Create and Expose SamVisionModel as public for better accessibility (#36493)

* move encoder below

* auto modeling

* write SamVisionTester

* fix vision attention shape

* fix SamVisionTest

* minor changes to SamVisionTest

* Revert "fix vision attention shape"

This reverts commit d2a4083.

* fix attention output shape in new tests

* remove encoder examples

* run modular on got_ocr2

* code formatting

* fix got_ocr2

* ruff fixes

* code quality

* add sam_vision in auto modeling and auto configuration

* remove composite test

* updated index.md

* add TFSamVisionEncoder to __init__

* fix public TFSamVisionEncoder

* remove outdated todo comment

* set test_torch_exportable

Co-authored-by: Pavel Iakubovskii <qubvel@gmail.com>

* rename: VisionEncoder -> VisionModel

* bring back original SamVisionEncoder

* rename back: VisionEncoderOutput -> VisionModelOutput

* undo changes in SamModelTester

* reuse SamVisionEncoder in SamVisionModel

---------

Co-authored-by: Pavel Iakubovskii <qubvel@gmail.com>

Author: geetu040

docs/source/en/model_doc/sam.md

@@ -149,12 +149,24 @@ alt="drawing" width="900"/>
 [[autodoc]] SamImageProcessor
 
 
+## SamVisionModel
+
+[[autodoc]] SamVisionModel
+    - forward
+
+
 ## SamModel
 
 [[autodoc]] SamModel
     - forward
 
 
+## TFSamVisionModel
+
+[[autodoc]] TFSamVisionModel
+    - call
+
+
 ## TFSamModel
 
 [[autodoc]] TFSamModel

src/transformers/__init__.py

@@ -3589,6 +3589,7 @@
         [
             "SamModel",
             "SamPreTrainedModel",
+            "SamVisionModel",
         ]
     )
     _import_structure["models.seamless_m4t"].extend(
@@ -4757,6 +4758,7 @@
         [
             "TFSamModel",
             "TFSamPreTrainedModel",
+            "TFSamVisionModel",
         ]
     )
     _import_structure["models.segformer"].extend(
@@ -8431,6 +8433,7 @@
         from .models.sam import (
             SamModel,
             SamPreTrainedModel,
+            SamVisionModel,
         )
         from .models.seamless_m4t import (
             SeamlessM4TCodeHifiGan,
@@ -9372,6 +9375,7 @@
         from .models.sam import (
             TFSamModel,
             TFSamPreTrainedModel,
+            TFSamVisionModel,
         )
         from .models.segformer import (
             TFSegformerDecodeHead,

src/transformers/models/auto/configuration_auto.py

@@ -273,6 +273,7 @@
         ("rt_detr_v2", "RTDetrV2Config"),
         ("rwkv", "RwkvConfig"),
         ("sam", "SamConfig"),
+        ("sam_vision_model", "SamVisionConfig"),
         ("seamless_m4t", "SeamlessM4TConfig"),
         ("seamless_m4t_v2", "SeamlessM4Tv2Config"),
         ("segformer", "SegformerConfig"),
@@ -630,6 +631,7 @@
         ("rt_detr_v2", "RT-DETRv2"),
         ("rwkv", "RWKV"),
         ("sam", "SAM"),
+        ("sam_vision_model", "SamVisionModel"),
         ("seamless_m4t", "SeamlessM4T"),
         ("seamless_m4t_v2", "SeamlessM4Tv2"),
         ("segformer", "SegFormer"),
@@ -773,6 +775,7 @@
         ("chinese_clip_vision_model", "chinese_clip"),
         ("rt_detr_resnet", "rt_detr"),
         ("granitevision", "llava_next"),
+        ("sam_vision_model", "sam"),
     ]
 )
 

src/transformers/models/auto/modeling_auto.py

@@ -249,6 +249,7 @@
         ("rt_detr_v2", "RTDetrV2Model"),
         ("rwkv", "RwkvModel"),
         ("sam", "SamModel"),
+        ("sam_vision_model", "SamVisionModel"),
         ("seamless_m4t", "SeamlessM4TModel"),
         ("seamless_m4t_v2", "SeamlessM4Tv2Model"),
         ("segformer", "SegformerModel"),

src/transformers/models/auto/modeling_tf_auto.py

@@ -80,6 +80,7 @@
         ("roberta-prelayernorm", "TFRobertaPreLayerNormModel"),
         ("roformer", "TFRoFormerModel"),
         ("sam", "TFSamModel"),
+        ("sam_vision_model", "TFSamVisionModel"),
         ("segformer", "TFSegformerModel"),
         ("speech_to_text", "TFSpeech2TextModel"),
         ("swiftformer", "TFSwiftFormerModel"),

