birder 0.3.1__py3-none-any.whl → 0.3.3__py3-none-any.whl

birder/adversarial/deepfool.py

@@ -2,6 +2,8 @@
 DeepFool
 
 Paper "DeepFool: a simple and accurate method to fool deep neural networks", https://arxiv.org/abs/1511.04599
+
+Generated by gpt-5.2-codex xhigh.
 """
 
 from typing import Optional

birder/adversarial/simba.py

@@ -2,6 +2,8 @@
 SimBA (Simple Black-box Attack)
 
 Paper "Simple Black-box Adversarial Attacks", https://arxiv.org/abs/1905.07121
+
+Generated by gpt-5.2-codex xhigh.
 """
 
 from typing import Optional

birder/common/masking.py

@@ -84,8 +84,8 @@ def mask_tensor(
 
     (B, H, W, _) = x.size()
 
-    shaped_mask = mask.reshape(-1, H // patch_factor, W // patch_factor)
-    shaped_mask = shaped_mask.repeat_interleave(patch_factor, axis=1).repeat_interleave(patch_factor, axis=2)
+    shaped_mask = mask.reshape(B, H // patch_factor, W // patch_factor)
+    shaped_mask = shaped_mask.repeat_interleave(patch_factor, dim=1).repeat_interleave(patch_factor, dim=2)
     shaped_mask = shaped_mask.unsqueeze(3).type_as(x)
 
     if mask_token is not None:
@@ -228,14 +228,23 @@ class Masking:
 
 
 class UniformMasking(Masking):
-    def __init__(self, input_size: tuple[int, int], mask_ratio: float, device: Optional[torch.device] = None) -> None:
+    def __init__(
+        self,
+        input_size: tuple[int, int],
+        mask_ratio: float,
+        min_mask_size: int = 1,
+        device: Optional[torch.device] = None,
+    ) -> None:
         self.h = input_size[0]
         self.w = input_size[1]
         self.mask_ratio = mask_ratio
+        self.min_mask_size = min_mask_size
         self.device = device
 
     def __call__(self, batch_size: int) -> torch.Tensor:
-        return uniform_mask(batch_size, self.h, self.w, self.mask_ratio, device=self.device)[0]
+        return uniform_mask(
+            batch_size, self.h, self.w, self.mask_ratio, min_mask_size=self.min_mask_size, device=self.device
+        )[0]
 
 
 class BlockMasking(Masking):

birder/inference/classification.py

@@ -85,7 +85,7 @@ def infer_batch(
             logits = net(t(tta_input), **kwargs)
             outs.append(logits if return_logits is True else F.softmax(logits, dim=1))
 
-        out = torch.stack(outs).mean(axis=0)
+        out = torch.stack(outs).mean(dim=0)
 
     else:
         logits = net(inputs, **kwargs)

birder/introspection/__init__.py

@@ -1,5 +1,6 @@
 from birder.introspection.attention_rollout import AttentionRollout
 from birder.introspection.base import InterpretabilityResult
+from birder.introspection.feature_pca import FeaturePCA
 from birder.introspection.gradcam import GradCAM
 from birder.introspection.guided_backprop import GuidedBackprop
 from birder.introspection.transformer_attribution import TransformerAttribution
@@ -7,6 +8,7 @@ from birder.introspection.transformer_attribution import TransformerAttribution
 __all__ = [
     "InterpretabilityResult",
     "AttentionRollout",
+    "FeaturePCA",
     "GradCAM",
     "GuidedBackprop",
     "TransformerAttribution",

birder/introspection/base.py

@@ -2,7 +2,6 @@ from collections.abc import Callable
 from dataclasses import dataclass
 from pathlib import Path
 from typing import Optional
-from typing import Protocol
 
 import matplotlib
 import matplotlib.pyplot as plt
@@ -27,12 +26,6 @@ class InterpretabilityResult:
         plt.show()
 
