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

birder/net/efficientvit_mit.py

@@ -54,7 +54,7 @@ class DSConv(nn.Module):
             in_channels,
             kernel_size=kernel_size,
             stride=stride,
-            padding=(kernel_size[0] // 2, kernel_size[1] // 2),
+            padding=((kernel_size[0] - 1) // 2, (kernel_size[1] - 1) // 2),
             groups=in_channels,
             norm_layer=norm_layer[0],
             activation_layer=act_layer[0],
@@ -96,7 +96,7 @@ class ConvBlock(nn.Module):
             mid_channels,
             kernel_size=kernel_size,
             stride=stride,
-            padding=(kernel_size[0] // 2, kernel_size[1] // 2),
+            padding=((kernel_size[0] - 1) // 2, (kernel_size[1] - 1) // 2),
             norm_layer=norm_layer[0],
             activation_layer=act_layer[0],
             inplace=None,
@@ -106,7 +106,7 @@ class ConvBlock(nn.Module):
             out_channels,
             kernel_size=kernel_size,
             stride=(1, 1),
-            padding=(kernel_size[0] // 2, kernel_size[1] // 2),
+            padding=((kernel_size[0] - 1) // 2, (kernel_size[1] - 1) // 2),
             norm_layer=norm_layer[1],
             activation_layer=act_layer[1],
             inplace=None,
@@ -148,7 +148,7 @@ class MBConv(nn.Module):
             mid_channels,
             kernel_size=kernel_size,
             stride=stride,
-            padding=(kernel_size[0] // 2, kernel_size[1] // 2),
+            padding=((kernel_size[0] - 1) // 2, (kernel_size[1] - 1) // 2),
             groups=mid_channels,
             norm_layer=norm_layer[1],
             activation_layer=act_layer[1],
@@ -192,7 +192,7 @@ class FusedMBConv(nn.Module):
             mid_channels,
             kernel_size=kernel_size,
             stride=stride,
-            padding=(kernel_size[0] // 2, kernel_size[1] // 2),
+            padding=((kernel_size[0] - 1) // 2, (kernel_size[1] - 1) // 2),
             norm_layer=norm_layer[0],
             activation_layer=act_layer[0],
             inplace=None,

birder/net/fasternet.py

@@ -50,10 +50,10 @@ class MLPBlock(nn.Module):
         mlp_hidden_dim = int(dim * mlp_ratio)
         self.spatial_mixing = PartialConv(dim, n_div)
         self.mlp = nn.Sequential(
-            nn.Conv2d(dim, mlp_hidden_dim, 1, bias=False),
+            nn.Conv2d(dim, mlp_hidden_dim, kernel_size=(1, 1), stride=(1, 1), padding=(0, 0), bias=False),
             nn.BatchNorm2d(mlp_hidden_dim),
             act_layer(),
-            nn.Conv2d(mlp_hidden_dim, dim, 1, bias=False),
+            nn.Conv2d(mlp_hidden_dim, dim, kernel_size=(1, 1), stride=(1, 1), padding=(0, 0), bias=False),
         )
         self.drop_path = StochasticDepth(drop_path, mode="row")
 

