqual(
+                    len(fx_outputs_loaded), len(pt_outputs), "Output lengths differ between Flax and PyTorch"
+                )
+                for fx_output_loaded, pt_output in zip(fx_outputs_loaded[:4], pt_outputs[:4]):
+                    self.assert_almost_equals(fx_output_loaded, pt_output.numpy(), 4e-2)
+
+    # overwrite from common since FlaxCLIPSegModel returns nested output
+    # which is not supported in the common test
+    @is_pt_flax_cross_test
+    def test_equivalence_flax_to_pt(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__):
+                # load corresponding PyTorch class
+                pt_model = model_class(config).eval()
+
+                # So we disable `use_cache` here for PyTorch model.
+                pt_model.config.use_cache = False
+
+                fx_model_class_name = "Flax" + model_class.__name__
+
+                if not hasattr(transformers, fx_model_class_name):
+                    # no flax model exists for this class
+                    return
+
+                fx_model_class = getattr(transformers, fx_model_class_name)
+
+                # load Flax class
+                fx_model = fx_model_class(config, dtype=jnp.float32)
+                # make sure only flax inputs are forward that actually exist in function args
+                fx_input_keys = inspect.signature(fx_model.__call__).parameters.keys()
+
+                pt_model = load_flax_weights_in_pytorch_model(pt_model, fx_model.params)
+
+                # make sure weights are tied in PyTorch
+                pt_model.tie_weights()
+
+                # prepare inputs
+                pt_inputs = self._prepare_for_class(inputs_dict, model_class)
+
+                # remove function args that don't exist in Flax
+                pt_inputs = {k: v for k, v in pt_inputs.items() if k in fx_input_keys}
+
+                with torch.no_grad():
+                    pt_outputs = pt_model(**pt_inputs).to_tuple()
+
+                fx_inputs = {k: np.array(v) for k, v in pt_inputs.items() if torch.is_tensor(v)}
+
+                fx_outputs = fx_model(**fx_inputs).to_tuple()
+                self.assertEqual(len(fx_outputs), len(pt_outputs), "Output lengths differ between Flax and PyTorch")
+
+                for fx_output, pt_output in zip(fx_outputs[:4], pt_outputs[:4]):
+                    self.assert_almost_equals(fx_output, pt_output.numpy(), 4e-2)
+
+                with tempfile.TemporaryDirectory() as tmpdirname:
+                    fx_model.save_pretrained(tmpdirname)
+                    pt_model_loaded = model_class.from_pretrained(tmpdirname, from_flax=True)
+
+                with torch.no_grad():
+                    pt_outputs_loaded = pt_model_loaded(**pt_inputs).to_tuple()
+
+                self.assertEqual(
+                    len(fx_outputs), len(pt_outputs_loaded), "Output lengths differ between Flax and PyTorch"
+                )
+                for fx_output, pt_output in zip(fx_outputs[:4], pt_outputs_loaded[:4]):
+                    self.assert_almost_equals(fx_output, pt_output.numpy(), 4e-2)
+
+    def test_training(self):
+        if not self.model_tester.is_training:
+            return
+
+        for model_class in self.all_model_classes:
+            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+            config.return_dict = True
+
+            if model_class in get_values(MODEL_MAPPING):
+                continue
+
+            print("Model class:", model_class)
+
+            model = model_class(config)
+            model.to(torch_device)
+            model.train()
+            inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
+            for k, v in inputs.items():
+                print(k, v.shape)
+            loss = model(**inputs).loss
+            loss.backward()
+
+    @slow
+    def test_model_from_pretrained(self):
+        for model_name in CLIPSEG_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+            model = CLIPSegModel.from_pretrained(model_name)
+            self.assertIsNotNone(model)
+
+
+# We will verify our results on an image of cute cats
+def prepare_img():
+    url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+    image = Image.open(requests.get(url, stream=True).raw)
+    return image
+
+
+@require_vision