src/transformers/models/sam/configuration_sam.py

@@ -183,9 +183,27 @@ class SamVisionConfig(PretrainedConfig):
         mlp_dim (`int`, *optional*):
             The dimensionality of the MLP layer in the Transformer encoder. If `None`, defaults to `mlp_ratio *
             hidden_size`.
-    """
+
+    Example:
+
+    ```python
+    >>> from transformers import (
+    ...     SamVisionConfig,
+    ...     SamVisionModel,
+    ... )
+
+    >>> # Initializing a SamVisionConfig with `"facebook/sam-vit-huge"` style configuration
+    >>> configuration = SamVisionConfig()
+
+    >>> # Initializing a SamVisionModel (with random weights) from the `"facebook/sam-vit-huge"` style configuration
+    >>> model = SamVisionModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
 
     base_config_key = "vision_config"
+    model_type = "sam_vision_model"
 
     def __init__(
         self,

src/transformers/models/sam/modeling_sam.py

@@ -27,7 +27,13 @@
 from ...activations import ACT2FN
 from ...modeling_outputs import BaseModelOutput
 from ...modeling_utils import PreTrainedModel
-from ...utils import ModelOutput, add_start_docstrings, add_start_docstrings_to_model_forward, logging
+from ...utils import (
+    ModelOutput,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    logging,
+    replace_return_docstrings,
+)
 from .configuration_sam import SamConfig, SamMaskDecoderConfig, SamPromptEncoderConfig, SamVisionConfig
 
 
@@ -1280,6 +1286,61 @@ def _init_weights(self, module):
 """
 
 
+SAM_VISION_INPUTS_DOCSTRING = r"""
+    Args:
+        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            Pixel values. Pixel values can be obtained using [`SamProcessor`]. See [`SamProcessor.__call__`] for
+            details.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+@add_start_docstrings(
+    """The vision model from Sam without any head or projection on top.""",
+    SAM_START_DOCSTRING,
+)
+class SamVisionModel(SamPreTrainedModel):
+    config_class = SamVisionConfig
+    main_input_name = "pixel_values"
+
+    def __init__(self, config: SamVisionConfig):
+        super().__init__(config)
+        self.vision_encoder = SamVisionEncoder(config)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self) -> nn.Module:
+        return self.vision_encoder.patch_embed
+
+    @add_start_docstrings_to_model_forward(SAM_VISION_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=SamVisionEncoderOutput, config_class=SamVisionConfig)
+    def forward(
+        self,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, SamVisionEncoderOutput]:
+        r"""
+        Returns:
+
+        """
+        return self.vision_encoder(
+            pixel_values,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+
 @add_start_docstrings(
     "Segment Anything Model (SAM) for generating segmentation masks, given an input image and ",
     " optional 2D location and bounding boxes.",
@@ -1522,4 +1583,4 @@ def forward(
         )
 
 
-__all__ = ["SamModel", "SamPreTrainedModel"]
+__all__ = ["SamVisionModel", "SamModel", "SamPreTrainedModel"]

src/transformers/models/sam/modeling_tf_sam.py

@@ -30,7 +30,13 @@
 from ...modeling_tf_outputs import TFBaseModelOutput
 from ...modeling_tf_utils import TFModelInputType, TFPreTrainedModel, keras, shape_list, unpack_inputs
 from ...tf_utils import flatten, functional_layernorm
-from ...utils import ModelOutput, add_start_docstrings, add_start_docstrings_to_model_forward, logging
+from ...utils import (
+    ModelOutput,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    logging,
+    replace_return_docstrings,
+)
 from .configuration_sam import SamConfig, SamMaskDecoderConfig, SamPromptEncoderConfig, SamVisionConfig
 
 
@@ -1400,6 +1406,70 @@ class TFSamPreTrainedModel(TFPreTrainedModel):
 """
 
 
+SAM_VISION_INPUTS_DOCSTRING = r"""
+    Args:
+        pixel_values (`tf.Tensor` of shape `(batch_size, num_channels, height, width)`):
+            Pixel values. Pixel values can be obtained using [`SamProcessor`]. See [`SamProcessor.__call__`] for
+            details.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+@add_start_docstrings(
+    """The vision model from Sam without any head or projection on top.""",
+    SAM_START_DOCSTRING,
+)
+class TFSamVisionModel(TFSamPreTrainedModel):
+    config_class = SamVisionConfig
+    main_input_name = "pixel_values"
+
+    def __init__(self, config: SamVisionConfig, **kwargs):
+        super().__init__(config, **kwargs)
+        self.vision_encoder = TFSamVisionEncoder(config, name="vision_encoder")
+
+    def build(self, input_shape=None):
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "vision_encoder", None) is not None:
+            with tf.name_scope(self.vision_encoder.name):
+                self.vision_encoder.build(None)
+
+    def get_input_embeddings(self):
+        return self.vision_encoder.patch_embed
+
+    @unpack_inputs
+    @add_start_docstrings_to_model_forward(SAM_VISION_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=TFSamVisionEncoderOutput, config_class=SamVisionConfig)
+    def call(
+        self,
+        pixel_values: TFModelInputType | None = None,
+        output_attentions: bool | None = None,
+        output_hidden_states: bool | None = None,
+        return_dict: bool | None = None,
+        training: bool = False,
+        **kwargs,
+    ) -> TFSamVisionEncoderOutput | Tuple[tf.Tensor]:
+        r"""
+        Returns:
+
+        """
+        return self.vision_encoder(
+            pixel_values,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            training=training,
+        )
+
+
 @add_start_docstrings(
     "Segment Anything Model (SAM) for generating segmentation masks, given an input image and ",
     " optional 2D location and bounding boxes.",
@@ -1653,4 +1723,4 @@ def build(self, input_shape=None):
                 self.mask_decoder.build(None)
 
 
-__all__ = ["TFSamModel", "TFSamPreTrainedModel"]
+__all__ = ["TFSamVisionModel", "TFSamModel", "TFSamPreTrainedModel"]

src/transformers/utils/dummy_pt_objects.py

@@ -8836,6 +8836,13 @@ def __init__(self, *args, **kwargs):
         requires_backends(self, ["torch"])
 
 
+class SamVisionModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
 class SeamlessM4TCodeHifiGan(metaclass=DummyObject):
     _backends = ["torch"]
 

src/transformers/utils/dummy_tf_objects.py

@@ -2375,6 +2375,13 @@ def __init__(self, *args, **kwargs):
         requires_backends(self, ["tf"])
 
 
+class TFSamVisionModel(metaclass=DummyObject):
+    _backends = ["tf"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["tf"])
+
+
 class TFSegformerDecodeHead(metaclass=DummyObject):
     _backends = ["tf"]