sysinternals 源码(swin_transformer源码详解)
注:为了更加实例化的说明 ,本文假设输入图像大小为(224,224,3)
整体架构
对于一张224*224的图像 ,首先,经过4*4的卷积 ,将图像维度化为 4,56,56,128的特征图 ,对特征图维度进行变换 ,得到4*3136*128的图像 ,即对图像进行了embeding ,然后将图像输入transforer block ,将特征图转变为8*8的窗口 ,进行注意力机制的计算 ,一个transformer block包含窗口自注意力W-MAS ,计算8*8窗口内部的特征和滑动窗口自注意力SW-MSA,计算窗口间的特征 ,经过transformer的计算后 ,再进行patch mergeing将特征图大小减半,类似于卷积 。
1.图像数据patch编码
首先,对于输入的图像 ,假设为224*224 ,我们采用4*4的卷积,然后将图像进行flatten ,形成一个个patch,最后输出维度为batch_size * HW * Channels,H=W=224/4
代码如下:
class PatchEmbed(nn.Module): r""" Image to Patch Embedding Args: img_size (int): Image size. Default: 224. patch_size (int): Patch token size. Default: 4. in_chans (int): Number of input image channels. Default: 3. embed_dim (int): Number of linear projection output channels. Default: 96. norm_layer (nn.Module, optional): Normalization layer. Default: None """ 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 # in_channels:3,out_channels:128 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}) doesnt match model ({self.img_size[0]}*{self.img_size[1]})." # 卷积 x = self.proj(x).flatten(2).transpose(1, 2) # B Ph*Pw C # print(x.shape) #4 3136 96 其中3136就是 224/4 * 224/4 相当于有这么长的序列 ,其中每个元素是96维向量 if self.norm is not None: x = self.norm(x) # print(x.shape) 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 flops3.transformer block
一个transformer block由w-MSA和SW-MSA组成
W-MSA/SW-MSA
输入维度为4,3136,128的序列x ,首先将其维度变换为4,56,56,128 ,再经过维度变换 ,将维度变成 256, 49, 128 ,即表示 ,有256个特征图 ,每个特征图有49个tokens ,每个token是128维的向量 。
首先做W-MSA,对于W-SMA ,不对窗口进行偏移 ,经过多头注意力的计算,得到结果 ,对于SW-MSA ,窗口进行偏移,加入mask后 ,做相同的多头注意力的计算 。最后将窗口再偏移回去 。
多头注意力:首先构造维度为256, 4, 49, 32的q,k,v辅助向量 ,256表示有256个特征图 ,4表示有4个head,49表示有49个tokens ,32表示 ,每个头32个向量 ,然后经过多头注意力的计算 ,其中 ,会加入相对位置编码
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. window_size (tuple[int]): The height and width of the window. num_heads (int): Number of attention heads. qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set 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, window_size, num_heads, qkv_bias=True, qk_scale=None, attn_drop=0., proj_drop=0.): super().__init__() self.dim = dim self.window_size = window_size # Wh, Ww self.num_heads = num_heads head_dim = dim // num_heads self.scale = qk_scale or head_dim ** -0.5 # define a parameter table of relative position bias 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 # get pair-wise relative position index for each token inside the window 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 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] += self.window_size[0] - 1 # shift to start from 0 relative_coords[:, :, 1] += self.window_size[1] - 1 relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1 relative_position_index = relative_coords.sum(-1) # Wh*Ww, Wh*Ww self.register_buffer("relative_position_index", relative_position_index) self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias) self.attn_drop = nn.Dropout(attn_drop) self.proj = nn.Linear(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 forward(self, x, mask=None): """ Args: x: input features with shape of (num_windows*B, N, C) mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None """ # num_windows, Wh*Ww, Wh*Ww B_, N, C = x.shape # 3, 256, 4, 49, 32 qkv = self.qkv(x).reshape(B_, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4) # print(qkv.shape) # 256, 4, 49, 32 q, k, v = qkv[0], qkv[1], qkv[2] # make