Upload model

config.json

@@ -55,7 +55,7 @@
   "pretrained": null,
   "qk_scale": null,
   "qkv_bias": true,
-  "semantic_weight": 0.2,
+  "semantic_weight": 0.5,
   "strides": [
     4,
     2,

configuration_solider.py

@@ -39,7 +39,7 @@ class SOLIDERConfig(PretrainedConfig):
         convert_weights=False,
         frozen_stages=-1,
         init_cfg=None,
-        semantic_weight=0.2, # NOTE: I modified this from the implemenation of SOLIDER
+        semantic_weight=0.5, # NOTE: I modified this from the implemenation of SOLIDER
         name="solider_small",
         **kwargs,
     ):

modeling_solider.py

@@ -1654,7 +1654,7 @@ SOLIDER_BASE_MODEL_CONFIG_PARAMETERS = {
     "convert_weights": False,
     "frozen_stages": -1,
     "init_cfg": None,
-    "semantic_weight": 0.2,
+    "semantic_weight": 0.5,
     "name": "solider_base",
 }
 
@@ -1683,7 +1683,7 @@ SOLIDER_SMALL_MODEL_CONFIG_PARAMETERS = {
     "convert_weights": False,
     "frozen_stages": -1,
     "init_cfg": None,
-    "semantic_weight": 0.2,
+    "semantic_weight": 0.5,
     "name": "solider_small",
 }
 
@@ -1712,7 +1712,7 @@ SOLIDER_TINY_MODEL_CONFIG_PARAMETERS = {
     "convert_weights": False,
     "frozen_stages": -1,
     "init_cfg": None,
-    "semantic_weight": 0.2,
+    "semantic_weight": 0.5,
     "name": "solider_tiny",
 }
 
@@ -1729,7 +1729,7 @@ def build_solider_vision_encoder(weight_path, name="swin_small_patch4_window7_22
                 "name": name,
                 "img_size": [384, 128],
                 "pretrained": weight_path,
-                "semantic_weight": 0.2,
+                "semantic_weight": 0.5,
             }
         ),
         vision_width,
@@ -1776,10 +1776,12 @@ class SOLIDERModel(PreTrainedModel):
         # self.init_weights()
 
     def forward(self, x, semantic_weight=None):
-        # if semantic_weight is None, use the default value from config
+        if semantic_weight is None:
+            semantic_weight = self.config.semantic_weight
         return self.solider(x, semantic_weight)
 
 
+# NOTE: Currently not used!
 class SoliderEncoder(SwinTransformer):
     options = [
         "swin_tiny_patch4_window7_224",
@@ -1789,6 +1791,7 @@ class SoliderEncoder(SwinTransformer):
 
     @classmethod
     def from_config(cls, cfg, from_pretrained=None):
+        
         name = cfg.get("name", "swin_small_patch4_window7_224")
         img_size = cfg.get("img_size", [384, 128])
         drop_path_rate = cfg.get("drop_path_rate", 0.1)
@@ -1834,7 +1837,11 @@ class SoliderEncoder(SwinTransformer):
             state_dict_vision_encoder = torch.load(from_pretrained, map_location="cpu")
             msg = model.load_state_dict(state_dict_vision_encoder)
             print(msg)
+        model.config = cfg
         return model
 
     def forward_features(self, x, semantic_weight=None):
+        if semantic_weight is None:
+            semantic_weight = self.config.semantic_weight
         return SwinTransformer.forward(self, x, semantic_weight)
+