+@require_torch
+class CLIPSegModelIntegrationTest(unittest.TestCase):
+    @slow
+    def test_inference_image_segmentation(self):
+        model_name = "CIDAS/clipseg-rd64-refined"
+        processor = CLIPSegProcessor.from_pretrained(model_name)
+        model = CLIPSegForImageSegmentation.from_pretrained(model_name).to(torch_device)
+
+        image = prepare_img()
+        texts = ["a cat", "a remote", "a blanket"]
+        inputs = processor(text=texts, images=[image] * len(texts), padding=True, return_tensors="pt").to(torch_device)
+
+        # forward pass
+        with torch.no_grad():
+            outputs = model(**inputs)
+
+        # verify the predicted masks
+        self.assertEqual(
+            outputs.logits.shape,
+            torch.Size((3, 352, 352)),
+        )
+        expected_masks_slice = torch.tensor(
+            [[-7.4577, -7.4952, -7.4072], [-7.3115, -7.0969, -7.1624], [-6.9472, -6.7641, -6.8911]]
+        )
+        self.assertTrue(torch.allclose(outputs.logits[0, :3, :3], expected_masks_slice, atol=1e-3))
+
+        # verify conditional and pooled output
+        expected_conditional = torch.tensor([0.5601, -0.0314, 0.1980])
+        expected_pooled_output = torch.tensor([0.2692, -0.7197, -0.1328])
+        self.assertTrue(torch.allclose(outputs.conditional_embeddings[0, :3], expected_conditional, atol=1e-3))
+        self.assertTrue(torch.allclose(outputs.pooled_output[0, :3], expected_pooled_output, atol=1e-3))

tests/models/clipseg/test_processor_clipseg.py

+# Copyright 2022 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 json
+import os
+import shutil
+import tempfile
+import unittest
+
+import numpy as np
+import pytest
+
+from transformers import CLIPTokenizer, CLIPTokenizerFast
+from transformers.models.clip.tokenization_clip import VOCAB_FILES_NAMES
+from transformers.testing_utils import require_vision
+from transformers.utils import FEATURE_EXTRACTOR_NAME, is_vision_available
+
+
+if is_vision_available():
+    from PIL import Image
+
+    from transformers import CLIPSegProcessor, ViTFeatureExtractor
+
+
+@require_vision
+class CLIPSegProcessorTest(unittest.TestCase):
+    def setUp(self):
+        self.tmpdirname = tempfile.mkdtemp()
+
+        # fmt: off
+        vocab = ["l", "o", "w", "e", "r", "s", "t", "i", "d", "n", "lo", "l</w>", "w</w>", "r</w>", "t</w>", "low</w>", "er</w>", "lowest</w>", "newer</w>", "wider", "<unk>", "<|startoftext|>", "<|endoftext|>"]
+        # fmt: on
+        vocab_tokens = dict(zip(vocab, range(len(vocab))))
+        merges = ["#version: 0.2", "l o", "lo w</w>", "e r</w>", ""]
+        self.special_tokens_map = {"unk_token": "<unk>"}
+
+        self.vocab_file = os.path.join(self.tmpdirname, VOCAB_FILES_NAMES["vocab_file"])
+        self.merges_file = os.path.join(self.tmpdirname, VOCAB_FILES_NAMES["merges_file"])
+        with open(self.vocab_file, "w", encoding="utf-8") as fp:
+            fp.write(json.dumps(vocab_tokens) + "\n")
+        with open(self.merges_file, "w", encoding="utf-8") as fp:
+            fp.write("\n".join(merges))
+
+        feature_extractor_map = {
+            "do_resize": True,
+            "size": 20,
+            "do_center_crop": True,
+            "crop_size": 18,
+            "do_normalize": True,
+            "image_mean": [0.48145466, 0.4578275, 0.40821073],
+            "image_std": [0.26862954, 0.26130258, 0.27577711],
+        }
+        self.feature_extractor_file = os.path.join(self.tmpdirname, FEATURE_EXTRACTOR_NAME)
+        with open(self.feature_extractor_file, "w", encoding="utf-8") as fp:
+            json.dump(feature_extractor_map, fp)
+
+    def get_tokenizer(self, **kwargs):
+        return CLIPTokenizer.from_pretrained(self.tmpdirname, **kwargs)
+
+    def get_rust_tokenizer(self, **kwargs):
+        return CLIPTokenizerFast.from_pretrained(self.tmpdirname, **kwargs)
+
+    def get_feature_extractor(self, **kwargs):
+        return ViTFeatureExtractor.from_pretrained(self.tmpdirname, **kwargs)
+
+    def tearDown(self):
+        shutil.rmtree(self.tmpdirname)
+
+    def prepare_image_inputs(self):
+        """This function prepares a list of PIL images, or a list of numpy arrays if one specifies numpify=True,
+        or a list of PyTorch tensors if one specifies torchify=True."""