birder/net/flexivit.py

@@ -22,6 +22,8 @@ from birder.layers import MultiHeadAttentionPool
 from birder.layers import SwiGLU_FFN
 from birder.layers.activations import get_activation_module
 from birder.model_registry import registry
+from birder.net._vit_configs import BASE
+from birder.net._vit_configs import SMALL
 from birder.net.base import DetectorBackbone
 from birder.net.base import MaskedTokenOmissionMixin
 from birder.net.base import MaskedTokenRetentionMixin
@@ -583,70 +585,47 @@ class FlexiViT(DetectorBackbone, PreTrainEncoder, MaskedTokenOmissionMixin, Mask
 registry.register_model_config(
     "flexivit_s16",
     FlexiViT,
-    config={
-        "patch_size": 16,
-        "num_layers": 12,
-        "num_heads": 6,
-        "hidden_dim": 384,
-        "mlp_dim": 1536,
-        "drop_path_rate": 0.0,
-    },
+    config={"patch_size": 16, **SMALL},
 )
 registry.register_model_config(
     "flexivit_s16_ls",
     FlexiViT,
-    config={
-        "patch_size": 16,
-        "num_layers": 12,
-        "num_heads": 6,
-        "hidden_dim": 384,
-        "mlp_dim": 1536,
-        "layer_scale_init_value": 1e-5,
-        "drop_path_rate": 0.0,
-    },
+    config={"patch_size": 16, **SMALL, "layer_scale_init_value": 1e-5},
 )
+registry.register_model_config(
+    "flexivit_b16",
+    FlexiViT,
+    config={"patch_size": 16, **BASE},
+)
+
+# With registers
+####################
+
 registry.register_model_config(
     "flexivit_reg1_s16",
     FlexiViT,
-    config={
-        "patch_size": 16,
-        "num_layers": 12,
-        "num_heads": 6,
-        "hidden_dim": 384,
-        "mlp_dim": 1536,
-        "num_reg_tokens": 1,
-        "drop_path_rate": 0.0,
-    },
+    config={"patch_size": 16, **SMALL, "num_reg_tokens": 1},
 )
 registry.register_model_config(
     "flexivit_reg1_s16_rms_ls",
     FlexiViT,
     config={
         "patch_size": 16,
-        "num_layers": 12,
-        "num_heads": 6,
-        "hidden_dim": 384,
-        "mlp_dim": 1536,
+        **SMALL,
         "layer_scale_init_value": 1e-5,
         "num_reg_tokens": 1,
         "norm_layer_type": "RMSNorm",
-        "drop_path_rate": 0.0,
     },
 )
+registry.register_model_config(
+    "flexivit_reg4_b16",
+    FlexiViT,
+    config={"patch_size": 16, **BASE, "num_reg_tokens": 4},
+)
 registry.register_model_config(
     "flexivit_reg8_b14_ap",
     FlexiViT,
-    config={
-        "patch_size": 14,
-        "num_layers": 12,
-        "num_heads": 12,
-        "hidden_dim": 768,
-        "mlp_dim": 3072,
-        "num_reg_tokens": 8,
-        "class_token": False,
-        "attn_pool_head": True,
-        "drop_path_rate": 0.1,
-    },
+    config={"patch_size": 14, **BASE, "num_reg_tokens": 8, "class_token": False, "attn_pool_head": True},
 )
 
 registry.register_weights(
@@ -661,7 +640,7 @@ registry.register_weights(
         "formats": {
             "pt": {
                 "file_size": 83.6,
-                "sha256": "8d11fb14630f2a54632aeebd09c5a9c2b3b7de1099e09de5e91f433ed915b784",
+                "sha256": "8285f4fe56401f169491cb2399d2a7c82f3a0cfbe8a5a8d3c27163024a274800",
             },
         },
         "net": {"network": "flexivit_reg1_s16_rms_ls", "tag": "dino-v2-il-all"},

birder/net/groupmixformer.py

@@ -160,7 +160,7 @@ class ConvPosEnc(nn.Module):
             dim,
             kernel_size=kernel_size,
             stride=(1, 1),
-            padding=(kernel_size[0] // 2, kernel_size[1] // 2),
+            padding=((kernel_size[0] - 1) // 2, (kernel_size[1] - 1) // 2),
             groups=dim,
         )
 

birder/net/hgnet_v1.py

@@ -23,8 +23,8 @@ from birder.net.base import DetectorBackbone
 class LearnableAffineBlock(nn.Module):
     def __init__(self, scale_value: float, bias_value: float) -> None:
         super().__init__()
-        self.scale = nn.Parameter(torch.tensor([scale_value]), requires_grad=True)
-        self.bias = nn.Parameter(torch.tensor([bias_value]), requires_grad=True)
+        self.scale = nn.Parameter(torch.tensor([scale_value]))
+        self.bias = nn.Parameter(torch.tensor([bias_value]))
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:
         return self.scale * x + self.bias
@@ -90,7 +90,7 @@ class LightConvBNAct(nn.Module):
             out_channels,
             kernel_size=kernel_size,
             stride=(1, 1),
-            padding=(kernel_size[0] // 2, kernel_size[1] // 2),
+            padding=((kernel_size[0] - 1) // 2, (kernel_size[1] - 1) // 2),
             groups=out_channels,
             use_act=True,
             use_lab=use_lab,
@@ -180,7 +180,7 @@ class HighPerfGPUBlock(nn.Module):
                         mid_channels,
                         kernel_size=kernel_size,
                         stride=(1, 1),
-                        padding=(kernel_size[0] // 2, kernel_size[1] // 2),
+                        padding=((kernel_size[0] - 1) // 2, (kernel_size[1] - 1) // 2),
                         groups=1,
                         use_act=True,
                         use_lab=use_lab,
@@ -267,7 +267,7 @@ class HighPerfGPUStage(nn.Module):
                 in_channels,
                 kernel_size=(3, 3),
                 stride=stride,
-                padding=(kernel_size[0] // 2, kernel_size[1] // 2),
+                padding=(1, 1),
                 groups=in_channels,
                 use_act=False,
                 use_lab=False,

