swim_transformer的pytorch实现

#swin_transformer.py
import math
import torch
import torch.nn as nn
from torch import _assert
from timm.models.layers import DropPath, to_2tuple, to_ntuple, trunc_normal_
from utils.helpers import checkpoint_seq

class Mlp(nn.Module):
    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, bias=True, drop=0.):
        super().__init__()
        out_features = out_features or in_features
        hidden_features = hidden_features or in_features
        bias = to_2tuple(bias)
        drop_probs = to_2tuple(drop)

        self.fc1 = nn.Linear(in_features, hidden_features, bias=bias[0])
        self.act = act_layer()
        self.drop1 = nn.Dropout(drop_probs[0])
        self.fc2 = nn.Linear(hidden_features, out_features, bias=bias[1])
        self.drop2 = nn.Dropout(drop_probs[1])

    def forward(self, x):
        x = self.fc1(x)
        x = self.act(x)
        x = self.drop1(x)
        x = self.fc2(x)
        x = self.drop2(x)
        return x

#patch嵌入
class PatchEmbed(nn.Module):
    """ 2D Image to Patch Embedding """
    def __init__(self, img_size=224, patch_size=4, in_chans=3, embed_dim=96, norm_layer=nn.LayerNorm, flatten=True):
        super().__init__()
        img_size = to_2tuple(img_size)
        patch_size = to_2tuple(patch_size)
        self.img_size = img_size
        self.patch_size = patch_size
        self.grid_size = (img_size[0] // patch_size[0], img_size[1] // patch_size[1]) #224/4=56
        self.num_patches = self.grid_size[0] * self.grid_size[1] #56*56
        self.flatten = flatten

        self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)
        self.norm = norm_layer(embed_dim) if norm_layer else nn.Identity()

    def forward(self, x):
        x = self.proj(x)
        if self.flatten:
            x = x.flatten(2).transpose(1, 2)  # BCHW -> BC(H*W) -> B(H*W)C
        x = self.norm(x)
        return x

#patch聚合 尺寸减半 通道加倍
class PatchMerging(nn.Module):
    r""" Patch Merging Layer.
    Args:
        input_resolution (tuple[int]): Resolution of input feature.
        dim (int): Number of input channels.
        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
    """

    def __init__(self, input_resolution, dim, out_dim=None, norm_layer=nn.LayerNorm):
        super().__init__()
        self.input_resolution = input_resolution
        self.dim = dim
        self.out_dim = out_dim or 2 * dim
        self.norm = norm_layer(4 * dim)
        self.reduction = nn.Linear(4 * dim, self.out_dim, bias=False)

    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  0::2 表示从0开始到最后 间隔为1 取元素
        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  在通道维度拼接 通道数量翻了4倍
        x = x.view(B, -1, 4 * C)  # B H/2*W/2 4*C

        x = self.norm(x)
        x = self.reduction(x)  #通过线性层降维 B H/2*W/2 2*C

        return x

#将一个layer分成若干个windows，然后在每个windows内attention计算
def window_partition(x, window_size):  # window_size = 7
    """
    Args:
        x: (B, H, W, C)
        window_size (int): window size
    Returns:
        windows: (num_windows*B, window_size, window_size, C)
    """
    B, H, W, C = x.shape #torch.Size([1, 56, 56, 96])
    x = x.view(B, H // window_size, window_size, W // window_size, window_size, C) #torch.Size([1, 8, 7, 8, 7, 96])
    windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C) #torch.Size([64, 7, 7, 96])
    return windows

#将若干个windows合并为一个layer
def window_reverse(windows, window_size, H, W):
    """
    Args:
        windows: (num_windows*B, window_size, window_size, C)     num_windows = 56*56/(7*7)
        window_size (int): Window size
        H (int): Height of image
        W (int): Width of image
    Returns:
        x: (B, H, W, C)
    """
    B = int(windows.shape[0] / (H * W / window_size / window_size))
    x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1)
    x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
    return x

