birder 0.3.0__py3-none-any.whl → 0.3.2__py3-none-any.whl

birder/net/densenet.py

@@ -140,14 +140,14 @@ class DenseNet(DetectorBackbone):
 
     def freeze_stages(self, up_to_stage: int) -> None:
         for param in self.stem.parameters():
-            param.requires_grad = False
+            param.requires_grad_(False)
 
         for idx, module in enumerate(self.body.children()):
             if idx >= up_to_stage:
                 break
 
             for param in module.parameters():
-                param.requires_grad = False
+                param.requires_grad_(False)
 
     def forward_features(self, x: torch.Tensor) -> torch.Tensor:
         x = self.stem(x)

birder/net/detection/deformable_detr.py

@@ -633,11 +633,11 @@ class Deformable_DETR(DetectionBaseNet):
 
     def freeze(self, freeze_classifier: bool = True) -> None:
         for param in self.parameters():
-            param.requires_grad = False
+            param.requires_grad_(False)
 
         if freeze_classifier is False:
             for param in self.class_embed.parameters():
-                param.requires_grad = True
+                param.requires_grad_(True)
 
     def _get_src_permutation_idx(self, indices: list[torch.Tensor]) -> tuple[torch.Tensor, torch.Tensor]:
         batch_idx = torch.concat([torch.full_like(src, i) for i, (src, _) in enumerate(indices)])

birder/net/detection/detr.py

@@ -356,11 +356,11 @@ class DETR(DetectionBaseNet):
 
     def freeze(self, freeze_classifier: bool = True) -> None:
         for param in self.parameters():
-            param.requires_grad = False
+            param.requires_grad_(False)
 
         if freeze_classifier is False:
             for param in self.class_embed.parameters():
-                param.requires_grad = True
+                param.requires_grad_(True)
 
     def _get_src_permutation_idx(self, indices: list[torch.Tensor]) -> tuple[torch.Tensor, torch.Tensor]:
         batch_idx = torch.concat([torch.full_like(src, i) for i, (src, _) in enumerate(indices)])

birder/net/detection/efficientdet.py

@@ -601,11 +601,11 @@ class EfficientDet(DetectionBaseNet):
 
     def freeze(self, freeze_classifier: bool = True) -> None:
         for param in self.parameters():
-            param.requires_grad = False
+            param.requires_grad_(False)
 
         if freeze_classifier is False:
             for param in self.class_net.parameters():
-                param.requires_grad = True
+                param.requires_grad_(True)
 
     def compute_loss(
         self,

birder/net/detection/faster_rcnn.py

@@ -851,11 +851,11 @@ class Faster_RCNN(DetectionBaseNet):
 
     def freeze(self, freeze_classifier: bool = True) -> None:
         for param in self.parameters():
-            param.requires_grad = False
+            param.requires_grad_(False)
 
         if freeze_classifier is False:
             for param in self.roi_heads.box_predictor.parameters():
-                param.requires_grad = True
+                param.requires_grad_(True)
 
     def forward(
         self,

birder/net/detection/fcos.py

@@ -338,11 +338,11 @@ class FCOS(DetectionBaseNet):
 
     def freeze(self, freeze_classifier: bool = True) -> None:
         for param in self.parameters():
-            param.requires_grad = False
+            param.requires_grad_(False)
 
         if freeze_classifier is False:
             for param in self.head.classification_head.parameters():
-                param.requires_grad = True
+                param.requires_grad_(True)
 
     def compute_loss(
         self,

birder/net/detection/retinanet.py

@@ -332,11 +332,11 @@ class RetinaNet(DetectionBaseNet):
 
     def freeze(self, freeze_classifier: bool = True) -> None:
         for param in self.parameters():
-            param.requires_grad = False
+            param.requires_grad_(False)
 
         if freeze_classifier is False:
             for param in self.head.classification_head.parameters():
-                param.requires_grad = True
+                param.requires_grad_(True)
 
     @torch.jit.unused  # type: ignore[untyped-decorator]
     @torch.compiler.disable()  # type: ignore[untyped-decorator]

birder/net/detection/rt_detr_v1.py

@@ -742,16 +742,16 @@ class RT_DETR_v1(DetectionBaseNet):
 
     def freeze(self, freeze_classifier: bool = True) -> None:
         for param in self.parameters():
-            param.requires_grad = False
+            param.requires_grad_(False)
 
         if freeze_classifier is False:
             for param in self.decoder.class_embed.parameters():
-                param.requires_grad = True
+                param.requires_grad_(True)
             for param in self.decoder.enc_score_head.parameters():
-                param.requires_grad = True
+                param.requires_grad_(True)
             if self.num_denoising > 0:
                 for param in self.denoising_class_embed.parameters():
-                    param.requires_grad = True
+                    param.requires_grad_(True)
 