+
+        image_inputs = [np.random.randint(255, size=(3, 30, 400), dtype=np.uint8)]
+
+        image_inputs = [Image.fromarray(np.moveaxis(x, 0, -1)) for x in image_inputs]
+
+        return image_inputs
+
+    def test_save_load_pretrained_default(self):
+        tokenizer_slow = self.get_tokenizer()
+        tokenizer_fast = self.get_rust_tokenizer()
+        feature_extractor = self.get_feature_extractor()
+
+        processor_slow = CLIPSegProcessor(tokenizer=tokenizer_slow, feature_extractor=feature_extractor)
+        processor_slow.save_pretrained(self.tmpdirname)
+        processor_slow = CLIPSegProcessor.from_pretrained(self.tmpdirname, use_fast=False)
+
+        processor_fast = CLIPSegProcessor(tokenizer=tokenizer_fast, feature_extractor=feature_extractor)
+        processor_fast.save_pretrained(self.tmpdirname)
+        processor_fast = CLIPSegProcessor.from_pretrained(self.tmpdirname)
+
+        self.assertEqual(processor_slow.tokenizer.get_vocab(), tokenizer_slow.get_vocab())
+        self.assertEqual(processor_fast.tokenizer.get_vocab(), tokenizer_fast.get_vocab())
+        self.assertEqual(tokenizer_slow.get_vocab(), tokenizer_fast.get_vocab())
+        self.assertIsInstance(processor_slow.tokenizer, CLIPTokenizer)
+        self.assertIsInstance(processor_fast.tokenizer, CLIPTokenizerFast)
+
+        self.assertEqual(processor_slow.feature_extractor.to_json_string(), feature_extractor.to_json_string())
+        self.assertEqual(processor_fast.feature_extractor.to_json_string(), feature_extractor.to_json_string())
+        self.assertIsInstance(processor_slow.feature_extractor, ViTFeatureExtractor)
+        self.assertIsInstance(processor_fast.feature_extractor, ViTFeatureExtractor)
+
+    def test_save_load_pretrained_additional_features(self):
+        processor = CLIPSegProcessor(tokenizer=self.get_tokenizer(), feature_extractor=self.get_feature_extractor())
+        processor.save_pretrained(self.tmpdirname)
+
+        tokenizer_add_kwargs = self.get_tokenizer(bos_token="(BOS)", eos_token="(EOS)")
+        feature_extractor_add_kwargs = self.get_feature_extractor(do_normalize=False, padding_value=1.0)
+
+        processor = CLIPSegProcessor.from_pretrained(
+            self.tmpdirname, bos_token="(BOS)", eos_token="(EOS)", do_normalize=False, padding_value=1.0
+        )
+
+        self.assertEqual(processor.tokenizer.get_vocab(), tokenizer_add_kwargs.get_vocab())
+        self.assertIsInstance(processor.tokenizer, CLIPTokenizerFast)
+
+        self.assertEqual(processor.feature_extractor.to_json_string(), feature_extractor_add_kwargs.to_json_string())
+        self.assertIsInstance(processor.feature_extractor, ViTFeatureExtractor)
+
+    def test_feature_extractor(self):
+        feature_extractor = self.get_feature_extractor()
+        tokenizer = self.get_tokenizer()
+
+        processor = CLIPSegProcessor(tokenizer=tokenizer, feature_extractor=feature_extractor)
+
+        image_input = self.prepare_image_inputs()
+
+        input_feat_extract = feature_extractor(image_input, return_tensors="np")
+        input_processor = processor(images=image_input, return_tensors="np")
+
+        for key in input_feat_extract.keys():
+            self.assertAlmostEqual(input_feat_extract[key].sum(), input_processor[key].sum(), delta=1e-2)
+
+    def test_tokenizer(self):