torchscript happy (cannot use tensor as tuple) # print(q.shape) # print(k.shape) # print(v.shape) # 256, 4, 49, 49 q = q * self.scale attn = (q @ k.transpose(-2, -1)) # print(attn.shape) # 相对位置编码 49*49*4 relative_position_bias = self.relative_position_bias_table[self.relative_position_index.view(-1)].view( self.window_size[0] * self.window_size[1], self.window_size[0] * self.window_size[1], -1) # Wh*Ww,Wh*Ww,nH # print(relative_position_bias.shape) # 4, 49, 49 relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous() # nH, Wh*Ww, Wh*Ww # print(relative_position_bias.shape) # 加入位置编码 256, 4, 49, 49 attn = attn + relative_position_bias.unsqueeze(0) # print(attn.shape) if mask is not None: nW = mask.shape[0] attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0) attn = attn.view(-1, self.num_heads, N, N) attn = self.softmax(attn) else: attn = self.softmax(attn) # dropout层 attn = self.attn_drop(attn) # print(attn.shape) # qkv x = (attn @ v).transpose(1, 2).reshape(B_, N, C) # print(x.shape) # 全连接层 x = self.proj(x) # print(x.shape) # dropout层 x = self.proj_drop(x) # print(x.shape) return x4.下采样
下采样操作 ,但是不同于池化,这个相当于间接的 (对H和W维度进行间隔采样后拼接在一起 ,得到H/2,W/2,C*4)
代码如下:
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, 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) # 间隔采样 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 flops5.相对位置编码
有关swin transformer相对位置编码的理解:
假设window_size是7*7
那么窗口中共有49个patch ,共有49*49个相对位置,每个相对位置有两个索引对应x和y两个方向 ,每个索引值的取值范围是[-6,6] 。(第0行相对第6行 ,x索引相对值为-6;第6行相对第0行,x索引相对值为6;所以索引取值范围是[-6,6])
# get pair-wise relative position index for each token inside the window
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
# 2, Wh*Ww, Wh*Ww, https://www.cnblogs.com/sgdd123/p/7603004.html
relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]
# Wh*Ww, Wh*Ww, 2, [i,j,:]表示窗口内第i个patch相对于第j个patch的坐标
relative_coords = relative_coords.permute(1, 2, 0).contiguous()此时 ,构建出来的relative_coords的shape是[49, 49, 2] ,[i, j, :]表示窗口内第i个patch相对于第j个patch的坐标 。
由于此时索引取值范围中包含负值 ,可分别在每个方向上加上6 ,使得索引取值从0开始 。此时 ,索引取值范围为[0,12]
relative_coords[:, :, 0] += self.window_size[0] - 1 # shift to start from 0
relative_coords[:, :, 1] += self.window_size[1] - 1有了这些相对位置坐标之后 ,就可以根据这些坐标获取对应的position bias ,即论文中公式(4)中的B:
这个时候可以构建一个shape为[13,13]的table ,则当相对位置为(i ,j)时,B=table[i, j] 。(i ,j的取值范围都是[0, 12])由于论文中使用的时multi-head-self-attention ,所以table[i, j]的值应该是一个维度为num_heads的一维向量 。
在代码中,实现如下:(注意 ,此时的table将二维的位置关系 ,合并为了一维的位置关系)
# 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 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads))为了与table对应,根据相对位置坐标取值时 ,也需要将二维相对坐标(i, j)映射为一维相对坐标(i*13+j) , 在代码中体现为:
relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
relative_position_index = relative_coords.sum(-1) # Wh*Ww, Wh*Ww最后 ,就可以根据映射后的坐标来对B进行取值了:
relative_position_bias = self.relative_position_bias_table[self.relative_position_index.view(-1)].view(
self.window_size[0] * self.window_size[1], self.window_size[0] * self.window_size[1], -1) # Wh*Ww,Wh*Ww,nH
relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous() # nH, Wh*Ww, Wh*Ww附注:
将二维相对坐标(i, j)映射为一维相对坐标时 ,最简单的映射方式是将i和j相加 ,但这样无法区分(0, 2)和(2, 0) ,因为相加的结果都是2;所以作者采用了i*13+j这种方式 ,其中13 = 2*window_size - 1, 即j取值的最大值。类似于将一个二维数组打平后 ,每个元素的位置 。
参考信息:
https://blog.csdn.net/weixin_42364196/article/details/119954379
创心域SEO版权声明:以上内容作者已申请原创保护,未经允许不得转载,侵权必究!授权事宜、对本内容有异议或投诉,敬请联系网站管理员,我们将尽快回复您,谢谢合作!