birder/net/inception_next.py

@@ -110,7 +110,7 @@ class InceptionNeXtBlock(nn.Module):
             nn.BatchNorm2d(channels),
             ConvMLP(channels, hidden_features=int(mlp_ratio * channels), out_features=channels),
         )
-        self.layer_scale = nn.Parameter(torch.ones(channels, 1, 1) * layer_scale, requires_grad=True)
+        self.layer_scale = nn.Parameter(torch.ones(channels, 1, 1) * layer_scale)
         self.stochastic_depth = StochasticDepth(stochastic_depth_prob, mode="row")
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:

birder/net/inception_resnet_v1.py

@@ -33,7 +33,7 @@ class InceptionBlockA(nn.Module):
             Conv2dNormActivation(32, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False),
         )
 
-        self.conv2d = nn.Conv2d(96, 256, kernel_size=(1, 1), stride=(1, 1), padding=(0, 0), bias=False)
+        self.conv2d = nn.Conv2d(96, 256, kernel_size=(1, 1), stride=(1, 1), padding=(0, 0))
         self.relu = nn.ReLU(inplace=True)
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:
@@ -84,7 +84,7 @@ class InceptionBlockB(nn.Module):
             Conv2dNormActivation(128, 128, kernel_size=(7, 1), stride=(1, 1), padding=(3, 0), bias=False),
         )
 
-        self.conv2d = nn.Conv2d(256, 896, kernel_size=(1, 1), stride=(1, 1), padding=(0, 0), bias=False)
+        self.conv2d = nn.Conv2d(256, 896, kernel_size=(1, 1), stride=(1, 1), padding=(0, 0))
         self.relu = nn.ReLU(inplace=True)
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:
@@ -140,7 +140,7 @@ class InceptionBlockC(nn.Module):
             Conv2dNormActivation(192, 192, kernel_size=(3, 1), stride=(1, 1), padding=(1, 0), bias=False),
         )
 
-        self.conv2d = nn.Conv2d(384, 1792, kernel_size=(1, 1), stride=(1, 1), padding=(0, 0), bias=False)
+        self.conv2d = nn.Conv2d(384, 1792, kernel_size=(1, 1), stride=(1, 1), padding=(0, 0))
         self.relu = nn.ReLU(inplace=True)
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:

birder/net/inception_resnet_v2.py

@@ -4,6 +4,9 @@ https://github.com/huggingface/pytorch-image-models/blob/main/timm/models/incept
 
 Paper "Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning",
 https://arxiv.org/abs/1602.07261
+
+Changes from original:
+* Using nn.BatchNorm2d with eps 1e-5 instead of 1e-3
 """
 
 # Reference license: Apache-2.0
@@ -35,7 +38,7 @@ class StemBlock(nn.Module):
             Conv2dNormActivation(96, 96, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False),
         )
         self.branch_pool = nn.Sequential(
-            nn.AvgPool2d(kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
+            nn.AvgPool2d(kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), count_include_pad=False),
             Conv2dNormActivation(in_channels, 64, kernel_size=(1, 1), stride=(1, 1), padding=(0, 0), bias=False),
         )
 
@@ -66,7 +69,7 @@ class InceptionBlockA(nn.Module):
             Conv2dNormActivation(48, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False),
         )
 
-        self.conv2d = nn.Conv2d(128, 320, kernel_size=(1, 1), stride=(1, 1), padding=(0, 0), bias=False)
+        self.conv2d = nn.Conv2d(128, 320, kernel_size=(1, 1), stride=(1, 1), padding=(0, 0))
         self.relu = nn.ReLU(inplace=True)
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:
@@ -117,7 +120,7 @@ class InceptionBlockB(nn.Module):
             Conv2dNormActivation(160, 192, kernel_size=(7, 1), stride=(1, 1), padding=(3, 0), bias=False),
         )
 
