Swin Transformer代码阅读注释

Posted 2022-02-16 HollowKnightZ

tags:

篇首语：本文由小常识网(cha138.com)小编为大家整理，主要介绍了Swin Transformer代码阅读注释相关的知识，希望对你有一定的参考价值。

Swin Transformer代码阅读注释

前言
Swin Transformer

前言

上一篇博文以论文中的内容介绍了Swin Transformer的网络结构和一些细节。本篇博文将从官方代码中的swin_transformer.py去详细介绍Swin Transformer结构，并补充代码中才有而论文中没有的细节。

|| 如果对 Swin Transformer 不了解建议先看论文介绍再看源码 ||
Swin Transformer介绍博客：论文阅读笔记：Swin Transformer

Swin Transformer

代码中实现的网络结构与论文中的结构如如下：

如上图所示，代码中使用PatchEmbed来实现Patch Partition + Linear Embedding,使用BasicLayer来实现Swin Transformer Block + PatchMerging，对于最后一个BasicLayer不使用PatchMerging来降采样。

Swin-T 的配置如下：

Swin-T参数配置表

网络结构介绍可看Swin Transformer介绍博客：论文阅读笔记：Swin Transformer
整体结构代码和注释如下（代码大部分和和 Vision Transformer 是一样）：

class SwinTransformer(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], num_heads=[3, 6, 12, 24],
                 window_size=7, mlp_ratio=4., qkv_bias=True, qk_scale=None,
                 drop_rate=0., attn_drop_rate=0., drop_path_rate=0.1,
                 norm_layer=nn.LayerNorm, ape=False, patch_norm=True,
                 use_checkpoint=False, **kwargs):
		'''
		img_size(int | tuple(int)): 输入图像尺寸. 默认: 224
		patch_size (int | tuple(int)): Patch尺寸. 默认: Swin-T参数配置表中 stage1中的96
	    in_chans (int): 输入图像通道. 默认: 3
	    num_classes (int): 分类数. 默认: 1000
	    embed_dim (int): Patch embedding的输出通道. 默认: 96 (Swin-T参数配置表中 stage 1 中的 96-d)
	    depths (tuple(int)):Swin Transformer Block 的个数. 默认:[2, 2, 6, 2] (Swin-T参数配置表中的[×2, ×2, ×6, ×2])
	    num_heads (tuple(int)): 不同层 MSA 计算中的 head 数. 默认：[3, 6, 12, 24] (Swin-T参数配置表中的[head 3，head 6，head 12，head 24])
	    window_size (int): W-MSA 和 SW-MSA 的 Window 尺寸. 默认: 7 (Swin-T参数配置表中的 “win.sz. 7×7”)
	    mlp_ratio (float): 通过MLP的输出通道倍数. 默认: 4 (Swin-T参数配置表中的“dim 96”,“dim 192”,“dim 384”,“dim 768”可以看出)
	    qkv_bias (bool): 使用 Linear 将输入映射到 qkv 时，Linear是否使用偏置. 默认: True
	    qk_scale (float):qk缩放比例，如果是 None 则使用根号 dim_k 分之一. 默认: None
	    drop_rate (float): dropout概率. 默认: 0
	    attn_drop_rate (float): attention 中的 dropout 概率. 默认: 0
	    drop_path_rate (float): attention 中的 droppath 概率. 默认: 0.1
	    norm_layer (nn.Module): 归一化方式. 默认: nn.LayerNorm.
	    ape (bool): 是否在 patch embedding 后使用绝对位置编码. 默认: False
	    patch_norm (bool): 是否在 patch embedding 后使用归一化. 默认: True
	    use_checkpoint (bool): 是否 checkpointing 节省内存. 默认: False
		'''    
		            
        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
        '''
        经过4个stage后的通道数（从96->768 即：96*2^(4-1)=768）
        '''
        self.num_features = int(embed_dim * 2 ** (self.num_layers - 1))
        self.mlp_ratio = mlp_ratio

        '''
        将图片划分成没有重叠的多个patch
        PatchEmbed代码在下文中介绍
        patches_resolution = [img_size[0]//patch_size[0], img_size[1]//patch_size[1]] = [56,56]
        num_patches = patches_resolution[0] * patches_resolution[1] = 3136
        '''
        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