     def _get_src_permutation_idx(self, indices: list[torch.Tensor]) -> tuple[torch.Tensor, torch.Tensor]:
         batch_idx = torch.concat([torch.full_like(src, i) for i, (src, _) in enumerate(indices)])

birder/net/detection/ssd.py

@@ -341,11 +341,11 @@ class SSD(DetectionBaseNet):
 
     def freeze(self, freeze_classifier: bool = True) -> None:
         for param in self.parameters():
-            param.requires_grad = False
+            param.requires_grad_(False)
 
         if freeze_classifier is False:
             for param in self.head.classification_head.parameters():
-                param.requires_grad = True
+                param.requires_grad_(True)
 
     # pylint: disable=too-many-locals
     def compute_loss(

birder/net/detection/ssdlite.py

@@ -197,8 +197,8 @@ class SSDLite(SSD):
 
     def freeze(self, freeze_classifier: bool = True) -> None:
         for param in self.parameters():
-            param.requires_grad = False
+            param.requires_grad_(False)
 
         if freeze_classifier is False:
             for param in self.head.classification_head.parameters():
-                param.requires_grad = True
+                param.requires_grad_(True)

birder/net/detection/yolo_v2.py

@@ -270,11 +270,11 @@ class YOLO_v2(DetectionBaseNet):
 
     def freeze(self, freeze_classifier: bool = True) -> None:
         for param in self.parameters():
-            param.requires_grad = False
+            param.requires_grad_(False)
 
         if freeze_classifier is False:
             for param in self.head.parameters():
-                param.requires_grad = True
+                param.requires_grad_(True)
 
     def _compute_anchor_iou(self, box_wh: torch.Tensor, anchor_wh: torch.Tensor) -> torch.Tensor:
         """

birder/net/detection/yolo_v3.py

@@ -376,11 +376,11 @@ class YOLO_v3(DetectionBaseNet):
 
     def freeze(self, freeze_classifier: bool = True) -> None:
         for param in self.parameters():
-            param.requires_grad = False
+            param.requires_grad_(False)
 
         if freeze_classifier is False:
             for param in self.head.parameters():
-                param.requires_grad = True
+                param.requires_grad_(True)
 
     def _compute_anchor_iou(self, box_wh: torch.Tensor, anchor_wh: torch.Tensor) -> torch.Tensor:
         """

birder/net/detection/yolo_v4.py

@@ -444,11 +444,11 @@ class YOLO_v4(DetectionBaseNet):
 
     def freeze(self, freeze_classifier: bool = True) -> None:
         for param in self.parameters():
-            param.requires_grad = False
+            param.requires_grad_(False)
 
         if freeze_classifier is False:
             for param in self.head.parameters():
-                param.requires_grad = True
+                param.requires_grad_(True)
 
     def _compute_anchor_iou(self, box_wh: torch.Tensor, anchor_wh: torch.Tensor) -> torch.Tensor:
         """

birder/net/edgenext.py

@@ -336,14 +336,14 @@ class EdgeNeXt(DetectorBackbone):
 
     def freeze_stages(self, up_to_stage: int) -> None:
         for param in self.stem.parameters():
-            param.requires_grad = False
+            param.requires_grad_(False)
 
         for idx, module in enumerate(self.body.children()):
             if idx >= up_to_stage:
                 break
 
             for param in module.parameters():
-                param.requires_grad = False
+                param.requires_grad_(False)
 
     def forward_features(self, x: torch.Tensor) -> torch.Tensor:
         x = self.stem(x)

birder/net/edgevit.py

@@ -354,7 +354,7 @@ class EdgeViT(DetectorBackbone):
                 break
 
             for param in module.parameters():
-                param.requires_grad = False
+                param.requires_grad_(False)
 
     def forward_features(self, x: torch.Tensor) -> torch.Tensor:
         return self.body(x)

birder/net/efficientformer_v1.py

@@ -307,18 +307,18 @@ class EfficientFormer_v1(BaseNet):
 
     def freeze(self, freeze_classifier: bool = True, unfreeze_features: bool = False) -> None:
         for param in self.parameters():
-            param.requires_grad = False
+            param.requires_grad_(False)
 
         if freeze_classifier is False:
             for param in self.classifier.parameters():
-                param.requires_grad = True
+                param.requires_grad_(True)
 
             for param in self.dist_classifier.parameters():
-                param.requires_grad = True
+                param.requires_grad_(True)
 