#获得相对位置编码  未看
def get_relative_position_index(win_h, win_w):
    # get pair-wise relative position index for each token inside the window
    coords = torch.stack(torch.meshgrid([torch.arange(win_h), torch.arange(win_w)]))  # 2, Wh, Ww
    coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
    relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]  # 2, Wh*Ww, Wh*Ww
    relative_coords = relative_coords.permute(1, 2, 0).contiguous()  # Wh*Ww, Wh*Ww, 2
    relative_coords[:, :, 0] += win_h - 1  # shift to start from 0
    relative_coords[:, :, 1] += win_w - 1
    relative_coords[:, :, 0] *= 2 * win_w - 1
    return relative_coords.sum(-1)  # Wh*Ww, Wh*Ww

#窗口注意力
class WindowAttention(nn.Module):
    r""" Window based multi-head self attention (W-MSA) module with relative position bias.
    It supports both of shifted and non-shifted window.
    Args:
        dim (int): Number of input channels.
        num_heads (int): Number of attention heads. NUM_HEADS: [ 3, 6, 12, 24 ]
        head_dim (int): Number of channels per head (dim // num_heads if not set)
        window_size (tuple[int]): The height and width of the window.
        qkv_bias (bool, optional):  If True, add a learnable bias to query, key, value. Default: True
        attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0
        proj_drop (float, optional): Dropout ratio of output. Default: 0.0
    """

    def __init__(self, dim, num_heads, head_dim=None, window_size=7, qkv_bias=True, attn_drop=0., proj_drop=0.):

        super().__init__()
        self.dim = dim
        self.window_size = to_2tuple(window_size)  # Wh, Ww
        win_h, win_w = self.window_size
        self.window_area = win_h * win_w
        self.num_heads = num_heads
        head_dim = head_dim or dim // num_heads
        attn_dim = head_dim * num_heads
        self.scale = head_dim ** -0.5

        # define a parameter table of relative position bias, shape: 2*Wh-1 * 2*Ww-1, nH
        self.relative_position_bias_table = nn.Parameter(torch.zeros((2 * win_h - 1) * (2 * win_w - 1), num_heads))

        # get pair-wise relative position index for each token inside the window
        self.register_buffer("relative_position_index", get_relative_position_index(win_h, win_w))

        self.qkv = nn.Linear(dim, attn_dim * 3, bias=qkv_bias)
        self.attn_drop = nn.Dropout(attn_drop)
        self.proj = nn.Linear(attn_dim, dim)
        self.proj_drop = nn.Dropout(proj_drop)

        trunc_normal_(self.relative_position_bias_table, std=.02)
        self.softmax = nn.Softmax(dim=-1)

    def _get_rel_pos_bias(self) -> torch.Tensor:
        relative_position_bias = self.relative_position_bias_table[
            self.relative_position_index.view(-1)].view(self.window_area, self.window_area, -1)  # Wh*Ww,Wh*Ww,nH
        relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()  # nH, Wh*Ww, Wh*Ww
        return relative_position_bias.unsqueeze(0)

    def forward(self, x, mask = None):
        """
        Args:
            x: input features with shape of (num_windows*B, N, C) = torch.Size([64, 7*7, 96])
            mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None
        """
        B_, N, C = x.shape
        qkv = self.qkv(x).reshape(B_, N, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)   #[64,7*7,96]->[64,7*7,96*3]->[64,7*7,3,12,8]->[3,64,12,49,8]
        q, k, v = qkv.unbind(0)  # make torchscript happy (cannot use tensor as tuple)

        q = q * self.scale
        attn = (q @ k.transpose(-2, -1))
        attn = attn + self._get_rel_pos_bias()  #相对位置编码是在注意力的时候直接加上去

        if mask is not None: #移动窗口  需要mask
            num_win = mask.shape[0]
            attn = attn.view(B_ // num_win, num_win, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0)
            attn = attn.view(-1, self.num_heads, N, N)  #不想要的区域mask设置负数 比如-100
            attn = self.softmax(attn)  #与注意力相加之后  做softmax就会等于0  从而控制不需要做自注意力的区域
        else:
            attn = self.softmax(attn)

        attn = self.attn_drop(attn)

        x = (attn @ v).transpose(1, 2).reshape(B_, N, -1)
        x = self.proj(x) #通道不变  做一次线性投影
        x = self.proj_drop(x)
        return x