        '''
        如果使用绝对位置编码则构建可学习的绝对位置编码参数：
        self.absolute_pos_embed : [1,3136,96]
        默认不使用
        '''
        if self.ape:
            self.absolute_pos_embed = nn.Parameter(torch.zeros(1, num_patches, embed_dim))
            trunc_normal_(self.absolute_pos_embed, std=.02)
		
		'''
		pos_drop 以 drop_rate 概率进行 Dropout
		'''
        self.pos_drop = nn.Dropout(p=drop_rate)

        '''
		构建首项为0，长度为depths(2+2+6+2=12)的等差数列，且最后一项小于drop_path_rate
		也就是说 传入 BasicLayer 的 droppath 概率是递增的。
		代码这里是让 drop_path_ratio 默认等于0.1
		最后利用参数构建 depth(12) 层 BasicLayer 层 
		BasicLayer 的代码在下文中介绍
		'''
        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))]  # stochastic depth decay rule
        # build layers
        self.layers = nn.ModuleList()
        for i_layer in range(self.num_layers):
            layer = BasicLayer(dim=int(embed_dim * 2 ** i_layer), #每个Basiclayer模块后通道数都翻倍
                               '''
                               每个Basiclayer都进行了降采样
                               所以input_resolution每一次都要除以2
                               '''
                               input_resolution=(patches_resolution[0] // (2 ** i_layer), 
                                                 patches_resolution[1] // (2 ** i_layer)), 
                               depth=depths[i_layer],
                               num_heads=num_heads[i_layer],
                               window_size=window_size,
                               mlp_ratio=self.mlp_ratio,
                               qkv_bias=qkv_bias, qk_scale=qk_scale,
                               drop=drop_rate, attn_drop=attn_drop_rate,
                               drop_path=dpr[sum(depths[:i_layer]):sum(depths[:i_layer + 1])],
                               norm_layer=norm_layer,
                               '''
                               如果i_layer 是最后一层则不使用PatchMerging 来降采样
                               '''
                               downsample=PatchMerging if (i_layer < self.num_layers - 1) else None,
                               use_checkpoint=use_checkpoint)
            self.layers.append(layer)
		
		'''
		进行归一化和平均池化
		最后用一个Linear做预测head
		'''
        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)
	
	'''
	初始化权重
	'''
    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):
    	'''
    	如图所示先进行patch embedding
    	如果使用绝对位置偏置就加上绝对位置编码
    	'''
        x = self.patch_embed(x)
        if self.ape:
            x = x + self.absolute_pos_embed
        x = self.pos_drop(x)
		
		'''
		循环执行Blocks
		'''
        for layer in self.layers:
            x = layer(x)
		
		'''
		归一化并平均池化
		'''
        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)
        '''
        对swin transformer的特征提取进行预测
        '''
        x = self.head(x)
        return x

1 PatchEmbed

Swin Transformer中的PatchEmbed模块和 VIT 中的 Linear Projection of Flattened Patches: PatchEmbed模块差不多，可查看博文Vision Transformer（Pytorch版）代码阅读注释查看，其主要思想是通过感受野大小等于步距大小的卷积来实现，与 VIT 不同的是其使用了nn.LayerNorm。

PatchEmbed代码和注释如下：

class PatchEmbed(nn.Module):
    """ Image to Patch Embedding

    Args:
        img_size (int): 图像尺寸.  默认: 224.
        patch_size (int): token尺寸. 默认: 4.
        in_chans (int): 图像通道. 默认: 3.
        embed_dim (int): patch embed通道. 默认: 96.
        norm_layer (nn.Module, optional): 归一化. Default: None
    """
    def __init__(self, img_size=224, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None):
        super().__init__()
        '''
        self.image_size = (224,224)
        self.patch_size = (4,4)
        self.patches_resolution = [56,56]
        self.num_patches = 56*56=3136
        '''
        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 = 3
        self.embed_dim = 96
		'''
        self.in_chans = in_chans
        self.embed_dim = embed_dim
		
		'''
		self.proj = nn.Conv2d(3,96,(4,4),4)
		self.norm = nn.LayerNorm(96)
		'''
        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])."
            
