Upload model

modeling.py

@@ -12,25 +12,20 @@ from torch.nn.attention.flex_attention import (
     create_block_mask,
     flex_attention,
 )
-from torch.nn.utils.parametrizations import _Orthogonal, orthogonal
+from torch.nn.utils.parametrizations import _is_orthogonal, orthogonal
 from transformers.modeling_utils import PreTrainedModel
 from transformers.utils.backbone_utils import load_backbone
 
 from .configuration import LSPDetrConfig, STAConfig
 
 
-def _meta_safe_right_inverse(self, x: torch.Tensor) -> torch.Tensor:
-    if x.device.type == "meta":
-        # Skip validation; assume OK (unsafe but necessary for meta init)
-        return x
+def _meta_safe_is_orthogonal(Q, eps=None):
+    return True if Q.device == torch.device("meta") else _is_orthogonal(Q, eps=eps)
 
-    return _right_inverse(self, x)
 
-
-_right_inverse = _Orthogonal.right_inverse
 patch(
-    "torch.nn.utils.parametrizations._Orthogonal.right_inverse",
-    _meta_safe_right_inverse,
+    "torch.nn.utils.parametrizations._is_orthogonal",
+    _meta_safe_is_orthogonal,
 ).start()
 
 
@@ -384,7 +379,7 @@ class LSPTransformer(nn.Module):
     def init_weights(self) -> None:
         prior_prob = 0.01
         bias_value = -math.log((1 - prior_prob) / prior_prob)
-        self.class_head.bias.data = torch.ones(self.num_classes) * bias_value
+        nn.init.constant_(self.class_head.bias, bias_value)
 
         # initialize regression layers
         for head in self.point_head:
@@ -530,3 +525,6 @@ class LSPDetrModel(PreTrainedModel):
         )
 
         return self.decode_head(tgt, ref_points, features, h, w)
+
+
+LSPDetrModel.from_pretrained("RationAI/LSP-DETR", trust_remote_code=True)