-        self.conv2d = nn.Conv2d(384, 1088, kernel_size=(1, 1), stride=(1, 1), padding=(0, 0), bias=False)
+        self.conv2d = nn.Conv2d(384, 1088, kernel_size=(1, 1), stride=(1, 1), padding=(0, 0))
         self.relu = nn.ReLU(inplace=True)
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:
@@ -174,7 +177,7 @@ class InceptionBlockC(nn.Module):
             Conv2dNormActivation(224, 256, kernel_size=(3, 1), stride=(1, 1), padding=(1, 0), bias=False),
         )
 
-        self.conv2d = nn.Conv2d(448, 2080, kernel_size=(1, 1), stride=(1, 1), padding=(0, 0), bias=False)
+        self.conv2d = nn.Conv2d(448, 2080, kernel_size=(1, 1), stride=(1, 1), padding=(0, 0))
         self.relu = nn.ReLU(inplace=True)
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:

birder/net/inception_v3.py

@@ -3,6 +3,9 @@ Inception v3, adapted from
 https://github.com/pytorch/vision/blob/main/torchvision/models/inception.py
 
 Paper "Rethinking the Inception Architecture for Computer Vision", https://arxiv.org/abs/1512.00567
+
+Changes from original:
+* Using nn.BatchNorm2d with eps 1e-5 instead of 1e-3
 """
 
 # Reference license: BSD 3-Clause

birder/net/inception_v4.py

@@ -4,6 +4,9 @@ https://github.com/huggingface/pytorch-image-models/blob/main/timm/models/incept
 
 Paper "Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning",
 https://arxiv.org/abs/1602.07261
+
+Changes from original:
+* Using nn.BatchNorm2d with eps 1e-5 instead of 1e-3
 """
 
 # Reference license: Apache-2.0

birder/net/maxvit.py

@@ -82,7 +82,7 @@ class MBConv(nn.Module):
 
         if stride[0] != 1 or stride[1] != 1 or in_channels != out_channels:
             self.proj = nn.Sequential(
-                nn.AvgPool2d(kernel_size=(2, 2), stride=stride, padding=(0, 0)),
+                nn.AvgPool2d(kernel_size=(3, 3), stride=stride, padding=(1, 1)),
                 nn.Conv2d(in_channels, out_channels, kernel_size=(1, 1), stride=(1, 1), padding=(0, 0), bias=True),
             )
         else:

birder/net/metaformer.py

@@ -81,7 +81,7 @@ class ConvMLP(nn.Module):
 class Scale(nn.Module):
     def __init__(self, dim: int, init_value: float) -> None:
         super().__init__()
-        self.scale = nn.Parameter(init_value * torch.ones(dim, 1, 1), requires_grad=True)
+        self.scale = nn.Parameter(init_value * torch.ones(dim, 1, 1))
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:
         return x * self.scale
@@ -104,8 +104,8 @@ class StarReLU(nn.Module):
     def __init__(self, scale_value: float = 1.0, bias_value: float = 0.0, inplace: bool = False) -> None:
         super().__init__()
         self.relu = nn.ReLU(inplace=inplace)
-        self.scale = nn.Parameter(scale_value * torch.ones(1), requires_grad=True)
-        self.bias = nn.Parameter(bias_value * torch.ones(1), requires_grad=True)
+        self.scale = nn.Parameter(scale_value * torch.ones(1))
+        self.bias = nn.Parameter(bias_value * torch.ones(1))
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:
         return self.scale * self.relu(x) ** 2 + self.bias

birder/net/mim/crossmae.py

@@ -113,7 +113,7 @@ class CrossMAE(MIMBaseNet):
             dim=decoder_embed_dim,
             num_special_tokens=0,
         ).unsqueeze(0)
-        self.decoder_pos_embed = nn.Parameter(pos_embedding, requires_grad=False)
+        self.decoder_pos_embed = nn.Buffer(pos_embedding)
 
         self.decoder_layers = nn.ModuleList()
         for _ in range(decoder_depth):

birder/net/mim/mae_vit.py

@@ -67,7 +67,7 @@ class MAE_ViT(MIMBaseNet):
                 dim=decoder_embed_dim,
                 num_special_tokens=self.encoder.num_special_tokens,
             )
-            self.decoder_pos_embed = nn.Parameter(pos_embedding, requires_grad=False)
+            self.decoder_pos_embed = nn.Buffer(pos_embedding)
 
         layers = []
         for _ in range(decoder_depth):

birder/net/mim/simmim.py

@@ -83,7 +83,7 @@ class SimMIM(MIMBaseNet):
             bias=True,
         )
 