        '''
        self.proj(x):[B,3,224,224]->[B,96,56,56]
        flatten(2):[B,96,56,56]->[B,96,56*56]=[B,96,3136]
        transpose(1, 2):[B,96,3136]->[B,3136,96]
        self.norm(x):[B,3136,96]->[B,3136,96]
        '''    
        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

BasicLayer

代码中使用BasicLayer来实现论文中的Swin Transformer Block + PatchMerging，对于最后一个BasicLayer不使用PatchMerging来降采样。

BasicLayer的代码和注释如下：

class BasicLayer(nn.Module):
    """ A basic Swin Transformer layer for one stage.

    Args:
        dim (int): 输入特征图的通道数.
        input_resolution (tuple[int]): 输入特征图的分辨率大小.
        depth (int): SwinTransformerBlock的个数.
        num_heads (int): Muti-Head Self-Attention 中的head个数.
        window_size (int): window 大小.
        mlp_ratio (float): patch embedding通过MLP的通道倍数.
        qkv_bias (bool): 使用 Linear 将输入映射到 qkv 时，Linear是否使用偏置. 默认: True
	    qk_scale (float):qk缩放比例，如果是 None 则使用根号 dim_k 分之一. 默认: None
	    drop (float, optional): dropout概率. 默认: 0
	    attn_drop (float, optional): attention 中的 dropout 概率. 默认: 0
	    drop_path (float | tuple[float], optional): attention 中的 droppath 概率. 默认: 0.1
	    norm_layer (nn.Module): 归一化方式. 默认: nn.LayerNorm.
        downsample (nn.Module | None, optional): 降采样层. 默认: None 代码使用PatchMerging
	    use_checkpoint (bool): 是否 checkpointing 节省内存. 默认: False
    """

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

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

        '''
        构建SwinTransformerBlock
        SwinTransformerBlock代码在下文介绍
        '''
        self.blocks = nn.ModuleList([
            SwinTransformerBlock(dim=dim, input_resolution=input_resolution,
                                 num_heads=num_heads, window_size=window_size,
                                 '''
                                 如果i是偶数，则表示是W-MSA，shift_size =0
                                 如果i是奇数，则表示是SW-MSA，shift_size = window_size // 2
                                 '''
                                 shift_size=0 if (i % 2 == 0) else window_size // 2,
                                 mlp_ratio=mlp_ratio,
                                 qkv_bias=qkv_bias, qk_scale=qk_scale,
                                 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)])

        '''
        使用PatchMerging进行降采样
        PatchMerging代码在下文介绍
        '''
        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

SwinTransformerBlock

SwinTransformerBlock的结构如下：

Mlp

此处和 VIT 中MLP一模一样，可查看Vision Transformer（Pytorch版）代码阅读注释。代码也很简单，就不再做任何赘述了。
Mlp的代码如下：

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

window_partition

W-MSA和SW-MSA首先需要将特征图拆分成多个windows。

window_partition的代码和注释如下：

def window_partition(x, window_size):
    """
    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
    '''
    [B, H, W, C] -> [BHW//(window_size*window_size), window_size, window_size, C]
    '''
    x = x.view(B, H // window_size, window_size, W // window_size, window_size, C)
    windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
    return windows

window_reverse

在对每一个windows进行WSA计算以后需要将其还原成正常的特征图传入下一模块中。其实就是window_partition的逆过程。
window_reverse的代码和注释如下：

def window_reverse(windows, window_size, H, W):
    """
    Args:
        windows: (num_windows*B, window_size, window_size, C)
        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

WindowAttention

WindowAttention就是在 Vision Transformer 模块的Attention基础上加入了相对位置偏移relative_position_bias_table（即论文中提出的 Relative Position Bias）来提升精度：

生成相对位置偏置的过程（以一个head为例，假设window_h = 2,window_w = 2,相关代码已在图中标出）：
1.随机生成相对位置偏置表relative_position_bias_table：

self.relative_position_bias_table = nn.Parameter(
		torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads))  # 2*Wh-1 * 2*Ww-1, nH

2.首先windows内部每个像素都有自己的位置编码，其绝对位置编码的坐标abs_coords如果以左上角为原点，则如下图：

coords_h = torch.arange(self.window_size[0])
coords_w = torch.arange(self.window_size[1])
coords = torch.stack(torch.meshgrid([coords_h, coords_w]))  # 2, Wh, Ww
coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww

3.为了获得每个绝对坐标相对于其他坐标的相对位置，则需要用每个绝对坐标减去其他绝对坐标，即:

用 $(0, 0)$ 分别减去 $(0, 1)$ ，