#单个swimtransformer的block  包括窗口的多头注意力和移动窗口的多头注意力
class SwinTransformerBlock(nn.Module):
    r""" Swin Transformer Block.
    Args:
        dim (int): Number of input channels. 输入通道
        input_resolution (tuple[int]): Input resulotion. 输入分辨率
        window_size (int): Window size. 窗口大小
        num_heads (int): Number of attention heads. 多头的数量
        head_dim (int): Enforce the number of channels per head  每个头的通道数
        shift_size (int): Shift size for SW-MSA.  #移动几个patchs
        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.  #一般是两层FC  先将通道数放大4倍  再变回原通道数
        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
        drop (float, optional): Dropout rate. Default: 0.0
        attn_drop (float, optional): Attention dropout rate. Default: 0.0
        drop_path (float, optional): Stochastic depth rate. Default: 0.0
        act_layer (nn.Module, optional): Activation layer. Default: nn.GELU
        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
    """

    def __init__(
            self, dim, input_resolution, num_heads, head_dim=None, window_size=7, shift_size=0,
            mlp_ratio=4., qkv_bias=True, drop=0., attn_drop=0., drop_path=0.,
            act_layer=nn.GELU, norm_layer=nn.LayerNorm):
        super().__init__()
        self.dim = dim
        self.input_resolution = input_resolution
        self.window_size = window_size
        self.shift_size = shift_size
        self.mlp_ratio = mlp_ratio
        if min(self.input_resolution) <= self.window_size:
            # if window size is larger than input resolution, we don't partition windows
            self.shift_size = 0
            self.window_size = min(self.input_resolution)
        assert 0 <= self.shift_size < self.window_size, "shift_size must in 0-window_size"

        self.norm1 = norm_layer(dim)
        self.attn = WindowAttention(
            dim, num_heads=num_heads, head_dim=head_dim, window_size=to_2tuple(self.window_size),
            qkv_bias=qkv_bias, attn_drop=attn_drop, proj_drop=drop)

        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity() #残差
        self.norm2 = norm_layer(dim)
        self.mlp = Mlp(in_features=dim, hidden_features=int(dim * mlp_ratio), act_layer=act_layer, drop=drop)

        if self.shift_size > 0:
            # calculate attention mask for SW-MSA
            H, W = self.input_resolution
            img_mask = torch.zeros((1, H, W, 1))  # 1 H W 1
            cnt = 0
            for h in (
                    slice(0, -self.window_size),
                    slice(-self.window_size, -self.shift_size),
                    slice(-self.shift_size, None)):
                for w in (
                        slice(0, -self.window_size),
                        slice(-self.window_size, -self.shift_size),
                        slice(-self.shift_size, None)):
                    img_mask[:, h, w, :] = cnt
                    cnt += 1
            mask_windows = window_partition(img_mask, self.window_size)  # num_win, window_size, window_size, 1
            mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
            attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
            attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0))
        else:
            attn_mask = None

        self.register_buffer("attn_mask", attn_mask)

    def forward(self, x):
        H, W = self.input_resolution  #[56,56]>>[28,28]>>[14,14]>>[7,7]
        B, L, C = x.shape
        _assert(L == H * W, "input feature has wrong size")

        shortcut = x
        x = self.norm1(x)
        x = x.view(B, H, W, C)

        # cyclic shift   大于0再移位  在dims(1,2)维度上移位shift_size
        if self.shift_size > 0:
            shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
        else:
            shifted_x = x

        # partition windows
        x_windows = window_partition(shifted_x, self.window_size)  # num_win*B, window_size, window_size, C
        x_windows = x_windows.view(-1, self.window_size * self.window_size, C)  # num_win*B, window_size*window_size, C

        # W-MSA/SW-MSA
        attn_windows = self.attn(x_windows, mask=self.attn_mask)  # num_win*B, window_size*window_size, C

        # merge windows
        attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
        shifted_x = window_reverse(attn_windows, self.window_size, H, W)  # B H' W' C

        # reverse cyclic shift
        if self.shift_size > 0:
            x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
        else:
            x = shifted_x
        x = x.view(B, H * W, C)

        # FFN
        x = shortcut + self.drop_path(x)
        x = x + self.drop_path(self.mlp(self.norm2(x)))

        return x  #x = [B L C]