         if unfreeze_features is True:
             for param in self.features.parameters():
-                param.requires_grad = True
+                param.requires_grad_(True)
 
     def forward_features(self, x: torch.Tensor) -> torch.Tensor:
         x = self.stem(x)

birder/net/efficientformer_v2.py

@@ -469,18 +469,18 @@ class EfficientFormer_v2(DetectorBackbone):
 
     def freeze(self, freeze_classifier: bool = True, unfreeze_features: bool = False) -> None:
         for param in self.parameters():
-            param.requires_grad = False
+            param.requires_grad_(False)
 
         if freeze_classifier is False:
             for param in self.classifier.parameters():
-                param.requires_grad = True
+                param.requires_grad_(True)
 
             for param in self.dist_classifier.parameters():
-                param.requires_grad = True
+                param.requires_grad_(True)
 
         if unfreeze_features is True:
             for param in self.features.parameters():
-                param.requires_grad = True
+                param.requires_grad_(True)
 
     def transform_to_backbone(self) -> None:
         self.features = nn.Identity()
@@ -500,14 +500,14 @@ class EfficientFormer_v2(DetectorBackbone):
 
     def freeze_stages(self, up_to_stage: int) -> None:
         for param in self.stem.parameters():
-            param.requires_grad = False
+            param.requires_grad_(False)
 
         for idx, module in enumerate(self.body.children()):
             if idx >= up_to_stage:
                 break
 
             for param in module.parameters():
-                param.requires_grad = False
+                param.requires_grad_(False)
 
     def forward_features(self, x: torch.Tensor) -> torch.Tensor:
         x = self.stem(x)

birder/net/efficientnet_lite.py

@@ -242,14 +242,14 @@ class EfficientNet_Lite(DetectorBackbone, PreTrainEncoder, MaskedTokenRetentionM
 
     def freeze_stages(self, up_to_stage: int) -> None:
         for param in self.stem.parameters():
-            param.requires_grad = False
+            param.requires_grad_(False)
 
         for idx, module in enumerate(self.body.children()):
             if idx >= up_to_stage:
                 break
 
             for param in module.parameters():
-                param.requires_grad = False
+                param.requires_grad_(False)
 
     def masked_encoding_retention(
         self,

birder/net/efficientnet_v1.py

@@ -249,14 +249,14 @@ class EfficientNet_v1(DetectorBackbone, PreTrainEncoder, MaskedTokenRetentionMix
 
     def freeze_stages(self, up_to_stage: int) -> None:
         for param in self.stem.parameters():
-            param.requires_grad = False
+            param.requires_grad_(False)
 
         for idx, module in enumerate(self.body.children()):
             if idx >= up_to_stage:
                 break
 
             for param in module.parameters():
-                param.requires_grad = False
+                param.requires_grad_(False)
 
     def masked_encoding_retention(
         self,

birder/net/efficientnet_v2.py

@@ -257,14 +257,14 @@ class EfficientNet_v2(DetectorBackbone, PreTrainEncoder, MaskedTokenRetentionMix
 
     def freeze_stages(self, up_to_stage: int) -> None:
         for param in self.stem.parameters():
-            param.requires_grad = False
+            param.requires_grad_(False)
 
         for idx, module in enumerate(self.body.children()):
             if idx >= up_to_stage:
                 break
 
             for param in module.parameters():
-                param.requires_grad = False
+                param.requires_grad_(False)
 
     def masked_encoding_retention(
         self,

birder/net/efficientvim.py

@@ -418,21 +418,21 @@ class EfficientViM(DetectorBackbone):
 
     def freeze_stages(self, up_to_stage: int) -> None:
         for param in self.stem.parameters():
-            param.requires_grad = False
+            param.requires_grad_(False)
 
         for idx, module in enumerate(self.body.children()):
             if idx >= up_to_stage:
                 break
 
             for param in module.parameters():
-                param.requires_grad = False
+                param.requires_grad_(False)
 
         for idx, module in enumerate(self.norm.children()):
             if idx >= up_to_stage:
                 break
 
             for param in module.parameters():
-                param.requires_grad = False
+                param.requires_grad_(False)
 
     def forward_features(self, x: torch.Tensor) -> tuple[torch.Tensor, list[torch.Tensor]]:
         x = self.stem(x)

birder/net/efficientvit_mit.py

@@ -619,14 +619,14 @@ class EfficientViT_MIT(DetectorBackbone):
 
     def freeze_stages(self, up_to_stage: int) -> None:
         for param in self.stem.parameters():
-            param.requires_grad = False
+            param.requires_grad_(False)
 
         for idx, module in enumerate(self.body.children()):
             if idx >= up_to_stage:
                 break
 
