birder 0.3.3__py3-none-any.whl → 0.4.0__py3-none-any.whl

birder/net/deit.py

@@ -16,6 +16,9 @@ import torch
 from torch import nn
 
 from birder.model_registry import registry
+from birder.net._vit_configs import BASE
+from birder.net._vit_configs import SMALL
+from birder.net._vit_configs import TINY
 from birder.net.base import BaseNet
 from birder.net.vit import Encoder
 from birder.net.vit import PatchEmbed
@@ -94,6 +97,10 @@ class DeiT(BaseNet):
         self.classifier = self.create_classifier()
         self.distillation_output = False
 
+        self.max_stride = patch_size
+        self.stem_stride = patch_size
+        self.stem_width = hidden_dim
+
         # Weight initialization
         if isinstance(self.conv_proj, nn.Conv2d):
             # Init the patchify stem
@@ -200,38 +207,17 @@ class DeiT(BaseNet):
 registry.register_model_config(
     "deit_t16",
     DeiT,
-    config={
-        "patch_size": 16,
-        "num_layers": 12,
-        "num_heads": 3,
-        "hidden_dim": 192,
-        "mlp_dim": 768,
-        "drop_path_rate": 0.0,
-    },
+    config={"patch_size": 16, **TINY},
 )
 registry.register_model_config(
     "deit_s16",
     DeiT,
-    config={
-        "patch_size": 16,
-        "num_layers": 12,
-        "num_heads": 6,
-        "hidden_dim": 384,
-        "mlp_dim": 1536,
-        "drop_path_rate": 0.1,
-    },
+    config={"patch_size": 16, **SMALL, "drop_path_rate": 0.1},  # Override the SMALL definition
 )
 registry.register_model_config(
     "deit_b16",
     DeiT,
-    config={
-        "patch_size": 16,
-        "num_layers": 12,
-        "num_heads": 12,
-        "hidden_dim": 768,
-        "mlp_dim": 3072,
-        "drop_path_rate": 0.1,
-    },
+    config={"patch_size": 16, **BASE},
 )
 
 registry.register_weights(
@@ -242,7 +228,7 @@ registry.register_weights(
         "formats": {
             "pt": {
                 "file_size": 21.7,
-                "sha256": "ac124122dec9f1bceff383a6a555ca375ca1b613caf486dac3f29d87afac03b3",
+                "sha256": "68b33aba0c1be5e78d4a33e74a7c1ea72b6abb232d59f0048ff9b8342e43246e",
             }
         },
         "net": {"network": "deit_t16", "tag": "il-common"},
@@ -258,7 +244,7 @@ registry.register_weights(
         "formats": {
             "pt": {
                 "file_size": 21.7,
-                "sha256": "fafd0c3c65f9c35318f449f60485f640917736ee7b44056be55c2226909ffdb8",
+                "sha256": "f693e89fc350341141c55152bec9f499df63738e8423071f3b8e71801c3e5415",
             }
         },
         "net": {"network": "deit_t16", "tag": "dist-il-common"},