-        self.mask_token = nn.Parameter(torch.zeros(1, 1, 1, self.encoder.stem_width), requires_grad=True)
+        self.mask_token = nn.Parameter(torch.zeros(1, 1, 1, self.encoder.stem_width))
 
         # Weights initialization
         nn.init.trunc_normal_(self.mask_token, mean=0.0, std=0.02)

birder/net/mobilenet_v1.py

@@ -107,15 +107,6 @@ class MobileNet_v1(DetectorBackbone):
 
         self.body = nn.Sequential(stages)
         self.features = nn.Sequential(
-            Conv2dNormActivation(
-                base * 32,
-                base * 32,
-                kernel_size=(1, 1),
-                stride=(1, 1),
-                padding=(0, 0),
-                bias=False,
-                activation_layer=None,
-            ),
             nn.AdaptiveAvgPool2d(output_size=(1, 1)),
             nn.Flatten(1),
         )

birder/net/mobilenet_v2.py

@@ -37,36 +37,44 @@ class InvertedResidual(nn.Module):
         num_expfilter = int(round(in_channels * expansion_factor))
 
         self.shortcut = shortcut
-        self.block = nn.Sequential(
-            Conv2dNormActivation(
-                in_channels,
-                num_expfilter,
-                kernel_size=(1, 1),
-                stride=(1, 1),
-                padding=(0, 0),
-                bias=False,
-                activation_layer=activation_layer,
-            ),
-            Conv2dNormActivation(
-                num_expfilter,
-                num_expfilter,
-                kernel_size=kernel_size,
-                stride=stride,
-                padding=padding,
-                groups=num_expfilter,
-                bias=False,
-                activation_layer=activation_layer,
-            ),
-            Conv2dNormActivation(
-                num_expfilter,
-                out_channels,
-                kernel_size=(1, 1),
-                stride=(1, 1),
-                padding=(0, 0),
-                bias=False,
-                activation_layer=None,
-            ),
+        layers = []
+        if expansion_factor != 1.0:
+            layers.append(
+                Conv2dNormActivation(
+                    in_channels,
+                    num_expfilter,
+                    kernel_size=(1, 1),
+                    stride=(1, 1),
+                    padding=(0, 0),
+                    bias=False,
+                    activation_layer=activation_layer,
+                )
+            )
+
+        layers.extend(
+            [
+                Conv2dNormActivation(
+                    num_expfilter,
+                    num_expfilter,
+                    kernel_size=kernel_size,
+                    stride=stride,
+                    padding=padding,
+                    groups=num_expfilter,
+                    bias=False,
+                    activation_layer=activation_layer,
+                ),
+                Conv2dNormActivation(
+                    num_expfilter,
+                    out_channels,
+                    kernel_size=(1, 1),
+                    stride=(1, 1),
+                    padding=(0, 0),
+                    bias=False,
+                    activation_layer=None,
+                ),
+            ]
         )
+        self.block = nn.Sequential(*layers)
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:
         if self.shortcut is True:
@@ -171,6 +179,7 @@ class MobileNet_v2(DetectorBackbone):
             ),
             nn.AdaptiveAvgPool2d(output_size=(1, 1)),
             nn.Flatten(1),
+            nn.Dropout(0.2),
         )
         self.return_channels = return_channels[1:5]
         self.embedding_size = last_channels
@@ -230,18 +239,3 @@ registry.register_model_config("mobilenet_v2_1_25", MobileNet_v2, config={"alpha
 registry.register_model_config("mobilenet_v2_1_5", MobileNet_v2, config={"alpha": 1.5})
 registry.register_model_config("mobilenet_v2_1_75", MobileNet_v2, config={"alpha": 1.75})
 registry.register_model_config("mobilenet_v2_2_0", MobileNet_v2, config={"alpha": 2.0})
-
-registry.register_weights(
-    "mobilenet_v2_1_0_il-common",
-    {
-        "description": "MobileNet v2 (1.0 multiplier) model trained on the il-common dataset",
-        "resolution": (256, 256),
-        "formats": {
-            "pt": {
-                "file_size": 10.6,
-                "sha256": "d6182293e98c102026f7cdc0d446aaf0e511232173c4b98c1a882c9f147be6e7",
-            }
-        },
-        "net": {"network": "mobilenet_v2_1_0", "tag": "il-common"},
-    },
-)