             for param in module.parameters():
-                param.requires_grad = False
+                param.requires_grad_(False)
 
     def forward_features(self, x: torch.Tensor) -> torch.Tensor:
         x = self.stem(x)

birder/net/efficientvit_msft.py

@@ -436,14 +436,14 @@ class EfficientViT_MSFT(DetectorBackbone):
 
     def freeze_stages(self, up_to_stage: int) -> None:
         for param in self.stem.parameters():
-            param.requires_grad = False
+            param.requires_grad_(False)
 
         for idx, module in enumerate(self.body.children()):
             if idx >= up_to_stage:
                 break
 
             for param in module.parameters():
-                param.requires_grad = False
+                param.requires_grad_(False)
 
     def forward_features(self, x: torch.Tensor) -> torch.Tensor:
         x = self.stem(x)

birder/net/fasternet.py

@@ -199,14 +199,14 @@ class FasterNet(DetectorBackbone):
 
     def freeze_stages(self, up_to_stage: int) -> None:
         for param in self.stem.parameters():
-            param.requires_grad = False
+            param.requires_grad_(False)
 
         for idx, module in enumerate(self.body.children()):
             if idx >= up_to_stage:
                 break
 
             for param in module.parameters():
-                param.requires_grad = False
+                param.requires_grad_(False)
 
     def forward_features(self, x: torch.Tensor) -> torch.Tensor:
         x = self.stem(x)

birder/net/fastvit.py

@@ -607,7 +607,6 @@ class AttentionBlock(nn.Module):
 
 
 class FastVitStage(nn.Module):
-    # pylint: disable=too-many-arguments,too-many-positional-arguments
     def __init__(
         self,
         dim: int,
@@ -843,14 +842,14 @@ class FastViT(DetectorBackbone, PreTrainEncoder, MaskedTokenRetentionMixin):
 
     def freeze_stages(self, up_to_stage: int) -> None:
         for param in self.stem.parameters():
-            param.requires_grad = False
+            param.requires_grad_(False)
 
         for idx, module in enumerate(self.body.children()):
             if idx >= up_to_stage:
                 break
 
             for param in module.parameters():
-                param.requires_grad = False
+                param.requires_grad_(False)
 
     def masked_encoding_retention(
         self,

birder/net/flexivit.py

@@ -98,6 +98,8 @@ class FlexiViT(DetectorBackbone, PreTrainEncoder, MaskedTokenOmissionMixin, Mask
         layer_scale_init_value: Optional[float] = self.config.get("layer_scale_init_value", None)
         pre_norm: bool = self.config.get("pre_norm", False)
         post_norm: bool = self.config.get("post_norm", True)
+        qkv_bias: bool = self.config.get("qkv_bias", True)
+        qk_norm: bool = self.config.get("qk_norm", False)
         num_reg_tokens: int = self.config.get("num_reg_tokens", 0)
         class_token: bool = self.config.get("class_token", True)
         attn_pool_head: bool = self.config.get("attn_pool_head", False)
@@ -186,6 +188,8 @@ class FlexiViT(DetectorBackbone, PreTrainEncoder, MaskedTokenOmissionMixin, Mask
             attention_dropout,
             dpr,
             pre_norm=pre_norm,
+            qkv_bias=qkv_bias,
+            qk_norm=qk_norm,
             activation_layer=act_layer,
             layer_scale_init_value=layer_scale_init_value,
             norm_layer=norm_layer,
@@ -224,6 +228,7 @@ class FlexiViT(DetectorBackbone, PreTrainEncoder, MaskedTokenOmissionMixin, Mask
             drop_path=0,
             activation_layer=act_layer,
             norm_layer=norm_layer,
+            norm_layer_eps=norm_layer_eps,
             mlp_layer=mlp_layer,
         )
 
@@ -258,16 +263,16 @@ class FlexiViT(DetectorBackbone, PreTrainEncoder, MaskedTokenOmissionMixin, Mask
 
     def freeze(self, freeze_classifier: bool = True, unfreeze_features: bool = False) -> None:
         for param in self.parameters():
-            param.requires_grad = False
+            param.requires_grad_(False)
 
         if freeze_classifier is False:
             for param in self.classifier.parameters():
-                param.requires_grad = True
+                param.requires_grad_(True)
 
         if unfreeze_features is True:
             if self.attn_pool is not None:
                 for param in self.attn_pool.parameters():
-                    param.requires_grad = True
+                    param.requires_grad_(True)
 
     def set_causal_attention(self, is_causal: bool = True) -> None:
         self.encoder.set_causal_attention(is_causal)
@@ -305,16 +310,16 @@ class FlexiViT(DetectorBackbone, PreTrainEncoder, MaskedTokenOmissionMixin, Mask
 
     def freeze_stages(self, up_to_stage: int) -> None:
         for param in self.conv_proj.parameters():
-            param.requires_grad = False
+            param.requires_grad_(False)
 
