Add ConvNeXt (#17)

tests/test_convnext.py

@@ -0,0 +1,20 @@
+import timm
+import torch
+
+from vision_toolbox.backbones import ConvNeXt
+
+
+def test_forward():
+    m = ConvNeXt.from_config("T")
+    m(torch.randn(1, 3, 224, 224))
+
+
+def test_from_pretrained():
+    m = ConvNeXt.from_config("T", True).eval()
+    x = torch.randn(1, 3, 224, 224)
+    out = m(x)
+
+    m_timm = timm.create_model("convnext_tiny.fb_in22k", pretrained=True, num_classes=0).eval()
+    out_timm = m_timm(x)
+
+    torch.testing.assert_close(out, out_timm)

vision_toolbox/backbones/__init__.py

@@ -1,3 +1,4 @@
+from .convnext import ConvNeXt
 from .darknet import Darknet, DarknetYOLOv5
 from .mlp_mixer import MLPMixer
 from .patchconvnet import PatchConvNet

vision_toolbox/backbones/convnext.py

@@ -0,0 +1,146 @@
+# https://arxiv.org/abs/2201.03545
+# https://github.com/facebookresearch/ConvNeXt
+
+from __future__ import annotations
+
+from functools import partial
+
+import torch
+from torch import Tensor, nn
+
+from ..components import Permute, StochasticDepth
+from .base import BaseBackbone, _act, _norm
+
+
+class ConvNeXtBlock(nn.Module):
+    def __init__(
+        self,
+        d_model: int,
+        expansion_ratio: float = 4.0,
+        bias: bool = True,
+        layer_scale_init: float = 1e-6,
+        stochastic_depth: float = 0.0,
+        norm: _norm = partial(nn.LayerNorm, eps=1e-6),
+        act: _act = nn.GELU,
+    ) -> None:
+        super().__init__()
+        hidden_dim = int(d_model * expansion_ratio)
+        self.layers = nn.Sequential(
+            Permute(0, 3, 1, 2),
+            nn.Conv2d(d_model, d_model, 7, padding=3, groups=d_model, bias=bias),
+            Permute(0, 2, 3, 1),
+            norm(d_model),
+            nn.Linear(d_model, hidden_dim, bias=bias),
+            act(),
+            nn.Linear(hidden_dim, d_model, bias=bias),
+        )
+        self.layer_scale = nn.Parameter(torch.full((d_model,), layer_scale_init)) if layer_scale_init > 0 else None
+        self.drop = StochasticDepth(stochastic_depth)
+
+    def forward(self, x: Tensor) -> Tensor:
+        return x + self.drop(self.layers(x) * self.layer_scale)
+
+
+class ConvNeXt(BaseBackbone):
+    def __init__(
+        self,
+        d_model: int,
+        depths: tuple[int, ...],
+        expansion_ratio: float = 4.0,
+        bias: bool = True,
+        layer_scale_init: float = 1e-6,
+        stochastic_depth: float = 0.0,
+        norm: _norm = partial(nn.LayerNorm, eps=1e-6),
+        act: _act = nn.GELU,
+    ) -> None:
+        super().__init__()
+        self.stem = nn.Sequential(nn.Conv2d(3, d_model, 4, 4), Permute(0, 2, 3, 1), norm(d_model))
+
+        stochastic_depth_rates = torch.linspace(0, stochastic_depth, sum(depths))
+        self.stages = nn.Sequential()
+
+        for stage_idx, depth in enumerate(depths):
+            stage = nn.Sequential()
+            if stage_idx > 0:
+                # equivalent to PatchMerging in SwinTransformer
+                downsample = nn.Sequential(
+                    norm(d_model),
+                    Permute(0, 3, 1, 2),
+                    nn.Conv2d(d_model, d_model * 2, 2, 2),
+                    Permute(0, 2, 3, 1),
+                )
+                d_model *= 2
+            else:
+                downsample = nn.Identity()
+            stage.append(downsample)
+
+            for block_idx in range(depth):
+                rate = stochastic_depth_rates[sum(depths[:stage_idx]) + block_idx]
+                block = ConvNeXtBlock(d_model, expansion_ratio, bias, layer_scale_init, rate, norm, act)
