scsc_swin.py

import torch
import torch.nn as nn
import torch.utils.checkpoint as checkpoint
import torch.nn.functional as F
from timm.models.layers import DropPath, to_2tuple, trunc_normal_

def channel_shuffle_v2(x, m):
    batchsize, num_channels, height, width = x.data.size()

    # reshape
    x = x.reshape(batchsize * num_channels // m,
        m, height * width)
    x = x.permute(1, 0, 2)
    # flatten
    x = x.reshape(m, batchsize, num_channels // m, height, width)

    xs = [x[i] for i in range(m)]

    return torch.cat(xs, dim=1)


class SELayer(nn.Module):
    def __init__(self, channel, reduction=16):
        super(SELayer, self).__init__()
        self.avg_pool = nn.AdaptiveAvgPool2d(1)
        self.channel = channel
        self.mid_channel = max(channel // reduction, 32)
        self.fc = nn.Sequential(
            nn.Linear(channel, self.mid_channel, bias=False),
            nn.ReLU(inplace=True),
            nn.Linear(self.mid_channel, channel, bias=False),
            nn.Sigmoid()
        )

    def forward(self, x):
        b, c, _, _ = x.size()
        y = self.avg_pool(x).view(b, c)
        y = self.fc(y).view(b, c, 1, 1)
        return x * y.expand_as(x)

    def flops(self, N):
        flops = 0
        # x = self.conv0(x)
        flops += N * self.channel * self.mid_channel *2

        return flops

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 DynamicDWConv(nn.Module):
    def __init__(self, dim, kernel_size, bias=True, stride=1, padding=1, groups=1, reduction=4):
        super().__init__()
        self.dim = dim
        self.kernel_size = kernel_size
        self.stride = stride
        self.padding = padding
        self.groups = groups

        self.pool = nn.AdaptiveAvgPool2d((1, 1))
        self.conv1 = nn.Conv2d(dim, dim // reduction, 1, bias=False)
        self.bn = nn.BatchNorm2d(dim // reduction)
        self.relu = nn.ReLU(inplace=True)
        self.conv2 = nn.Conv2d(dim // reduction, dim * kernel_size * kernel_size, 1)
        if bias:
            self.bias = nn.Parameter(torch.zeros(dim))
        else:
            self.bias = None

    def forward(self, x):
        b, c, h, w = x.shape
        weight = self.conv2(self.relu(self.bn(self.conv1(self.pool(x)))))
        weight = weight.view(b * self.dim, 1, self.kernel_size, self.kernel_size)
        x = F.conv2d(x.reshape(1, -1, h, w), weight, self.bias.repeat(b), stride=self.stride, padding=self.padding,
                     groups=b * self.groups)
        x = x.view(b, c, x.shape[-2], x.shape[-1])
        return x


class DWBlock(nn.Module):

    def __init__(self, dim, window_size, dynamic=False):
        super().__init__()
        self.dim = dim
        self.window_size = window_size  # Wh, Ww
        self.dynamic = dynamic

        # pw-linear
        self.conv0 = nn.Conv2d(dim, dim, 1, bias=False)
        self.bn0 = nn.BatchNorm2d(dim)

        if dynamic:
            self.conv = DynamicDWConv(dim, kernel_size=window_size, stride=1, padding=window_size // 2, groups=dim)
        else:
            self.conv = nn.Conv2d(dim, dim, kernel_size=window_size, stride=1, padding=window_size // 2, groups=dim)

        self.bn = nn.BatchNorm2d(dim)
        self.relu = nn.ReLU(inplace=True)

        # pw-linear
        self.conv2 = nn.Conv2d(dim, dim, 1, bias=False)
        self.bn2 = nn.BatchNorm2d(dim)

    def forward(self, x):
        B, H, W, C = x.shape

        x = x.permute(0, 3, 1, 2)
        x = self.conv0(x)
        x = self.bn0(x)
        x = self.relu(x)

        x = self.conv(x)
        x = self.bn(x)
        x = self.relu(x)

        x = self.conv2(x)
        x = self.bn2(x)

        x = x.permute(0, 2, 3, 1)
        return x

    def extra_repr(self) -> str:
        return f'dim={self.dim}, window_size={self.window_size}'