birder/net/deit3.py

@@ -15,6 +15,12 @@ from torch import nn
 
 from birder.common.masking import mask_tensor
 from birder.model_registry import registry
+from birder.net._vit_configs import BASE
+from birder.net._vit_configs import HUGE
+from birder.net._vit_configs import LARGE
+from birder.net._vit_configs import MEDIUM
+from birder.net._vit_configs import SMALL
+from birder.net._vit_configs import TINY
 from birder.net.base import DetectorBackbone
 from birder.net.base import MaskedTokenOmissionMixin
 from birder.net.base import MaskedTokenRetentionMixin
@@ -368,279 +374,126 @@ class DeiT3(DetectorBackbone, PreTrainEncoder, MaskedTokenOmissionMixin, MaskedT
 registry.register_model_config(
     "deit3_t16",
     DeiT3,
-    config={
-        "patch_size": 16,
-        "num_layers": 12,
-        "num_heads": 3,
-        "hidden_dim": 192,
-        "mlp_dim": 768,
-        "drop_path_rate": 0.0,
-    },
+    config={"patch_size": 16, **TINY},
+)
+registry.register_model_config(
+    "deit3_t14",
+    DeiT3,
+    config={"patch_size": 14, **TINY},
 )
 registry.register_model_config(
     "deit3_s16",
     DeiT3,
-    config={
-        "patch_size": 16,
-        "num_layers": 12,
-        "num_heads": 6,
-        "hidden_dim": 384,
-        "mlp_dim": 1536,
-        "drop_path_rate": 0.05,
-    },
+    config={"patch_size": 16, **SMALL, "drop_path_rate": 0.05},
 )
 registry.register_model_config(
     "deit3_s14",
     DeiT3,
-    config={
-        "patch_size": 14,
-        "num_layers": 12,
-        "num_heads": 6,
-        "hidden_dim": 384,
-        "mlp_dim": 1536,
-        "drop_path_rate": 0.05,
-    },
+    config={"patch_size": 14, **SMALL, "drop_path_rate": 0.05},
 )
 registry.register_model_config(
     "deit3_m16",
     DeiT3,
-    config={
-        "patch_size": 16,
-        "num_layers": 12,
-        "num_heads": 8,
-        "hidden_dim": 512,
-        "mlp_dim": 2048,
-        "drop_path_rate": 0.1,
-    },
+    config={"patch_size": 16, **MEDIUM, "drop_path_rate": 0.1},
 )
 registry.register_model_config(
     "deit3_m14",
     DeiT3,
-    config={
-        "patch_size": 14,
-        "num_layers": 12,
-        "num_heads": 8,
-        "hidden_dim": 512,
-        "mlp_dim": 2048,
-        "drop_path_rate": 0.1,
-    },
+    config={"patch_size": 14, **MEDIUM, "drop_path_rate": 0.1},
 )
 registry.register_model_config(
     "deit3_b16",
     DeiT3,
-    config={
-        "patch_size": 16,
-        "num_layers": 12,
-        "num_heads": 12,
-        "hidden_dim": 768,
-        "mlp_dim": 3072,
-        "drop_path_rate": 0.2,
-    },
+    config={"patch_size": 16, **BASE, "drop_path_rate": 0.2},
 )
 registry.register_model_config(
     "deit3_b14",
     DeiT3,
-    config={
-        "patch_size": 14,
-        "num_layers": 12,
-        "num_heads": 12,
-        "hidden_dim": 768,
-        "mlp_dim": 3072,
-        "drop_path_rate": 0.2,
-    },
+    config={"patch_size": 14, **BASE, "drop_path_rate": 0.2},
 )
 registry.register_model_config(
     "deit3_l16",
     DeiT3,
-    config={
-        "patch_size": 16,
-        "num_layers": 24,
-        "num_heads": 16,
-        "hidden_dim": 1024,
-        "mlp_dim": 4096,
-        "drop_path_rate": 0.45,
-    },
+    config={"patch_size": 16, **LARGE, "drop_path_rate": 0.45},
 )
 registry.register_model_config(
     "deit3_l14",
     DeiT3,
-    config={
-        "patch_size": 14,
-        "num_layers": 24,
-        "num_heads": 16,
-        "hidden_dim": 1024,
-        "mlp_dim": 4096,
-        "drop_path_rate": 0.45,
-    },
+    config={"patch_size": 14, **LARGE, "drop_path_rate": 0.45},
 )
 registry.register_model_config(
     "deit3_h16",
     DeiT3,
-    config={
-        "patch_size": 16,
-        "num_layers": 32,
-        "num_heads": 16,
-        "hidden_dim": 1280,
-        "mlp_dim": 5120,
-        "drop_path_rate": 0.55,
-    },
+    config={"patch_size": 16, **HUGE, "drop_path_rate": 0.55},
 )
 registry.register_model_config(
     "deit3_h14",
     DeiT3,
-    config={
-        "patch_size": 14,
-        "num_layers": 32,
-        "num_heads": 16,
-        "hidden_dim": 1280,
-        "mlp_dim": 5120,
-        "drop_path_rate": 0.55,
-    },
+    config={"patch_size": 14, **HUGE, "drop_path_rate": 0.55},
 )
 