#通过{2,2,6,2} 构造stage
class BasicLayer(nn.Module):
    """ A basic Swin Transformer layer for one stage.
    Args:
        dim (int): Number of input channels.
        input_resolution (tuple[int]): Input resolution.
        depth (int): Number of blocks.    block的数量
        num_heads (int): Number of attention heads.
        head_dim (int): Channels per head (dim // num_heads if not set)
        window_size (int): Local window size.
        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
        drop (float, optional): Dropout rate. Default: 0.0
        attn_drop (float, optional): Attention dropout rate. Default: 0.0
        drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0
        norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
        downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None Patch merging
    """

    def __init__(
            self, dim, out_dim, input_resolution, depth, num_heads, head_dim=None,
            window_size=7, mlp_ratio=4., qkv_bias=True, drop=0., attn_drop=0.,
            drop_path=0., norm_layer=nn.LayerNorm, downsample=None):

        super().__init__()
        self.dim = dim
        self.input_resolution = input_resolution
        self.depth = depth
        self.grad_checkpointing = False

        # build blocks #shift_size=0 if (i % 2 == 0) else window_size // 2 表示偶数使用移动窗口
        self.blocks = nn.Sequential(*[
            SwinTransformerBlock(
                dim=dim, input_resolution=input_resolution, num_heads=num_heads, head_dim=head_dim,
                window_size=window_size, shift_size=0 if (i % 2 == 0) else window_size // 2,
                mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, drop=drop, attn_drop=attn_drop,
                drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path, norm_layer=norm_layer)
            for i in range(depth)])

        # patch merging layer  使用patch merging来减半尺寸 翻倍通道
        if downsample is not None:
            self.downsample = downsample(input_resolution, dim=dim, out_dim=out_dim, norm_layer=norm_layer)
        else:
            self.downsample = None

    def forward(self, x):
        if self.grad_checkpointing and not torch.jit.is_scripting():
            x = checkpoint_seq(self.blocks, x)
        else:
            x = self.blocks(x)
        if self.downsample is not None:
            x = self.downsample(x)
        return x

class SwinTransformer(nn.Module):
    r""" Swin Transformer
        A PyTorch impl of : `Swin Transformer: Hierarchical Vision Transformer using Shifted Windows`  -
          https://arxiv.org/pdf/2103.14030
    Args:
        img_size (int | tuple(int)): Input image size. Default 224
        patch_size (int | tuple(int)): Patch size. Default: 4
        in_chans (int): Number of input image channels. Default: 3
        num_classes (int): Number of classes for classification head. Default: 1000
        embed_dim (int): Patch embedding dimension. Default: 96
        depths (tuple(int)): Depth of each Swin Transformer layer.
        num_heads (tuple(int)): Number of attention heads in different layers.
        head_dim (int, tuple(int)):每个头的通道数
        window_size (int): Window size. Default: 7
        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4
        qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True
        drop_rate (float): Dropout rate. Default: 0
        attn_drop_rate (float): Attention dropout rate. Default: 0
        drop_path_rate (float): Stochastic depth rate. Default: 0.1
        norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm.
        ape (bool): If True, add absolute position embedding to the patch embedding. Default: False
        patch_norm (bool): If True, add normalization after patch embedding. Default: True
    """

    def __init__(
            self, img_size=224, patch_size=4, in_chans=3, num_classes=1000, global_pool='avg',
            embed_dim=96, depths=(2, 2, 6, 2), num_heads=(3, 6, 12, 24), head_dim=None,
            window_size=7, mlp_ratio=4., qkv_bias=True,
            drop_rate=0., attn_drop_rate=0., drop_path_rate=0.1,
            norm_layer=nn.LayerNorm, ape=False, patch_norm=True, out_indices=(0, 1, 2, 3), frozen_stages=-1, **kwargs):
        super().__init__()
        assert global_pool in ('', 'avg')
        self.num_classes = num_classes
        self.global_pool = global_pool
        self.num_layers = len(depths) #层数  也就是stage数
        self.embed_dim = embed_dim
        self.num_features = int(embed_dim * 2 ** (self.num_layers - 1))
        self.frozen_stages = frozen_stages

        # 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 patch_norm else None)
        num_patches = self.patch_embed.num_patches  #56*56
        self.patch_grid = self.patch_embed.grid_size #[56,56]

        # absolute position embedding   if ape为Ture 使用绝对位置编码
        self.absolute_pos_embed = nn.Parameter(torch.zeros(1, num_patches, embed_dim)) if ape else None
        self.pos_drop = nn.Dropout(p=drop_rate)