 
-class Interpreter(Protocol):
-    def __call__(
-        self, image: str | Path | Image.Image, target_class: Optional[int] = None
-    ) -> InterpretabilityResult: ...
-
-
 def load_image(image: str | Path | Image.Image) -> Image.Image:
     if isinstance(image, (str, Path)):
         return Image.open(image)

birder/introspection/feature_pca.py

@@ -0,0 +1,101 @@
+from collections.abc import Callable
+from pathlib import Path
+from typing import Optional
+
+import numpy as np
+import torch
+from PIL import Image
+from sklearn.decomposition import PCA
+
+from birder.introspection.base import InterpretabilityResult
+from birder.introspection.base import preprocess_image
+from birder.net.base import DetectorBackbone
+
+
+class FeaturePCA:
+    """
+    Visualizes feature maps using Principal Component Analysis
+
+    This method extracts feature maps from a specified stage of a DetectorBackbone model,
+    applies PCA to reduce the channel dimension to 3 components, and visualizes them as an RGB image where:
+    - R channel = 1st principal component (most important)
+    - G channel = 2nd principal component
+    - B channel = 3rd principal component
+    """
+
+    def __init__(
+        self,
+        net: DetectorBackbone,
+        device: torch.device,
+        transform: Callable[..., torch.Tensor],
+        normalize: bool = False,
+        channels_last: bool = False,
+        stage: Optional[str] = None,
+    ) -> None:
+        self.net = net.eval()
+        self.device = device
+        self.transform = transform
+        self.normalize = normalize
+        self.channels_last = channels_last
+        self.stage = stage
+
+    def __call__(self, image: str | Path | Image.Image) -> InterpretabilityResult:
+        (input_tensor, rgb_img) = preprocess_image(image, self.transform, self.device)
+
+        with torch.inference_mode():
+            features_dict = self.net.detection_features(input_tensor)
+
+        if self.stage is not None:
+            features = features_dict[self.stage]
+        else:
+            features = list(features_dict.values())[-1]  # Use the last stage by default
+
+        features_np = features.cpu().numpy()
+
+        # Handle channels_last format (B, H, W, C) vs channels_first (B, C, H, W)
+        if self.channels_last is True:
+            (B, H, W, C) = features_np.shape
+            # Already in (B, H, W, C), just reshape to (B*H*W, C)
+            features_reshaped = features_np.reshape(-1, C)
+        else:
+            (B, C, H, W) = features_np.shape
+            # Reshape to (spatial_points, channels) for PCA
+            features_reshaped = features_np.reshape(B, C, -1)
+            features_reshaped = features_reshaped.transpose(0, 2, 1)  # (B, H*W, C)
+            features_reshaped = features_reshaped.reshape(-1, C)  # (B*H*W, C)
+
+        x = features_reshaped
+        if self.normalize is True:
+            x = x / (np.linalg.norm(x, axis=1, keepdims=True) + 1e-6)
+
+        pca = PCA(n_components=3)
+        pca_features = pca.fit_transform(x)
+        pca_features = pca_features.reshape(B, H, W, 3)
+
+        # Extract all 3 components (B=1)
+        pca_rgb = pca_features[0]  # (H, W, 3)
+
+        # Normalize each channel independently to [0, 1]
+        for i in range(3):
+            channel = pca_rgb[:, :, i]
+            channel = channel - channel.min()
+            channel = channel / (channel.max() + 1e-7)
+            pca_rgb[:, :, i] = channel
+
+        target_size = (input_tensor.size(-1), input_tensor.size(-2))  # PIL expects (width, height)
+        pca_rgb_resized = (
+            np.array(
+                Image.fromarray((pca_rgb * 255).astype(np.uint8)).resize(target_size, Image.Resampling.BILINEAR)
+            ).astype(np.float32)
+            / 255.0
+        )
+
+        visualization = (pca_rgb_resized * 255).astype(np.uint8)
+
+        return InterpretabilityResult(
+            original_image=rgb_img,
+            visualization=visualization,
+            raw_output=pca_rgb.astype(np.float32),
+            logits=None,
+            predicted_class=None,
+        )

birder/kernels/soft_nms/soft_nms.cpp

@@ -4,6 +4,9 @@
 * Taken from:
 * https://github.com/MrParosk/soft_nms
 * Licensed under the MIT License
+*
+* Modified by:
+* Ofer Hasson — 2026-01-10
 **************************************************************************************************
 */
 
@@ -40,8 +43,8 @@ torch::Tensor calculate_iou(const torch::Tensor& boxes, const torch::Tensor& are
     auto xx2 = torch::minimum(boxes.index({idx, 2}), boxes.index({Slice(idx + 1, None), 2}));
     auto yy2 = torch::minimum(boxes.index({idx, 3}), boxes.index({Slice(idx + 1, None), 3}));
 