 # With registers
+####################
+
 registry.register_model_config(
     "deit3_reg4_t16",
     DeiT3,
-    config={
-        "patch_size": 16,
-        "num_layers": 12,
-        "num_heads": 3,
-        "hidden_dim": 192,
-        "mlp_dim": 768,
-        "num_reg_tokens": 4,
-        "drop_path_rate": 0.0,
-    },
+    config={"patch_size": 16, **TINY, "num_reg_tokens": 4},
+)
+registry.register_model_config(
+    "deit3_reg4_t14",
+    DeiT3,
+    config={"patch_size": 14, **TINY, "num_reg_tokens": 4},
 )
 registry.register_model_config(
     "deit3_reg4_s16",
     DeiT3,
-    config={
-        "patch_size": 16,
-        "num_layers": 12,
-        "num_heads": 6,
-        "hidden_dim": 384,
-        "mlp_dim": 1536,
-        "num_reg_tokens": 4,
-        "drop_path_rate": 0.05,
-    },
+    config={"patch_size": 16, **SMALL, "num_reg_tokens": 4, "drop_path_rate": 0.05},
 )
 registry.register_model_config(
     "deit3_reg4_s14",
     DeiT3,
-    config={
-        "patch_size": 14,
-        "num_layers": 12,
-        "num_heads": 6,
-        "hidden_dim": 384,
-        "mlp_dim": 1536,
-        "num_reg_tokens": 4,
-        "drop_path_rate": 0.05,
-    },
+    config={"patch_size": 14, **SMALL, "num_reg_tokens": 4, "drop_path_rate": 0.05},
 )
 registry.register_model_config(
     "deit3_reg4_m16",
     DeiT3,
-    config={
-        "patch_size": 16,
-        "num_layers": 12,
-        "num_heads": 8,
-        "hidden_dim": 512,
-        "mlp_dim": 2048,
-        "num_reg_tokens": 4,
-        "drop_path_rate": 0.1,
-    },
+    config={"patch_size": 16, **MEDIUM, "num_reg_tokens": 4, "drop_path_rate": 0.1},
 )
 registry.register_model_config(
     "deit3_reg4_m14",
     DeiT3,
-    config={
-        "patch_size": 14,
-        "num_layers": 12,
-        "num_heads": 8,
-        "hidden_dim": 512,
-        "mlp_dim": 2048,
-        "num_reg_tokens": 4,
-        "drop_path_rate": 0.1,
-    },
+    config={"patch_size": 14, **MEDIUM, "num_reg_tokens": 4, "drop_path_rate": 0.1},
 )
 registry.register_model_config(
     "deit3_reg4_b16",
     DeiT3,
-    config={
-        "patch_size": 16,
-        "num_layers": 12,
-        "num_heads": 12,
-        "hidden_dim": 768,
-        "mlp_dim": 3072,
-        "num_reg_tokens": 4,
-        "drop_path_rate": 0.2,
-    },
+    config={"patch_size": 16, **BASE, "num_reg_tokens": 4, "drop_path_rate": 0.2},
 )
 registry.register_model_config(
     "deit3_reg4_b14",
     DeiT3,
-    config={
-        "patch_size": 14,
-        "num_layers": 12,
-        "num_heads": 12,
-        "hidden_dim": 768,
-        "mlp_dim": 3072,
-        "num_reg_tokens": 4,
-        "drop_path_rate": 0.2,
-    },
+    config={"patch_size": 14, **BASE, "num_reg_tokens": 4, "drop_path_rate": 0.2},
 )
 registry.register_model_config(
     "deit3_reg4_l16",
     DeiT3,
-    config={
-        "patch_size": 16,
-        "num_layers": 24,
-        "num_heads": 16,
-        "hidden_dim": 1024,
-        "mlp_dim": 4096,
-        "num_reg_tokens": 4,
-        "drop_path_rate": 0.45,
-    },
+    config={"patch_size": 16, **LARGE, "num_reg_tokens": 4, "drop_path_rate": 0.45},
 )
 registry.register_model_config(
     "deit3_reg4_l14",
     DeiT3,
-    config={
-        "patch_size": 14,
-        "num_layers": 24,
-        "num_heads": 16,
-        "hidden_dim": 1024,
-        "mlp_dim": 4096,
-        "num_reg_tokens": 4,
-        "drop_path_rate": 0.45,
-    },
+    config={"patch_size": 14, **LARGE, "num_reg_tokens": 4, "drop_path_rate": 0.45},
 )
 registry.register_model_config(
     "deit3_reg4_h16",
     DeiT3,
-    config={
-        "patch_size": 16,
-        "num_layers": 32,
-        "num_heads": 16,
-        "hidden_dim": 1280,
-        "mlp_dim": 5120,
-        "num_reg_tokens": 4,
-        "drop_path_rate": 0.55,
-    },
+    config={"patch_size": 16, **HUGE, "num_reg_tokens": 4, "drop_path_rate": 0.55},
 )
 registry.register_model_config(
     "deit3_reg4_h14",
     DeiT3,
-    config={
-        "patch_size": 14,
-        "num_layers": 32,
-        "num_heads": 16,
-        "hidden_dim": 1280,
-        "mlp_dim": 5120,
-        "num_reg_tokens": 4,
-        "drop_path_rate": 0.55,
-    },
+    config={"patch_size": 14, **HUGE, "num_reg_tokens": 4, "drop_path_rate": 0.55},
 )
 