-        self.pos_embedding.requires_grad = False
+        self.pos_embedding.requires_grad_(False)
 
         for idx, module in enumerate(self.encoder.children()):
             if idx >= up_to_stage:
                 break
 
             for param in module.parameters():
-                param.requires_grad = False
+                param.requires_grad_(False)
 
     # pylint: disable=too-many-branches
     def masked_encoding_omission(

birder/net/focalnet.py

@@ -245,7 +245,6 @@ class FocalNetBlock(nn.Module):
 
 
 class FocalNetStage(nn.Module):
-    # pylint: disable=too-many-arguments,too-many-positional-arguments
     def __init__(
         self,
         dim: int,
@@ -398,14 +397,14 @@ class FocalNet(DetectorBackbone, PreTrainEncoder, MaskedTokenRetentionMixin):
 
     def freeze_stages(self, up_to_stage: int) -> None:
         for param in self.stem.parameters():
-            param.requires_grad = False
+            param.requires_grad_(False)
 
         for idx, module in enumerate(self.body.children()):
             if idx >= up_to_stage:
                 break
 
             for param in module.parameters():
-                param.requires_grad = False
+                param.requires_grad_(False)
 
     def masked_encoding_retention(
         self,

birder/net/gc_vit.py

@@ -500,14 +500,14 @@ class GC_ViT(DetectorBackbone, PreTrainEncoder, MaskedTokenRetentionMixin):
 
     def freeze_stages(self, up_to_stage: int) -> None:
         for param in self.stem.parameters():
-            param.requires_grad = False
+            param.requires_grad_(False)
 
         for idx, module in enumerate(self.body.children()):
             if idx >= up_to_stage:
                 break
 
             for param in module.parameters():
-                param.requires_grad = False
+                param.requires_grad_(False)
 
     def set_dynamic_size(self, dynamic_size: bool = True) -> None:
         super().set_dynamic_size(dynamic_size)
@@ -669,3 +669,18 @@ registry.register_model_config(
         "drop_path_rate": 0.5,
     },
 )
+
+registry.register_weights(
+    "gc_vit_xxt_il-common",
+    {
+        "description": "GC ViT XX-Tiny model trained on the il-common dataset",
+        "resolution": (256, 256),
+        "formats": {
+            "pt": {
+                "file_size": 47.9,
+                "sha256": "5326a53903759e32178a6c2994639e6d0172faa51e1573a700f8d12b4f447c61",
+            }
+        },
+        "net": {"network": "gc_vit_xxt", "tag": "il-common"},
+    },
+)

birder/net/ghostnet_v1.py

@@ -254,14 +254,14 @@ class GhostNet_v1(DetectorBackbone):
 
     def freeze_stages(self, up_to_stage: int) -> None:
         for param in self.stem.parameters():
-            param.requires_grad = False
+            param.requires_grad_(False)
 
         for idx, module in enumerate(self.body.children()):
             if idx >= up_to_stage:
                 break
 
             for param in module.parameters():
-                param.requires_grad = False
+                param.requires_grad_(False)
 
     def forward_features(self, x: torch.Tensor) -> torch.Tensor:
         x = self.stem(x)

birder/net/ghostnet_v2.py

@@ -317,14 +317,14 @@ class GhostNet_v2(DetectorBackbone):
 
     def freeze_stages(self, up_to_stage: int) -> None:
         for param in self.stem.parameters():
-            param.requires_grad = False
+            param.requires_grad_(False)
 
         for idx, module in enumerate(self.body.children()):
             if idx >= up_to_stage:
                 break
 
             for param in module.parameters():
-                param.requires_grad = False
+                param.requires_grad_(False)
 
     def forward_features(self, x: torch.Tensor) -> torch.Tensor:
         x = self.stem(x)

birder/net/groupmixformer.py

@@ -391,14 +391,14 @@ class GroupMixFormer(DetectorBackbone):
 
     def freeze_stages(self, up_to_stage: int) -> None:
         for param in self.stem.parameters():
-            param.requires_grad = False
+            param.requires_grad_(False)
 
         for idx, module in enumerate(self.body.children()):
             if idx >= up_to_stage:
                 break
 
             for param in module.parameters():
-                param.requires_grad = False
+                param.requires_grad_(False)
 
     def forward_features(self, x: torch.Tensor) -> torch.Tensor:
         x = self.stem(x)