Add BackboneMixin

src/transformers/modeling_utils.py

@@ -15,6 +15,7 @@
 # limitations under the License.
 import collections
 import gc
+import inspect
 import json
 import os
 import re
@@ -932,6 +933,15 @@ def floating_point_ops(
         return 6 * self.estimate_tokens(input_dict) * self.num_parameters(exclude_embeddings=exclude_embeddings)
 
 
+class BackboneBaseModel(nn.Module):
+    def forward_with_filtered_kwargs(self, *args, **kwargs):
+
+        signature = dict(inspect.signature(self.forward).parameters)
+        filtered_kwargs = {k: v for k, v in kwargs.items() if k in signature}
+
+        return self(*args, **filtered_kwargs)
+
+
 class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMixin):
     r"""
     Base class for all models.

src/transformers/models/bit/modeling_bit.py

@@ -30,7 +30,7 @@
     BaseModelOutputWithPoolingAndNoAttention,
     ImageClassifierOutputWithNoAttention,
 )
-from ...modeling_utils import PreTrainedModel
+from ...modeling_utils import BackboneBaseModel, PreTrainedModel
 from ...utils import (
     add_code_sample_docstrings,
     add_start_docstrings,
@@ -842,7 +842,7 @@ def forward(
     """,
     BIT_START_DOCSTRING,
 )
-class BitBackbone(BitPreTrainedModel):
+class BitBackbone(BitPreTrainedModel, BackboneBaseModel):
     def __init__(self, config):
         super().__init__(config)
 

src/transformers/models/maskformer/modeling_maskformer_swin.py

@@ -27,7 +27,7 @@
 from ...activations import ACT2FN
 from ...file_utils import ModelOutput
 from ...modeling_outputs import BackboneOutput
-from ...modeling_utils import PreTrainedModel
+from ...modeling_utils import BackboneBaseModel, PreTrainedModel
 from ...pytorch_utils import find_pruneable_heads_and_indices, meshgrid, prune_linear_layer
 from .configuration_maskformer_swin import MaskFormerSwinConfig
 
@@ -837,7 +837,7 @@ def forward(
         )
 
 
-class MaskFormerSwinBackbone(MaskFormerSwinPreTrainedModel):
+class MaskFormerSwinBackbone(MaskFormerSwinPreTrainedModel, BackboneBaseModel):
     """
     MaskFormerSwin backbone, designed especially for the MaskFormer framework.
 

src/transformers/models/resnet/modeling_resnet.py

@@ -28,7 +28,7 @@
     BaseModelOutputWithPoolingAndNoAttention,
     ImageClassifierOutputWithNoAttention,
 )
-from ...modeling_utils import PreTrainedModel
+from ...modeling_utils import BackboneBaseModel, PreTrainedModel
 from ...utils import (
     add_code_sample_docstrings,
     add_start_docstrings,
@@ -431,7 +431,7 @@ def forward(
     """,
     RESNET_START_DOCSTRING,
 )
-class ResNetBackbone(ResNetPreTrainedModel):
+class ResNetBackbone(ResNetPreTrainedModel, BackboneBaseModel):
     def __init__(self, config):
         super().__init__(config)