-    auto w = torch::maximum(torch::zeros_like(xx1), xx2 - xx1);
-    auto h = torch::maximum(torch::zeros_like(yy1), yy2 - yy1);
+    auto w = (xx2 - xx1).clamp_min(0);
+    auto h = (yy2 - yy1).clamp_min(0);
 
     auto intersection = w * h;
     auto union_ = areas.index({idx}) + areas.index({Slice(idx + 1, None)}) - intersection;

birder/model_registry/model_registry.py

@@ -87,14 +87,15 @@ class ModelRegistry:
         no further registration is needed.
         """
 
+        alias_key = alias.lower()
         if net_type.auto_register is False:
             # Register the model manually, as the base class doesn't take care of that for us
-            registry.register_model(alias, type(alias, (net_type,), {"config": config}))
+            self.register_model(alias_key, type(alias, (net_type,), {"config": config}))
 
         if alias in self.aliases:
             warnings.warn(f"Alias {alias} is already registered", UserWarning)
 
-        self.aliases[alias] = type(alias, (net_type,), {"config": config})
+        self.aliases[alias_key] = type(alias, (net_type,), {"config": config})
 
     def register_weights(self, name: str, weights_info: manifest.ModelMetadataType) -> None:
         if name in self._pretrained_nets:

birder/net/convnext_v1.py

@@ -195,6 +195,21 @@ class ConvNeXt_v1(DetectorBackbone, PreTrainEncoder, MaskedTokenRetentionMixin):
         return self.features(x)
 
 
+registry.register_model_config(
+    "convnext_v1_atto",  # Not in the original v1, taken from v2
+    ConvNeXt_v1,
+    config={"in_channels": [40, 80, 160, 320], "num_layers": [2, 2, 6, 2], "drop_path_rate": 0.0},
+)
+registry.register_model_config(
+    "convnext_v1_femto",  # Not in the original v1, taken from v2
+    ConvNeXt_v1,
+    config={"in_channels": [48, 96, 192, 384], "num_layers": [2, 2, 6, 2], "drop_path_rate": 0.0},
+)
+registry.register_model_config(
+    "convnext_v1_pico",  # Not in the original v1, taken from v2
+    ConvNeXt_v1,
+    config={"in_channels": [64, 128, 256, 512], "num_layers": [2, 2, 6, 2], "drop_path_rate": 0.0},
+)
 registry.register_model_config(
     "convnext_v1_nano",  # Not in the original v1, taken from v2
     ConvNeXt_v1,
@@ -220,6 +235,11 @@ registry.register_model_config(
     ConvNeXt_v1,
     config={"in_channels": [192, 384, 768, 1536], "num_layers": [3, 3, 27, 3], "drop_path_rate": 0.5},
 )
+registry.register_model_config(
+    "convnext_v1_huge",  # Not in the original v1, taken from v2
+    ConvNeXt_v1,
+    config={"in_channels": [352, 704, 1408, 2816], "num_layers": [3, 3, 27, 3], "drop_path_rate": 0.5},
+)
 
 registry.register_weights(
     "convnext_v1_tiny_eu-common256px",

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,

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,
         )
 

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,

birder/net/hiera.py

@@ -301,14 +301,14 @@ class HieraBlock(nn.Module):
         self.dim = dim
         self.dim_out = dim_out
 
-        self.norm1 = nn.LayerNorm(dim)
+        self.norm1 = nn.LayerNorm(dim, eps=1e-6)
         if dim != dim_out:
             self.proj = nn.Linear(dim, dim_out)
         else:
             self.proj = None
 
         self.attn = MaskUnitAttention(dim, dim_out, heads, q_stride, window_size, use_mask_unit_attn)
-        self.norm2 = nn.LayerNorm(dim_out)
+        self.norm2 = nn.LayerNorm(dim_out, eps=1e-6)
         self.mlp = MLP(dim_out, [int(dim_out * mlp_ratio), dim_out], activation_layer=nn.GELU)
         self.drop_path = StochasticDepth(drop_path, mode="row")
 
@@ -450,7 +450,7 @@ class Hiera(DetectorBackbone, PreTrainEncoder, MaskedTokenOmissionMixin):
         self.body = nn.Sequential(stages)
         self.features = nn.Sequential(
             attn_pool if attn_pool is not None else AvgTokens(),
-            nn.LayerNorm(embed_dim),
+            nn.LayerNorm(embed_dim, eps=1e-6),
             nn.Flatten(1),
         )
         self.return_channels = return_channels

birder/net/hieradet.py

@@ -125,7 +125,7 @@ class MultiScaleBlock(nn.Module):
         self.dim = dim
         self.dim_out = dim_out
 