+                stage.append(block)
+
+            self.stages.append(stage)
+
+        self.head_norm = norm(d_model)
+
+    def get_feature_maps(self, x: Tensor) -> list[Tensor]:
+        out = [self.stem(x)]
+        for stage in self.stages:
+            out.append(stage(out[-1]))
+        return out[-1:]
+
+    def forward(self, x: Tensor) -> Tensor:
+        return self.head_norm(self.get_feature_maps(x)[-1].mean((1, 2)))
+
+    @staticmethod
+    def from_config(variant: str, pretrained: bool = False) -> ConvNeXt:
+        d_model, depths = dict(
+            T=(96, (3, 3, 9, 3)),
+            S=(96, (3, 3, 27, 3)),
+            B=(128, (3, 3, 27, 3)),
+            L=(192, (3, 3, 27, 3)),
+            XL=(256, (3, 3, 27, 3)),
+        )[variant]
+        m = ConvNeXt(d_model, depths)
+
+        if pretrained:
+            # TODO: also add torchvision checkpoints?
+            ckpt = dict(
+                T="convnext_tiny_22k_224.pth",
+                S="convnext_small_22k_224.pth",
+                B="convnext_base_22k_224.pth",
+                L="convnext_large_22k_224.pth",
+                XL="convnext_xlarge_22k_224.pth",
+            )[variant]
+            base_url = "https://dl.fbaipublicfiles.com/convnext/"
+            state_dict = torch.hub.load_state_dict_from_url(base_url + ckpt)["model"]
+            m.load_official_ckpt(state_dict)
+
+        return m
+
+    @torch.no_grad()
+    def load_official_ckpt(self, state_dict: dict[str, Tensor]) -> None:
+        def copy_(m: nn.Conv2d | nn.Linear | nn.LayerNorm, prefix: str):
+            m.weight.copy_(state_dict.pop(prefix + ".weight"))
+            m.bias.copy_(state_dict.pop(prefix + ".bias"))
+
+        copy_(self.stem[0], "downsample_layers.0.0")  # Conv2d
+        copy_(self.stem[2], "downsample_layers.0.1")  # LayerNorm
+
+        for stage_idx, stage in enumerate(self.stages):
+            if stage_idx > 0:
+                copy_(stage[0][0], f"downsample_layers.{stage_idx}.0")  # LayerNorm
+                copy_(stage[0][2], f"downsample_layers.{stage_idx}.1")  # Conv2d
+
+            for block_idx in range(1, len(stage)):
+                block: ConvNeXtBlock = stage[block_idx]
+                prefix = f"stages.{stage_idx}.{block_idx - 1}."
+
+                copy_(block.layers[1], prefix + "dwconv")
+                copy_(block.layers[3], prefix + "norm")
+                copy_(block.layers[4], prefix + "pwconv1")
+                copy_(block.layers[6], prefix + "pwconv2")
+                block.layer_scale.copy_(state_dict.pop(prefix + "gamma"))
+
+        copy_(self.head_norm, "norm")
+        assert len(state_dict) == 2

vision_toolbox/components.py

@@ -150,3 +150,31 @@ def forward(self, x: Tensor) -> Tensor:
             x = self.pool(x)
             outputs.append(x)
         return torch.cat(outputs, dim=1)
+
+
+class Permute(nn.Module):
+    def __init__(self, *dims: int) -> None:
+        super().__init__()
+        self.dims = dims
+
+    def forward(self, x: Tensor) -> Tensor:
+        return x.permute(self.dims)
+
+
+# https://arxiv.org/pdf/1603.09382.pdf
+class StochasticDepth(nn.Module):
+    def __init__(self, p: float) -> None:
+        assert 0.0 <= p <= 1.0
+        super().__init__()
+        self.p = p
+
+    def forward(self, x: Tensor) -> Tensor:
+        if not self.training or self.p == 0.0:
+            return x
+
+        shape = [x.shape[0]] + [1] * (x.ndim - 1)
+        keep_p = 1.0 - self.p
+        return x * x.new_empty(shape).bernoulli_(keep_p).div_(keep_p)
+
+    def extra_repr(self) -> str:
+        return f"p={self.p}"