        # build layers
        if not isinstance(embed_dim, (tuple, list)):
            embed_dim = [int(embed_dim * 2 ** i) for i in range(self.num_layers)]
        embed_out_dim = embed_dim[1:] + [None]
        head_dim = to_ntuple(self.num_layers)(head_dim)
        window_size = to_ntuple(self.num_layers)(window_size)
        mlp_ratio = to_ntuple(self.num_layers)(mlp_ratio)
        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))]  # stochastic depth decay rule
        layers = []
        for i in range(self.num_layers):
            layers += [BasicLayer(
                dim=embed_dim[i],
                out_dim=embed_out_dim[i],
                input_resolution=(self.patch_grid[0] // (2 ** i), self.patch_grid[1] // (2 ** i)),
                depth=depths[i],
                num_heads=num_heads[i],
                head_dim=head_dim[i],
                window_size=window_size[i],
                mlp_ratio=mlp_ratio[i],
                qkv_bias=qkv_bias,
                drop=drop_rate,
                attn_drop=attn_drop_rate,
                drop_path=dpr[sum(depths[:i]):sum(depths[:i + 1])],
                norm_layer=norm_layer,
                downsample=PatchMerging if (i < self.num_layers - 1) else None
            )]
        self.layers = nn.Sequential(*layers)

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


        self.apply(self._init_weights)
        self._freeze_stages()

    def _freeze_stages(self):
        if self.frozen_stages >= 0:
            self.patch_embed.eval()
            for param in self.patch_embed.parameters():
                param.requires_grad = False

        if self.frozen_stages >= 1 and self.ape:
            self.absolute_pos_embed.requires_grad = False

        if self.frozen_stages >= 2:
            self.pos_drop.eval()
            for i in range(0, self.frozen_stages - 1):
                m = self.layers[i]
                m.eval()
                for param in m.parameters():
                    param.requires_grad = False

    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)

    def forward_features(self, x):
        x = self.patch_embed(x)
        if self.absolute_pos_embed is not None:
            x = x + self.absolute_pos_embed
        x = self.pos_drop(x)
        x = self.layers(x)
        x = self.norm(x)  # B L C
        return x

    def forward_head(self, x, pre_logits: bool = False):
        if self.global_pool == 'avg':
            x = x.mean(dim=1)
        return x if pre_logits else self.head(x)

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

构建模型并打印模型

#build.py
from model.swin_transformer import SwinTransformer
import argparse


parser = argparse.ArgumentParser()
#相关参数设置
parser.add_argument('--img_size', type=int, default=224)
parser.add_argument('--patch_size', type=int, default=4)
parser.add_argument('--in_chans', type=int, default=3)
parser.add_argument('--global_pool', type=str, default='avg')
parser.add_argument('--embed_dim', type=int, default=96)
parser.add_argument('--depths', type=tuple, default=(2,2,6,2))
parser.add_argument('--num_heads', type=tuple, default=(3,6,12,24))
parser.add_argument('--window_size', type=int, default=7)
parser.add_argument('--mlp_ratio', type=int, default=4)
parser.add_argument("--qkv_bias", type=bool, default=True)
parser.add_argument('--drop_rate', type=float, default=0.)
parser.add_argument('--attn_drop_rate', type=float, default=0.)
parser.add_argument('--drop_path_rate', type=float, default=0.2)
parser.add_argument('--ape', type=bool, default=False)
parser.add_argument('--patch_norm', type=bool, default=True)
parser.add_argument('--resume', type=str, default="../pretrain/swin_tiny_patch4_window7_224.pth")

args = parser.parse_args()

if __name__ == '__main__':
    model = SwinTransformer(img_size=args.img_size,
                            patch_size=args.patch_size,
                            in_chans=args.in_chans,
                            embed_dim=args.embed_dim,
                            depths=args.depths,
                            num_heads=args.num_heads,
                            window_size=args.window_size,
                            mlp_ratio=args.mlp_ratio,
                            qkv_bias=args.qkv_bias,
                            drop_rate=args.drop_rate,
                            drop_path_rate=args.drop_path_rate,
                            ape=args.ape,
                            patch_norm=args.patch_norm)
    print(model)

model

       --swin_transformer.py

       --build.py