-        self.norm1 = nn.LayerNorm(dim)
+        self.norm1 = nn.LayerNorm(dim, eps=1e-6)
         if dim != dim_out:
             self.proj = nn.Linear(dim, dim_out)
         else:
@@ -144,7 +144,7 @@ class MultiScaleBlock(nn.Module):
             num_heads=num_heads,
             q_pool=copy.deepcopy(self.pool),
         )
-        self.norm2 = nn.LayerNorm(dim_out)
+        self.norm2 = nn.LayerNorm(dim_out, eps=1e-6)
         self.mlp = MLP(dim_out, [int(dim_out * mlp_ratio), dim_out], activation_layer=nn.GELU)
         self.drop_path = StochasticDepth(drop_path, mode="row")
 
@@ -173,11 +173,9 @@ class MultiScaleBlock(nn.Module):
         if self.q_stride is not None:
             # Shapes have changed due to Q pooling
             window_size = self.window_size // self.q_stride[0]
-            (H, W) = (shortcut.size(1), shortcut.size(2))
+            pad_hw = (pad_hw[0] // self.q_stride[0], pad_hw[1] // self.q_stride[1])
 
-            pad_h = (window_size - H % window_size) % window_size
-            pad_w = (window_size - W % window_size) % window_size
-            pad_hw = (H + pad_h, W + pad_w)
+            (H, W) = (shortcut.size(1), shortcut.size(2))
 
         # Reverse window partition
         if self.window_size > 0:
@@ -271,7 +269,7 @@ class HieraDet(DetectorBackbone, PreTrainEncoder, MaskedTokenRetentionMixin):
 
         self.body = nn.Sequential(stages)
         self.features = nn.Sequential(
-            nn.LayerNorm(embed_dim),
+            nn.LayerNorm(embed_dim, eps=1e-6),
             Permute([0, 3, 1, 2]),  # B H W C -> B C H W
             nn.AdaptiveAvgPool2d(output_size=(1, 1)),
             nn.Flatten(1),
@@ -415,7 +413,7 @@ registry.register_model_config(
         "num_heads": 1,
         "global_pos_size": (7, 7),
         "global_att_blocks": [5, 7, 9],
-        "window_spec": [8, 4, -16, -32],
+        "window_spec": [8, 4, 14, 7],
         "drop_path_rate": 0.1,
     },
 )
@@ -428,7 +426,7 @@ registry.register_model_config(
         "num_heads": 1,
         "global_pos_size": (7, 7),
         "global_att_blocks": [7, 10, 13],
-        "window_spec": [8, 4, -16, -32],
+        "window_spec": [8, 4, 14, 7],
         "drop_path_rate": 0.1,
     },
 )
@@ -441,7 +439,7 @@ registry.register_model_config(
         "num_heads": 1,
         "global_pos_size": (14, 14),
         "global_att_blocks": [12, 16, 20],
-        "window_spec": [8, 4, -16, -32],
+        "window_spec": [8, 4, 14, 7],
         "drop_path_rate": 0.1,
     },
 )
@@ -454,7 +452,7 @@ registry.register_model_config(
         "num_heads": 2,
         "global_pos_size": (14, 14),
         "global_att_blocks": [12, 16, 20],
-        "window_spec": [8, 4, -16, -32],
+        "window_spec": [8, 4, 14, 7],
         "drop_path_rate": 0.1,
     },
 )
@@ -467,17 +465,84 @@ registry.register_model_config(
         "num_heads": 2,
         "global_pos_size": (7, 7),
         "global_att_blocks": [23, 33, 43],
-        "window_spec": [8, 4, -16, -32],
+        "window_spec": [8, 4, 14, 7],
+        "drop_path_rate": 0.2,
+    },
+)
+
+# Dynamic window size
+registry.register_model_config(
+    "hieradet_d_tiny",
+    HieraDet,
+    config={
+        "depths": [1, 2, 7, 2],
+        "embed_dim": 96,
+        "num_heads": 1,
+        "global_pos_size": (7, 7),
+        "global_att_blocks": [5, 7, 9],
+        "window_spec": [8, 4, 0, 0],
+        "drop_path_rate": 0.1,
+    },
+)
+registry.register_model_config(
+    "hieradet_d_small",
+    HieraDet,
+    config={
+        "depths": [1, 2, 11, 2],
+        "embed_dim": 96,
+        "num_heads": 1,
+        "global_pos_size": (7, 7),
+        "global_att_blocks": [7, 10, 13],
+        "window_spec": [8, 4, 0, 0],
+        "drop_path_rate": 0.1,
+    },
+)
+registry.register_model_config(
+    "hieradet_d_base",
+    HieraDet,
+    config={
+        "depths": [2, 3, 16, 3],
+        "embed_dim": 96,
+        "num_heads": 1,
+        "global_pos_size": (14, 14),
+        "global_att_blocks": [12, 16, 20],
+        "window_spec": [8, 4, 0, 0],
+        "drop_path_rate": 0.1,
+    },
+)
+registry.register_model_config(
+    "hieradet_d_base_plus",
+    HieraDet,
+    config={
+        "depths": [2, 3, 16, 3],
+        "embed_dim": 112,
+        "num_heads": 2,
+        "global_pos_size": (14, 14),
+        "global_att_blocks": [12, 16, 20],
+        "window_spec": [8, 4, 0, 0],
+        "drop_path_rate": 0.1,
+    },
+)
+registry.register_model_config(
+    "hieradet_d_large",
+    HieraDet,
+    config={
+        "depths": [2, 6, 36, 4],
+        "embed_dim": 144,
+        "num_heads": 2,
+        "global_pos_size": (7, 7),
+        "global_att_blocks": [23, 33, 43],
+        "window_spec": [8, 4, 0, 0],
         "drop_path_rate": 0.2,
     },
 )
 