    def flops(self, N):
        # calculate flops for 1 window with token length of N
        flops = 0
        # x = self.conv0(x)
        flops += N * self.dim * self.dim
        # x = self.conv(x)
        if self.dynamic:
            flops += (
                        N * self.dim + self.dim * self.dim / 4 + self.dim / 4 * self.dim * self.window_size * self.window_size)
        flops += N * self.dim * self.window_size * self.window_size
        #  x = self.conv2(x)
        flops += N * self.dim * self.dim
        #  batchnorm + relu
        flops += 8 * self.dim * N
        return flops


class SCSC_block(nn.Module):

    def __init__(self, dim, window_sizes, mg):
        super().__init__()
        self.m = len(window_sizes)
        self.dim =dim
        self.window_sizes = window_sizes
        self.mg = mg
        self.chunk_planes = dim//self.m

        self.reduction_c = nn.Conv2d(dim, self.chunk_planes, kernel_size=1, bias=False)
        self.bn1 = nn.BatchNorm2d(self.chunk_planes)
        self.act1 = nn.GELU()

        self.layers = nn.ModuleList()
        for kernel_size in window_sizes:
            self.layers.append(nn.Conv2d(self.chunk_planes, self.chunk_planes, kernel_size, stride=1, padding=kernel_size//2, groups=self.chunk_planes, bias=False))
        self.bn2 = nn.BatchNorm2d(self.chunk_planes*self.m)
        self.act2 = nn.GELU()

        self.recover_c = nn.Conv2d(self.chunk_planes, dim, kernel_size=1, bias=False)
        self.bn3 = nn.BatchNorm2d(dim)
        #self.act3 = nn.GELU()

        self.routing_fc = nn.Conv2d(self.chunk_planes, self.m*self.mg, kernel_size=1, stride=1, bias=False)

        self.se = SELayer(dim)

    def forward(self, input):

        n, h, w, c = input.shape
        x = input.permute(0, 3, 1, 2)

        x = self.reduction_c(x)
        x = self.bn1(x)
        x_reduc = self.act1(x)

        tmp_xs = []
        for layer in self.layers:
            tmp_x = layer(x_reduc)
            tmp_xs.append(tmp_x)
        x = channel_shuffle_v2(torch.cat(tmp_xs, dim=1), self.m)

        x = self.bn2(x)
        x = self.act2(x)
        x = x.reshape(n, self.m, self.mg, self.chunk_planes//self.mg, h*w).permute(0,4,2,3,1).contiguous()

        routing_x = F.sigmoid(self.routing_fc(x_reduc))
        routing_x = routing_x.reshape(n, self.m, self.mg, h*w, 1).permute(0,3,2,1,4).contiguous()

        x = torch.matmul(x, routing_x)
        x = x.reshape(n, h, w, self.chunk_planes).permute(0,3,1,2).contiguous()

        x = self.recover_c(x)
        x = self.bn3(x)
        #x = self.act3(x)
        x = self.se(x)

        x = x.permute(0, 2, 3, 1)

        return x


    def flops(self, N):
        # calculate flops for 1 window with token length of N
        flops = 0
        # x = self.reduction_c(x)
        flops += N * self.dim * self.chunk_planes
        # x = self.conv(x)
        for window_size in self.window_sizes:
            flops += N * self.chunk_planes * window_size * window_size

        #  x = matmul(x)
        flops += N * self.mg * (self.chunk_planes//self.mg) * self.m

        #  x = self.recover_c(x)
        flops += N * self.chunk_planes * self.dim

        #  batchnorm + relu
        flops += 8 * self.dim * N

        #se
        flops +=  self.se.flops(N)

        return flops

class SpatialBlock(nn.Module):

    def __init__(self, dim, input_resolution, window_sizes, mg,
                 mlp_ratio=4., drop=0., drop_path=0., dynamic=False, act_layer=nn.GELU):
        super().__init__()
        self.dim = dim
        self.input_resolution = input_resolution
        self.window_sizes = window_sizes
        self.mlp_ratio = mlp_ratio
        self.dynamic = dynamic

        self.attn2conv = SCSC_block(dim, window_sizes, mg)

        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()

        mlp_hidden_dim = int(dim * mlp_ratio)
        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
        self.bn = nn.BatchNorm2d(dim)

    def forward(self, x):
        H, W = self.input_resolution
        B, L, C = x.shape
        assert L == H * W, "input feature has wrong size"

        shortcut = x
        x = self.attn2conv(x.view(B, H, W, C))
        x = x.view(B, H * W, C)