+        feature_extractor = self.get_feature_extractor()
+        tokenizer = self.get_tokenizer()
+
+        processor = CLIPSegProcessor(tokenizer=tokenizer, feature_extractor=feature_extractor)
+
+        input_str = "lower newer"
+
+        encoded_processor = processor(text=input_str)
+
+        encoded_tok = tokenizer(input_str)
+
+        for key in encoded_tok.keys():
+            self.assertListEqual(encoded_tok[key], encoded_processor[key])
+
+    def test_processor(self):
+        feature_extractor = self.get_feature_extractor()
+        tokenizer = self.get_tokenizer()
+
+        processor = CLIPSegProcessor(tokenizer=tokenizer, feature_extractor=feature_extractor)
+
+        input_str = "lower newer"
+        image_input = self.prepare_image_inputs()
+
+        inputs = processor(text=input_str, images=image_input)
+
+        self.assertListEqual(list(inputs.keys()), ["input_ids", "attention_mask", "pixel_values"])
+
+        # test if it raises when no input is passed
+        with pytest.raises(ValueError):
+            processor()
+
+    def test_tokenizer_decode(self):
+        feature_extractor = self.get_feature_extractor()
+        tokenizer = self.get_tokenizer()
+
+        processor = CLIPSegProcessor(tokenizer=tokenizer, feature_extractor=feature_extractor)
+
+        predicted_ids = [[1, 4, 5, 8, 1, 0, 8], [3, 4, 3, 1, 1, 8, 9]]
+
+        decoded_processor = processor.batch_decode(predicted_ids)
+        decoded_tok = tokenizer.batch_decode(predicted_ids)
+
+        self.assertListEqual(decoded_tok, decoded_processor)

utils/check_repo.py

@@ -46,6 +46,7 @@ PRIVATE_MODELS = [
 # Being in this list is an exception and should **not** be the rule.
 IGNORE_NON_TESTED = PRIVATE_MODELS.copy() + [
     # models to ignore for not tested
+    "CLIPSegDecoder",  # Building part of bigger (tested) model.
     "TableTransformerEncoder",  # Building part of bigger (tested) model.
     "TableTransformerDecoder",  # Building part of bigger (tested) model.
     "TimeSeriesTransformerEncoder",  # Building part of bigger (tested) model.
@@ -140,6 +141,9 @@ TEST_FILES_WITH_NO_COMMON_TESTS = [
 # should **not** be the rule.
 IGNORE_NON_AUTO_CONFIGURED = PRIVATE_MODELS.copy() + [
     # models to ignore for model xxx mapping
+    "CLIPSegForImageSegmentation",
+    "CLIPSegVisionModel",
+    "CLIPSegTextModel",
     "EsmForProteinFolding",
     "TimeSeriesTransformerForPrediction",
     "PegasusXEncoder",

utils/documentation_tests.txt

@@ -35,6 +35,7 @@ src/transformers/models/bloom/configuration_bloom.py
 src/transformers/models/camembert/configuration_camembert.py
 src/transformers/models/canine/configuration_canine.py
 src/transformers/models/clip/configuration_clip.py
+src/transformers/models/clipseg/modeling_clipseg.py
 src/transformers/models/codegen/configuration_codegen.py
 src/transformers/models/conditional_detr/configuration_conditional_detr.py
 src/transformers/models/conditional_detr/modeling_conditional_detr.py
@@ -187,4 +188,4 @@ src/transformers/models/xlnet/configuration_xlnet.py
 src/transformers/models/yolos/configuration_yolos.py
 src/transformers/models/yolos/modeling_yolos.py
 src/transformers/models/x_clip/modeling_x_clip.py
-src/transformers/models/yoso/configuration_yoso.py
+src/transformers/models/yoso/configuration_yoso.py
\ No newline at end of file