 registry.register_weights(
-    "hieradet_small_dino-v2",
+    "hieradet_d_small_dino-v2",
     {
-        "url": "https://huggingface.co/birder-project/hieradet_small_dino-v2/resolve/main",
+        "url": "https://huggingface.co/birder-project/hieradet_d_small_dino-v2/resolve/main",
         "description": (
-            "HieraDet small image encoder pre-trained using DINOv2. "
+            "HieraDet (d) small image encoder pre-trained using DINOv2. "
             "This model has not been fine-tuned for a specific classification task"
         ),
         "resolution": (224, 224),
@@ -487,14 +552,16 @@ registry.register_weights(
                 "sha256": "eb41b8a35445e7f350797094d5e365306b29351e64edd4a316420c23d1e17073",
             }
         },
-        "net": {"network": "hieradet_small", "tag": "dino-v2"},
+        "net": {"network": "hieradet_d_small", "tag": "dino-v2"},
     },
 )
 registry.register_weights(
-    "hieradet_small_dino-v2-inat21-256px",
+    "hieradet_d_small_dino-v2-inat21-256px",
     {
-        "url": "https://huggingface.co/birder-project/hieradet_small_dino-v2-inat21/resolve/main",
-        "description": "HieraDet small model pre-trained using DINOv2, then fine-tuned on the iNaturalist 2021 dataset",
+        "url": "https://huggingface.co/birder-project/hieradet_d_small_dino-v2-inat21/resolve/main",
+        "description": (
+            "HieraDet (d) small model pre-trained using DINOv2, then fine-tuned on the iNaturalist 2021 dataset"
+        ),
         "resolution": (256, 256),
         "formats": {
             "pt": {
@@ -502,14 +569,16 @@ registry.register_weights(
                 "sha256": "e1bdeba97eae816ec3ab9b3238d97decf2c34d29b70f9291116ce962b9a4f9df",
             }
         },
-        "net": {"network": "hieradet_small", "tag": "dino-v2-inat21-256px"},
+        "net": {"network": "hieradet_d_small", "tag": "dino-v2-inat21-256px"},
     },
 )
 registry.register_weights(
-    "hieradet_small_dino-v2-inat21",
+    "hieradet_d_small_dino-v2-inat21",
     {
-        "url": "https://huggingface.co/birder-project/hieradet_small_dino-v2-inat21/resolve/main",
-        "description": "HieraDet small model pre-trained using DINOv2, then fine-tuned on the iNaturalist 2021 dataset",
+        "url": "https://huggingface.co/birder-project/hieradet_d_small_dino-v2-inat21/resolve/main",
+        "description": (
+            "HieraDet (d) small model pre-trained using DINOv2, then fine-tuned on the iNaturalist 2021 dataset"
+        ),
         "resolution": (384, 384),
         "formats": {
             "pt": {
@@ -517,14 +586,14 @@ registry.register_weights(
                 "sha256": "271fa9ed6a9aa1f4d1fc8bbb4c4cac9d15b264f2ac544efb5cd971412691880d",
             }
         },
-        "net": {"network": "hieradet_small", "tag": "dino-v2-inat21"},
+        "net": {"network": "hieradet_d_small", "tag": "dino-v2-inat21"},
     },
 )
 registry.register_weights(
-    "hieradet_small_dino-v2-imagenet12k",
+    "hieradet_d_small_dino-v2-imagenet12k",
     {
-        "url": "https://huggingface.co/birder-project/hieradet_small_dino-v2-imagenet12k/resolve/main",
-        "description": "HieraDet small model pre-trained using DINOv2, then fine-tuned on the ImageNet-12K dataset",
+        "url": "https://huggingface.co/birder-project/hieradet_d_small_dino-v2-imagenet12k/resolve/main",
+        "description": "HieraDet (d) small model pre-trained using DINOv2, then fine-tuned on the ImageNet-12K dataset",
         "resolution": (256, 256),
         "formats": {
             "pt": {
@@ -532,6 +601,25 @@ registry.register_weights(
                 "sha256": "b89dd6c13d061fe8a09d051bb3d76e632e650067ca71578e37b02033107c9963",
             }
         },
-        "net": {"network": "hieradet_small", "tag": "dino-v2-imagenet12k"},
+        "net": {"network": "hieradet_d_small", "tag": "dino-v2-imagenet12k"},
+    },
+)
+
+registry.register_weights(  # SAM v2: https://arxiv.org/abs/2408.00714
+    "hieradet_small_sam2_1",
+    {
+        "url": "https://huggingface.co/birder-project/hieradet_small_sam2_1/resolve/main",
+        "description": (
+            "HieraDet small image encoder pre-trained by Meta AI using SAM v2. "
+            "This model has not been fine-tuned for a specific classification task"
+        ),
+        "resolution": (224, 224),
+        "formats": {
+            "pt": {
+                "file_size": 129.6,
+                "sha256": "79b6ffdfd4ea9f3b1489ce5a229fe9756b215fc3b52640d01d64136560c1d341",
+            }
+        },
+        "net": {"network": "hieradet_small", "tag": "sam2_1"},
     },
 )