birder/net/{mobilenet_v3_large.py → mobilenet_v3.py}

@@ -3,6 +3,9 @@ MobileNet v3, adapted from
 https://github.com/pytorch/vision/blob/main/torchvision/models/mobilenetv3.py
 
 Paper "Searching for MobileNetV3", https://arxiv.org/abs/1905.02244
+
+Changes from original:
+* Using nn.BatchNorm2d with eps 1e-5 instead of 1e-3
 """
 
 # Reference license: BSD 3-Clause
@@ -113,7 +116,7 @@ class InvertedResidual(nn.Module):
 
 
 # pylint: disable=invalid-name
-class MobileNet_v3_Large(DetectorBackbone):
+class MobileNet_v3(DetectorBackbone):
     def __init__(
         self,
         input_channels: int,
@@ -121,12 +124,12 @@ class MobileNet_v3_Large(DetectorBackbone):
         *,
         config: Optional[dict[str, Any]] = None,
         size: Optional[tuple[int, int]] = None,
-        large: bool = True,
     ) -> None:
         super().__init__(input_channels, num_classes, config=config, size=size)
         assert self.config is not None, "must set config"
 
         alpha: float = self.config["alpha"]
+        large: bool = self.config["large"]
 
         if large is True:
             last_channels = int(round(1280 * max(1.0, alpha)))
@@ -268,15 +271,39 @@ class MobileNet_v3_Large(DetectorBackbone):
         )
 
 
-registry.register_model_config("mobilenet_v3_large_0_25", MobileNet_v3_Large, config={"alpha": 0.25})
-registry.register_model_config("mobilenet_v3_large_0_5", MobileNet_v3_Large, config={"alpha": 0.5})
-registry.register_model_config("mobilenet_v3_large_0_75", MobileNet_v3_Large, config={"alpha": 0.75})
-registry.register_model_config("mobilenet_v3_large_1_0", MobileNet_v3_Large, config={"alpha": 1.0})
-registry.register_model_config("mobilenet_v3_large_1_25", MobileNet_v3_Large, config={"alpha": 1.25})
-registry.register_model_config("mobilenet_v3_large_1_5", MobileNet_v3_Large, config={"alpha": 1.5})
-registry.register_model_config("mobilenet_v3_large_1_75", MobileNet_v3_Large, config={"alpha": 1.75})
-registry.register_model_config("mobilenet_v3_large_2_0", MobileNet_v3_Large, config={"alpha": 2.0})
+registry.register_model_config("mobilenet_v3_small_0_25", MobileNet_v3, config={"alpha": 0.25, "large": False})
+registry.register_model_config("mobilenet_v3_small_0_5", MobileNet_v3, config={"alpha": 0.5, "large": False})
+registry.register_model_config("mobilenet_v3_small_0_75", MobileNet_v3, config={"alpha": 0.75, "large": False})
+registry.register_model_config("mobilenet_v3_small_1_0", MobileNet_v3, config={"alpha": 1.0, "large": False})
+registry.register_model_config("mobilenet_v3_small_1_25", MobileNet_v3, config={"alpha": 1.25, "large": False})
+registry.register_model_config("mobilenet_v3_small_1_5", MobileNet_v3, config={"alpha": 1.5, "large": False})
+registry.register_model_config("mobilenet_v3_small_1_75", MobileNet_v3, config={"alpha": 1.75, "large": False})
+registry.register_model_config("mobilenet_v3_small_2_0", MobileNet_v3, config={"alpha": 2.0, "large": False})
+
+registry.register_model_config("mobilenet_v3_large_0_25", MobileNet_v3, config={"alpha": 0.25, "large": True})
+registry.register_model_config("mobilenet_v3_large_0_5", MobileNet_v3, config={"alpha": 0.5, "large": True})
+registry.register_model_config("mobilenet_v3_large_0_75", MobileNet_v3, config={"alpha": 0.75, "large": True})
+registry.register_model_config("mobilenet_v3_large_1_0", MobileNet_v3, config={"alpha": 1.0, "large": True})
+registry.register_model_config("mobilenet_v3_large_1_25", MobileNet_v3, config={"alpha": 1.25, "large": True})
+registry.register_model_config("mobilenet_v3_large_1_5", MobileNet_v3, config={"alpha": 1.5, "large": True})
+registry.register_model_config("mobilenet_v3_large_1_75", MobileNet_v3, config={"alpha": 1.75, "large": True})
+registry.register_model_config("mobilenet_v3_large_2_0", MobileNet_v3, config={"alpha": 2.0, "large": True})
+
 