 registry.register_weights(
@@ -651,7 +504,7 @@ registry.register_weights(
         "formats": {
             "pt": {
                 "file_size": 21.5,
-                "sha256": "6cd9749a9522f8ff61088e38702553fb1c4d2547b417c499652e3bfa6a81e77a",
+                "sha256": "a04141c7f6c459ae075a48ccdee5b82d191bbaa82337673140c06ef82f0a8dc5",
             }
         },
         "net": {"network": "deit3_t16", "tag": "il-common"},
@@ -665,7 +518,7 @@ registry.register_weights(
         "formats": {
             "pt": {
                 "file_size": 21.5,
-                "sha256": "6806a5ae7d45f1c84b25e9869a9cbc7de94368fe9573dc3777acf2da8c83dc4e",
+                "sha256": "d26320462da64df6d62b307f7fb35d09c86a5f073002dfb24a51f014074e65c3",
             }
         },
         "net": {"network": "deit3_reg4_t16", "tag": "il-common"},

birder/net/densenet.py

@@ -104,19 +104,20 @@ class DenseNet(DetectorBackbone):
         num_features = num_init_features
         stages: OrderedDict[str, nn.Module] = OrderedDict()
         return_channels: list[int] = []
-        layers = []
         for i, num_layers in enumerate(layer_list):
+            stage_layers = []
+            if i != 0:
+                stage_layers.append(TransitionBlock(num_features, num_features // 2))
+                num_features = num_features // 2
 
-            layers.append(DenseBlock(num_features, num_layers=num_layers, growth_rate=growth_rate))
+            stage_layers.append(DenseBlock(num_features, num_layers=num_layers, growth_rate=growth_rate))
             num_features = num_features + (num_layers * growth_rate)
+            if i == len(layer_list) - 1:
+                stage_layers.append(nn.BatchNorm2d(num_features))
+                stage_layers.append(nn.ReLU(inplace=True))
 
-            stages[f"stage{i+1}"] = nn.Sequential(*layers)
+            stages[f"stage{i+1}"] = nn.Sequential(*stage_layers)
             return_channels.append(num_features)
-            layers = []
-
-            if i != len(layer_list) - 1:
-                layers.append(TransitionBlock(num_features, num_features // 2))
-                num_features = num_features // 2
 
         self.body = nn.Sequential(stages)
         self.features = nn.Sequential(

birder/net/detection/deformable_detr.py

@@ -9,7 +9,7 @@ https://arxiv.org/abs/2010.04159
 
 Changes from original:
 * Removed two stage support
-* Zero cost matrix elements on overflow (HungarianMatcher)
+* Penalize cost matrix elements on overflow (HungarianMatcher)
 """
 
