突破计算瓶颈：GFNet在图像分类、目标检测与语义分割中的高效创新应用

import torch
from torch import nn
import math
from timm.models.layers import DropPath, to_2tuple

# 论文地址：https://arxiv.org/pdf/2107.00645
# 论文：Global Filter Networks for Image Classification


class PatchEmbed(nn.Module):
    """ Image to Patch Embedding
    """
    def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768):
        super().__init__()
        img_size = to_2tuple(img_size)
        patch_size = to_2tuple(patch_size)
        num_patches = (img_size[1] // patch_size[1]) * (img_size[0] // patch_size[0])
        self.img_size = img_size
        self.patch_size = patch_size
        self.num_patches = num_patches
        self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)

    def forward(self, x):
        B, C, H, W = x.shape
        # FIXME look at relaxing size constraints
        assert H == self.img_size[0] and W == self.img_size[1], \
            f"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})."
        x = self.proj(x).flatten(2).transpose(1, 2)
        return x

class GlobalFilter(nn.Module):
    def __init__(self, dim, h=14, w=8):
        super().__init__()
        self.complex_weight = nn.Parameter(torch.randn(h, w, dim, 2, dtype=torch.float32) * 0.02)
        self.w = w
        self.h = h

    def forward(self, x, spatial_size=None):
        B, N, C = x.shape
        if spatial_size is None:
            a = b = int(math.sqrt(N))
        else:
            a, b = spatial_size

        x = x.view(B, a, b, C)

        x = x.to(torch.float32)

        x = torch.fft.rfft2(x, dim=(1, 2), norm='ortho')
        weight = torch.view_as_complex(self.complex_weight)
        x = x * weight
        x = torch.fft.irfft2(x, s=(a, b), dim=(1, 2), norm='ortho')

        x = x.reshape(B, N, C)
        return x

class Mlp(nn.Module):
    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
        super().__init__()
        out_features = out_features or in_features
        hidden_features = hidden_features or in_features
        self.fc1 = nn.Linear(in_features, hidden_features)
        self.act = act_layer()
        self.fc2 = nn.Linear(hidden_features, out_features)
        self.drop = nn.Dropout(drop)

    def forward(self, x):
        x = self.fc1(x)
        x = self.act(x)
        x = self.drop(x)
        x = self.fc2(x)
        x = self.drop(x)
        return x

class Block(nn.Module):
    def __init__(self, dim, mlp_ratio=4., drop=0., drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, h=14, w=8):
        super().__init__()
        self.norm1 = norm_layer(dim)
        self.filter = GlobalFilter(dim, h=h, w=w)
        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
        self.norm2 = norm_layer(dim)
        mlp_hidden_dim = int(dim * mlp_ratio)
        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)

    def forward(self, x):
        x = x + self.drop_path(self.mlp(self.norm2(self.filter(self.norm1(x)))))
        return x


class GFNet(nn.Module):
    def __init__(self, embed_dim=384, img_size=224, patch_size=16, mlp_ratio=4, depth=4, num_classes=1000):
        super().__init__()
        self.patch_embed = PatchEmbed(img_size=img_size, patch_size=patch_size, embed_dim=embed_dim)
        self.embedding = nn.Linear((patch_size ** 2) * 3, embed_dim)

        h = img_size // patch_size
        w = h // 2 + 1


        self.blocks = nn.ModuleList([
            Block(dim=embed_dim, mlp_ratio=mlp_ratio, h=h, w=w)
            for i in range(depth)
        ])

        self.head = nn.Linear(embed_dim, num_classes)
        self.softmax = nn.Softmax(1)

    def forward(self, x):
        x = self.patch_embed(x)
        for blk in self.blocks:
            x = blk(x)
        x = x.mean(dim=1)
        x = self.softmax(self.head(x))
        return x

if __name__ == '__main__':
    input = torch.randn(1, 3, 224, 224)
    block = GFNet(embed_dim=384, img_size=224, patch_size=16, num_classes=1000)
    out = block(input)    print(out.shape)