birder/net/rope_flexivit.py

@@ -69,6 +69,8 @@ class RoPE_FlexiViT(DetectorBackbone, PreTrainEncoder, MaskedTokenOmissionMixin,
         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)
@@ -118,6 +120,7 @@ class RoPE_FlexiViT(DetectorBackbone, PreTrainEncoder, MaskedTokenOmissionMixin,
         self.num_reg_tokens = num_reg_tokens
         self.attn_pool_special_tokens = attn_pool_special_tokens
         self.norm_layer = norm_layer
+        self.norm_layer_eps = norm_layer_eps
         self.mlp_layer = mlp_layer
         self.act_layer = act_layer
         self.rope_rot_type = rope_rot_type
@@ -190,6 +193,8 @@ class RoPE_FlexiViT(DetectorBackbone, PreTrainEncoder, MaskedTokenOmissionMixin,
             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,
@@ -231,6 +236,7 @@ class RoPE_FlexiViT(DetectorBackbone, PreTrainEncoder, MaskedTokenOmissionMixin,
             rope_temperature=rope_temperature,
             layer_scale_init_value=layer_scale_init_value,
             norm_layer=norm_layer,
+            norm_layer_eps=norm_layer_eps,
             mlp_layer=mlp_layer,
             rope_rot_type=rope_rot_type,
         )
@@ -588,6 +594,7 @@ class RoPE_FlexiViT(DetectorBackbone, PreTrainEncoder, MaskedTokenOmissionMixin,
             rope_temperature=self.rope_temperature,
             layer_scale_init_value=self.layer_scale_init_value,
             norm_layer=self.norm_layer,
+            norm_layer_eps=self.norm_layer_eps,
             mlp_layer=self.mlp_layer,
             rope_rot_type=self.rope_rot_type,
         )