 # Reference license: Apache-2.0 (both)
@@ -89,7 +89,10 @@ class HungarianMatcher(nn.Module):
             # Final cost matrix
             C = self.cost_bbox * cost_bbox + self.cost_class * cost_class + self.cost_giou * cost_giou
             C = C.view(B, num_queries, -1).cpu()
-            C[C.isnan() | C.isinf()] = 0.0
+            finite = torch.isfinite(C)
+            if not torch.all(finite):
+                penalty = C[finite].max().item() + 1.0 if finite.any().item() else 1.0
+                C.nan_to_num_(nan=penalty, posinf=penalty, neginf=penalty)
 
             sizes = [len(v["boxes"]) for v in targets]
             indices = [linear_sum_assignment(c[i]) for i, c in enumerate(C.split(sizes, -1))]

birder/net/detection/detr.py

@@ -6,7 +6,7 @@ Paper "End-to-End Object Detection with Transformers", https://arxiv.org/abs/200
 
 Changes from original:
 * Move background index to first from last (to be inline with the rest of Birder detectors)
-* Zero cost matrix elements on overflow (HungarianMatcher)
+* Penalize cost matrix elements on overflow (HungarianMatcher)
 """
 
 # Reference license: Apache-2.0
@@ -78,7 +78,10 @@ class HungarianMatcher(nn.Module):
             # Final cost matrix
             C = self.cost_bbox * cost_bbox + self.cost_class * cost_class + self.cost_giou * cost_giou
             C = C.view(B, num_queries, -1).cpu()
-            C[C.isnan() | C.isinf()] = 0.0
+            finite = torch.isfinite(C)
+            if not torch.all(finite):
+                penalty = C[finite].max().item() + 1.0 if finite.any().item() else 1.0
+                C.nan_to_num_(nan=penalty, posinf=penalty, neginf=penalty)
 
             sizes = [len(v["boxes"]) for v in targets]
             indices = [linear_sum_assignment(c[i]) for i, c in enumerate(C.split(sizes, -1))]

birder/net/detection/efficientdet.py

@@ -195,7 +195,7 @@ class FpnCombine(nn.Module):
             )
 
         if weight_method in {"attn", "fastattn"}:
-            self.edge_weights = nn.Parameter(torch.ones(len(inputs_offsets)), requires_grad=True)  # WSM
+            self.edge_weights = nn.Parameter(torch.ones(len(inputs_offsets)))  # WSM
         else:
             self.edge_weights = None
 

birder/net/dpn.py

@@ -181,9 +181,8 @@ class DPN(BaseNet):
                 layers.append(DualPathBlock(num_features, r, r, bw, inc, groups, "normal"))
                 num_features += inc
 
-        self.norm_act = nn.Sequential(nn.BatchNorm2d(num_features), nn.ELU())
-
         self.body = nn.Sequential(*layers)
+        self.norm_act = nn.Sequential(nn.BatchNorm2d(num_features), nn.ReLU())
         self.features = nn.Sequential(
             nn.AdaptiveAvgPool2d(output_size=(1, 1)),
             nn.Flatten(1),

birder/net/edgenext.py

@@ -10,6 +10,7 @@ https://arxiv.org/abs/2206.10589
 
 import math
 from collections import OrderedDict
+from functools import partial
 from typing import Any
 from typing import Optional
 
@@ -277,7 +278,7 @@ class EdgeNeXt(DetectorBackbone):
             stride=(4, 4),
             padding=(0, 0),
             bias=True,
-            norm_layer=LayerNorm2d,
+            norm_layer=partial(LayerNorm2d, eps=1e-6),
             activation_layer=None,
         )
 

birder/net/edgevit.py

@@ -4,6 +4,9 @@ https://github.com/saic-fi/edgevit/blob/master/src/edgevit.py
 
 Paper "EdgeViTs: Competing Light-weight CNNs on Mobile Devices with Vision Transformers",
 https://arxiv.org/abs/2205.03436
+
+Changes from original:
+* Removed classifier bias
 """
 