+registry.register_weights(
+    "mobilenet_v3_small_1_0_il-common",
+    {
+        "description": "MobileNet v3 small (1.0 multiplier) model trained on the il-common dataset",
+        "resolution": (256, 256),
+        "formats": {
+            "pt": {
+                "file_size": 7.4,
+                "sha256": "ac53227f7513fd0c0b5204ee57403de2ab6c74c4e4d1061b9168596c6b5cea48",
+            }
+        },
+        "net": {"network": "mobilenet_v3_small_1_0", "tag": "il-common"},
+    },
+)
 registry.register_weights(
     "mobilenet_v3_large_0_75_il-common",
     {

birder/net/mobilevit_v1.py

@@ -1,11 +1,14 @@
 """
-MobileViT, adapted from
+MobileViT v1, adapted from
 https://github.com/huggingface/pytorch-image-models/blob/main/timm/models/mobilevit.py
 and
 https://github.com/lucidrains/vit-pytorch/blob/main/vit_pytorch/mobile_vit.py
 
 Paper "MobileViT: Light-weight, General-purpose, and Mobile-friendly Vision Transformer",
 https://arxiv.org/abs/2110.02178
+
+Changes from original:
+* Removed classifier bias
 """
 
 # Reference license: Apache-2.0 and MIT
@@ -63,6 +66,7 @@ class MobileVitBlock(nn.Module):
                     attention_dropout=attn_drop,
                     drop_path=drop_path_rate,
                     activation_layer=nn.SiLU,
+                    norm_layer_eps=1e-5,
                 )
                 for _ in range(transformer_depth)
             ]
@@ -166,7 +170,6 @@ class MobileViT_v1(BaseNet):
             stride=(2, 2),
             padding=(1, 1),
             activation_layer=nn.SiLU,
-            bias=True,
         )
 
         layers = []
@@ -231,7 +234,6 @@ class MobileViT_v1(BaseNet):
                 stride=(1, 1),
                 padding=(0, 0),
                 activation_layer=nn.SiLU,
-                bias=True,
             ),
             nn.AdaptiveAvgPool2d(output_size=(1, 1)),
             nn.Flatten(1),
@@ -290,32 +292,3 @@ registry.register_model_config(
         "expansion": 4,
     },
 )
-
-registry.register_weights(
-    "mobilevit_v1_xxs_il-common",
-    {
-        "description": "MobileViT v1 XXS model trained on the il-common dataset",
-        "resolution": (256, 256),
-        "formats": {
-            "pt": {
-                "file_size": 4.2,
-                "sha256": "2b565a768ca21fd72d5ef5090ff0f8b725f3e1165cd8e56749815041e5254d26",
-            }
-        },
-        "net": {"network": "mobilevit_v1_xxs", "tag": "il-common"},
-    },
-)
-registry.register_weights(
-    "mobilevit_v1_xs_il-common",
-    {
-        "description": "MobileViT v1 XS model trained on the il-common dataset",
-        "resolution": (256, 256),
-        "formats": {
-            "pt": {
-                "file_size": 8.1,
-                "sha256": "193bcede7f0b9f4574673e95c23c6ca3b8eeb30254a32a85e93342f1d67db31b",
-            }
-        },
-        "net": {"network": "mobilevit_v1_xs", "tag": "il-common"},
-    },
-)