        # FFN
        x = shortcut + self.drop_path(x)
        shortcut = x
        x = self.bn(x.view(B, H, W, C).permute(0, 3, 1, 2)).permute(0, 2, 3, 1).view(B, H * W, C)
        x = shortcut + self.drop_path(self.mlp(x))
        return x

    def extra_repr(self) -> str:
        return f"dim={self.dim}, input_resolution={self.input_resolution} " \
               f"window_size={self.window_sizes}, mlp_ratio={self.mlp_ratio}"

    def flops(self):
        flops = 0
        H, W = self.input_resolution
        # batchnorm
        flops += 2 * self.dim * H * W
        # attn2conv
        flops += self.attn2conv.flops(H * W)
        # mlp
        flops += 2 * H * W * self.dim * self.dim * self.mlp_ratio
        return flops


class PatchMerging(nn.Module):

    def __init__(self, input_resolution, dim, norm_layer=nn.LayerNorm):
        super().__init__()
        self.input_resolution = input_resolution
        self.dim = dim
        self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False)
        self.norm = norm_layer(4 * dim)

    def forward(self, x):
        """
        x: B, H*W, C
        """
        H, W = self.input_resolution
        B, L, C = x.shape
        assert L == H * W, "input feature has wrong size"
        assert H % 2 == 0 and W % 2 == 0, f"x size ({H}*{W}) are not even."

        x = x.view(B, H, W, C)

        x0 = x[:, 0::2, 0::2, :]  # B H/2 W/2 C
        x1 = x[:, 1::2, 0::2, :]  # B H/2 W/2 C
        x2 = x[:, 0::2, 1::2, :]  # B H/2 W/2 C
        x3 = x[:, 1::2, 1::2, :]  # B H/2 W/2 C
        x = torch.cat([x0, x1, x2, x3], -1)  # B H/2 W/2 4*C
        x = x.view(B, -1, 4 * C)  # B H/2*W/2 4*C

        x = self.norm(x)
        x = self.reduction(x)

        return x

    def extra_repr(self) -> str:
        return f"input_resolution={self.input_resolution}, dim={self.dim}"

    def flops(self):
        H, W = self.input_resolution
        flops = H * W * self.dim
        flops += (H // 2) * (W // 2) * 4 * self.dim * 2 * self.dim
        return flops


class BasicLayer(nn.Module):

    def __init__(self, dim, input_resolution, depth, window_sizes, mg,
                 mlp_ratio=4., drop=0., drop_path=0., norm_layer=nn.LayerNorm,
                 downsample=None, use_checkpoint=False, dynamic=False):

        super().__init__()
        self.dim = dim
        self.input_resolution = input_resolution
        self.depth = depth
        self.use_checkpoint = use_checkpoint

        # build blocks
        self.blocks = nn.ModuleList([
            SpatialBlock(dim=dim, input_resolution=input_resolution,
                         window_sizes=window_sizes,
                         mg=mg,
                         mlp_ratio=mlp_ratio,
                         drop=drop,
                         dynamic=dynamic,
                         drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path)
            for i in range(depth)])

        # patch merging layer
        if downsample is not None:
            self.downsample = downsample(input_resolution, dim=dim, norm_layer=norm_layer)
        else:
            self.downsample = None

    def forward(self, x):
        for blk in self.blocks:
            if self.use_checkpoint:
                x = checkpoint.checkpoint(blk, x)
            else:
                x = blk(x)
        if self.downsample is not None:
            x = self.downsample(x)
        return x

    def extra_repr(self) -> str:
        return f"dim={self.dim}, input_resolution={self.input_resolution}, depth={self.depth}"

    def flops(self):
        flops = 0
        for blk in self.blocks:
            flops += blk.flops()
        if self.downsample is not None:
            flops += self.downsample.flops()
        return flops


class PatchEmbed(nn.Module):

    def __init__(self, img_size=224, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None):
        super().__init__()
        img_size = to_2tuple(img_size)
        patch_size = to_2tuple(patch_size)
        patches_resolution = [img_size[0] // patch_size[0], img_size[1] // patch_size[1]]
        self.img_size = img_size
        self.patch_size = patch_size
        self.patches_resolution = patches_resolution
        self.num_patches = patches_resolution[0] * patches_resolution[1]

        self.in_chans = in_chans
        self.embed_dim = embed_dim

        self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)
        if norm_layer is not None:
            self.norm = norm_layer(embed_dim)
        else:
            self.norm = None