 # Reference license: Apache-2.0

birder/net/efficientformer_v1.py

@@ -9,6 +9,7 @@ https://arxiv.org/abs/2206.01191
 
 Changes from original:
 * Removed attention bias cache
+* Stem bias term removed
 """
 
 # Reference license: Apache-2.0 (both)
@@ -76,7 +77,7 @@ class Downsample(nn.Module):
         stride: tuple[int, int],
     ) -> None:
         super().__init__()
-        padding = (kernel_size[0] // 2, kernel_size[1] // 2)
+        padding = ((kernel_size[0] - 1) // 2, (kernel_size[1] - 1) // 2)
         self.conv = nn.Conv2d(in_channels, out_channels, kernel_size=kernel_size, stride=stride, padding=padding)
         self.norm = nn.BatchNorm2d(out_channels)
 

birder/net/efficientformer_v2.py

@@ -9,6 +9,7 @@ https://arxiv.org/abs/2212.08059
 
 Changes from original:
 * Removed attention bias cache
+* Removed biases before norms
 """
 
 # Reference license: Apache-2.0 (both)
@@ -244,9 +245,24 @@ class ConvMLP(nn.Module):
         drop: float,
     ) -> None:
         super().__init__()
-        self.fc1 = Conv2dNormActivation(in_features, hidden_features, kernel_size=(1, 1), stride=(1, 1), padding=(0, 0))
+        self.fc1 = Conv2dNormActivation(
+            in_features,
+            hidden_features,
+            kernel_size=(1, 1),
+            stride=(1, 1),
+            padding=(0, 0),
+            activation_layer=nn.GELU,
+            inplace=None,
+        )
         self.mid = Conv2dNormActivation(
-            hidden_features, hidden_features, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=hidden_features
+            hidden_features,
+            hidden_features,
+            kernel_size=(3, 3),
+            stride=(1, 1),
+            padding=(1, 1),
+            groups=hidden_features,
+            activation_layer=nn.GELU,
+            inplace=None,
         )
         self.drop1 = nn.Dropout(drop)
         self.fc2 = Conv2dNormActivation(
@@ -676,32 +692,3 @@ registry.register_model_config(
         ],
     },
 )
-
-registry.register_weights(
-    "efficientformer_v2_s0_il-common",
-    {
-        "description": "EfficientFormer v2 S0 model trained on the il-common dataset",
-        "resolution": (256, 256),
-        "formats": {
-            "pt": {
-                "file_size": 13.2,
-                "sha256": "b5ba923d351d45a04686b5bda037438719e0f442a41a34207a7f19737a8edb45",
-            }
-        },
-        "net": {"network": "efficientformer_v2_s0", "tag": "il-common"},
-    },
-)
-registry.register_weights(
-    "efficientformer_v2_s1_il-common",
-    {
-        "description": "EfficientFormer v2 S1 model trained on the il-common dataset",
-        "resolution": (256, 256),
-        "formats": {
-            "pt": {
-                "file_size": 22.9,
-                "sha256": "6b7ce6bbf5aa83e222cd16d8f07e749cdbb703fd383f99e88362ec8401d81401",
-            }
-        },
-        "net": {"network": "efficientformer_v2_s1", "tag": "il-common"},
-    },
-)

birder/net/efficientnet_v2.py

@@ -4,6 +4,9 @@ https://github.com/pytorch/vision/blob/main/torchvision/models/efficientnet.py
 
 Paper "EfficientNetV2: Smaller Models and Faster Training",
 https://arxiv.org/abs/2104.00298
+
+Changes from original:
+* Using nn.BatchNorm2d with eps 1e-5 instead of 1e-3
 """
 
 # Reference license: BSD 3-Clause