Add kwargs for timm.create_model in TimmWrapper

src/transformers/models/timm_wrapper/configuration_timm_wrapper.py

@@ -15,7 +15,7 @@
 
 """Configuration for TimmWrapper models"""
 
-from typing import Any, Dict
+from typing import Any, Dict, Optional
 
 from ...configuration_utils import PretrainedConfig
 from ...utils import is_timm_available, logging, requires_backends
@@ -45,6 +45,9 @@ class TimmWrapperConfig(PretrainedConfig):
             The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
         do_pooling (`bool`, *optional*, defaults to `True`):
             Whether to do pooling for the last_hidden_state in `TimmWrapperModel` or not.
+        model_init_kwargs (`Dict[str, Any]`, *optional*):
+            Additional keyword arguments to pass to the `timm.create_model` function. e.g. `model_init_kwargs={"depth": 3}`
+            for `timm/vit_base_patch32_clip_448.laion2b_ft_in12k_in1k` to create a model with 3 blocks. Defaults to `None`.
 
     Example:
     ```python
@@ -60,9 +63,16 @@ class TimmWrapperConfig(PretrainedConfig):
 
     model_type = "timm_wrapper"
 
-    def __init__(self, initializer_range: float = 0.02, do_pooling: bool = True, **kwargs):
+    def __init__(
+        self,
+        initializer_range: float = 0.02,
+        do_pooling: bool = True,
+        model_init_kwargs: Optional[Dict[str, Any]] = None,
+        **kwargs,
+    ):
         self.initializer_range = initializer_range
         self.do_pooling = do_pooling
+        self.model_init_kwargs = model_init_kwargs
         super().__init__(**kwargs)
 
     @classmethod

src/transformers/models/timm_wrapper/modeling_timm_wrapper.py

@@ -116,7 +116,8 @@ class TimmWrapperModel(TimmWrapperPreTrainedModel):
     def __init__(self, config: TimmWrapperConfig):
         super().__init__(config)
         # using num_classes=0 to avoid creating classification head
-        self.timm_model = timm.create_model(config.architecture, pretrained=False, num_classes=0)
+        model_init_kwargs = config.model_init_kwargs or {}
+        self.timm_model = timm.create_model(config.architecture, pretrained=False, num_classes=0, **model_init_kwargs)
         self.post_init()
 
     @auto_docstring
@@ -233,7 +234,10 @@ def __init__(self, config: TimmWrapperConfig):
                 "or use `TimmWrapperModel` for feature extraction."
             )
 
-        self.timm_model = timm.create_model(config.architecture, pretrained=False, num_classes=config.num_labels)
+        model_init_kwargs = config.model_init_kwargs or {}
+        self.timm_model = timm.create_model(
+            config.architecture, pretrained=False, num_classes=config.num_labels, **model_init_kwargs
+        )
         self.num_labels = config.num_labels
         self.post_init()
 

tests/models/timm_wrapper/test_modeling_timm_wrapper.py

@@ -237,6 +237,17 @@ def test_timm_config_labels(self):
         self.assertEqual(config.id2label, restored_config.id2label)
         self.assertEqual(config.label2id, restored_config.label2id)
 
+    def test_model_init_kwargs(self):
+        config = TimmWrapperConfig.from_pretrained(
+            "timm/vit_base_patch32_clip_448.laion2b_ft_in12k_in1k",
+            model_init_kwargs={"depth": 3},
+        )
+        model = TimmWrapperModel(config)
+        self.assertEqual(len(model.timm_model.blocks), 3)
+
+        cls_model = TimmWrapperForImageClassification(config)
+        self.assertEqual(len(cls_model.timm_model.blocks), 3)
+
 
 # We will verify our results on an image of cute cats
 def prepare_img():