    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)  # B Ph*Pw C
        if self.norm is not None:
            x = self.norm(x)
        return x

    def flops(self):
        Ho, Wo = self.patches_resolution
        flops = Ho * Wo * self.embed_dim * self.in_chans * (self.patch_size[0] * self.patch_size[1])
        if self.norm is not None:
            flops += Ho * Wo * self.embed_dim
        return flops

class SCSC_Swin(nn.Module):

    def __init__(self, img_size=224, patch_size=4, in_chans=3, num_classes=1000,
                 embed_dim=96, depths=[2, 2, 6, 2], window_size_l=[[3,9,13],[3,7,11],[3,5,7],[3,5]], mg=4, mlp_ratio=4.,
                 drop_rate=0., drop_path_rate=0.1, norm_layer=nn.LayerNorm,
                 ape=False, patch_norm=True, use_checkpoint=False, dynamic=False, **kwargs):
        super().__init__()

        self.num_classes = num_classes
        self.num_layers = len(depths)
        self.embed_dim = embed_dim
        self.ape = ape
        self.patch_norm = patch_norm
        self.num_features = int(embed_dim * 2 ** (self.num_layers - 1))
        self.mlp_ratio = mlp_ratio

        # split image into non-overlapping patches
        self.patch_embed = PatchEmbed(
            img_size=img_size, patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim,
            norm_layer=norm_layer if self.patch_norm else None)
        num_patches = self.patch_embed.num_patches
        patches_resolution = self.patch_embed.patches_resolution
        self.patches_resolution = patches_resolution

        # absolute position embedding
        if self.ape:
            self.absolute_pos_embed = nn.Parameter(torch.zeros(1, num_patches, embed_dim))
            trunc_normal_(self.absolute_pos_embed, std=.02)

        self.pos_drop = nn.Dropout(p=drop_rate)

        # stochastic depth
        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))]  # stochastic depth decay rule

        # build layers
        self.layers = nn.ModuleList()
        assert self.num_layers==len(window_size_l)
        for i_layer, window_sizes in enumerate(window_size_l):
            layer = BasicLayer(dim=int(embed_dim * 2 ** i_layer),
                               input_resolution=(patches_resolution[0] // (2 ** i_layer),
                                                 patches_resolution[1] // (2 ** i_layer)),
                               depth=depths[i_layer],
                               window_sizes=window_sizes,
                               mg=mg,
                               mlp_ratio=self.mlp_ratio,
                               drop=drop_rate,
                               drop_path=dpr[sum(depths[:i_layer]):sum(depths[:i_layer + 1])],
                               norm_layer=norm_layer,
                               downsample=PatchMerging if (i_layer < self.num_layers - 1) else None,
                               use_checkpoint=use_checkpoint,
                               dynamic=dynamic)
            self.layers.append(layer)

        self.norm = norm_layer(self.num_features)
        self.avgpool = nn.AdaptiveAvgPool1d(1)
        self.head = nn.Linear(self.num_features, num_classes) if num_classes > 0 else nn.Identity()

        self.apply(self._init_weights)

    def _init_weights(self, m):
        if isinstance(m, nn.Linear):
            trunc_normal_(m.weight, std=.02)
            if isinstance(m, nn.Linear) and m.bias is not None:
                nn.init.constant_(m.bias, 0)
        elif isinstance(m, nn.LayerNorm):
            nn.init.constant_(m.bias, 0)
            nn.init.constant_(m.weight, 1.0)

    @torch.jit.ignore
    def no_weight_decay(self):
        return {'absolute_pos_embed'}

    @torch.jit.ignore
    def no_weight_decay_keywords(self):
        return {}

    def forward_features(self, x):
        x = self.patch_embed(x)
        if self.ape:
            x = x + self.absolute_pos_embed
        x = self.pos_drop(x)

        for layer in self.layers:
            x = layer(x)

        x = self.norm(x)  # B L C
        x = self.avgpool(x.transpose(1, 2))  # B C 1
        x = torch.flatten(x, 1)
        return x

    def forward(self, x):
        x = self.forward_features(x)
        x = self.head(x)
        return x

    def flops(self):
        flops = 0
        flops += self.patch_embed.flops()
        for i, layer in enumerate(self.layers):
            flops += layer.flops()
        flops += self.num_features * self.patches_resolution[0] * self.patches_resolution[1] // (2 ** self.num_layers)
        flops += self.num_features * self.num_classes
        return flops