7.卷积和Transformer结合的ViT

7.卷积和Transformer结合的ViT

一、回顾与简介

前面两节课我们讲了Swin Transformer的结构，以及其中的细节部分，进行了实现，其中由Swin Block 以及 Patch Merging等等，上节课讲了 SW-MSA的shift和mask，对于shift之后，其中window中需要的保留，不需要的去掉，用到了boardcasting等等。

这节课我们考虑卷积和Transformer如何结合，我们首先看一下conv是如何计算的，如果输入是5×5，卷积核是3×3，进行一个滑窗之后，最终输出是3×3的特征图

那我们看一下细节，如何实现

1×1的卷积可以对通道数进行修改，比如我们现在是4×4×3的图像，用1×1卷积核计算完，如果out_channels为4，那么输出的结果就是4×4×4的特征图

如果是4×4大小，3通道的图像，输出要求 图像大小不变，通道数变为4，Transformer有QKV三个映射，分别映射到embedd_dim的维度，通过进行QK’再乘以V

可以看出Transformer与CNN有点像，需要注意的是：

1.conv是利用滑窗来算，Attn是计算当前tensor中每一个位置对其他位置的关系，但是他们两个又可以同时完成一个状态，对输入的tensor求一个表征，并还可以给它改变一个维度

比如我们输入的尺寸没变，输入的通道数从3变为了4，Conv和Transformer都可以做到的

如果我们将它们内部实现封起来，当作黑盒来看，它们完成的事情是类似的，只不过它们具体的计算方式不一样：

conv是一个相对于固定的窗口在输入的tensor上进行滑动的操作，乘以的是固定的conv kernel，我们可以叫做相对local的状态，算的是局部的信息；而Transformer谁都可以看，它看到的是gloab全局的信息，这就是它俩的区别。我可以单独用gloab的信息或者local的信息，但也有一种可能性是我们将它俩结合，目前conv是比Transformer计算量更小的，如果需要更小模型的场景上conv是更work的。

今天讲的 MobileViT 就使用了下图的这样一种的结构，其中使用了Transformer，红线是Residual map path，我们要做的就是Conv+Transformer

整体结构中，前面有一部分是传统的Conv层，中间有一部分是Transformer层，其中起了MV2 Block和MViT Block，MV2 Block对应的MobileNetV2的一个block的结构，类似Resnet block的结构，MViT Block等会儿我们会讲。

通常我们会将前面的输入层叫做 Stem层（输入，3×3的卷积核，再加pooling）

最后在Global Pool之后叫 head层

在中间就是block层，看一下核心模块MV2 Block的网络结构，当Stride=1 或者 2的时候，架构是不一样的，其中用的激活函数是Silu函数在左侧显示，而Silu与Gelu的对比在右侧，蓝色线是Gelu，它们在计算时指数有一点点不一样。

我们可以看到Block通常接一个1×1的卷积，是为了调整tensor的通道数，而DConv是加了Group Conv，为了减少计算量，之后加BN和Silu，最后加1×1的卷积和BN是为了调回来，最后Add进行残差——这就是一个MV2 Block的一个架构。

MobileVit就是将其中MV2 Block中一些模块替换为ViT Block

二、论文

首先看Xl ，它是H×W×d，它是经过1×1卷积之后得到的tensor，经过一个flatten(non-overlapping)的切分后，得到了XU ，其中用到了Inter-patch relationship，这就是本篇论文的核心，得到经过Transformer的XG 之后，又做了point-wise convlution 以及 concatenation operation 还有N×N的卷积

读完文字后，脑子中可能还有点模糊，我们需要拿图像来更加直观和清晰地展现网络结构，上面3×3和1×1的卷积对应的就是第一段的内容，它对tensor无非就是做一些shape变换操作，之后还需要做一些操作才能进入Transformer，我们Transformer原则上是不改变tensor的尺度的，输出之后应该还是某一个tensor，我们再做后续的部分1×1的Conv等等

最关键的部分是画蓝色块的部分，比如卷积过后 H× W× d这么一个tensor，怎么样输入到Transformer中去？

具体来讲，最左边的输出是Conv后的结果，我们需要进行Patch Partition操作，这里其实可以叫做window操作，在H× W× d的这样一个尺寸上切出小window出来，比如我们切出9个window，A B C 到I，经过上一节课的学习，我们知道如果是Swin的话我们就在window上自注意力的计算，而MobileViT做的不一样

在Transformer做的操作是将 H× W× d 输入进行flatten拉直，比如我们可以将A扯平，变为1×d，可以将每个patch展开成为中间这样

其中会将A B C到 I的第一个token取出来，送到Transformer中去，当然Transformer不改变维度，第一个做完继续做第二个token，再送到Transformer中进行计算。

从整体看，我们假如是2×2的窗口，我们第一个取每个窗口中第一个元素，来算Self-attention；之后再算第二个位置的，如果是2×2的窗口，一共做4次，如果是3×3的窗口，那就做9次，但是MobileViT是固定2×2的窗口

这么做的目的是：为了看到全局的信息，还要减少计算量，和空洞卷积类似

三、代码实现

3.1 Swin完善

上一届对Swin Transformer的 WindowAttention的MSA，SW-MSA，SwinBlock进行了代码

今天我们对Swin Stage以及整个Swin 的Class代码进行编写

# CLass 7classSwinStage(nn.Module):def__init__(self, dim, input_resolution, depth, num_heads, window_size, patch_merging =None):super().__init__()
        self.blocks = nn.ModuleList()for i inrange(depth):
            self.blocks.append(
                SwinBlock(dim= dim,
                          input_reslution=input_resolution,
                          num_heads= num_heads,
                          window_size = window_size,
                          shift_size=0if(i %2==0)else window_size//2))if patch_merging isNone:# 最后一个 stage不做patch_merging
            self.patch_merging = Indentity()else:
            self.patch_merging = PatchMerging(input_resolution=input_resolution,dim = dim)defforward(self, x):for block in self.blocks:
            x = block(x)
        x = self.patch_merging(x)return x

Swin块：

# CLass 7classSwin(nn.Module):def__init__(self,
                 image_size=224,
                 patch_size=4,
                 in_channels=4,
                 embed_dim=96,
                 window_size=7,
                 num_heads=[3,6,12,24],
                 depths=[2,2,6,2],
                 num_classes=1000):super().__init__()
        self.num_classes = num_classes
        self.depths = depths
        self.num_heads = num_heads
        self.embed_dim = embed_dim
        self.num_stages =len(depths)
        self.num_features =int(self.embed_dim *2**(self.num_stages-1))
        self.patch_resolution =[image_size // patch_size , image_size // patch_size]

        self.patch_embedding = PatchEmbedding(patch_size=patch_size, embed_dim= embed_dim)
        self.stages =nn.ModuleList()for idx,(depth,num_heads)inenumerate(zip(self.depths, self.num_heads)):
            stage = SwinStage(dim=int(self.embed_dim *2**idx),
                              input_resolution=(self.patch_resolution[0]//(2** idx),
                                                 self.patch_resolution[0]//(2** idx)),
                              depth = depth,
                              num_heads = num_heads,
                              window_size = window_size,
                              patch_merging= PatchMerging if(idx < self.num_stages-1)elseNone)
            self.stages.append(stage)
        self.norm = nn.LayerNorm(self.num_features)
        self.avgpool = nn.AdaptiveAvgPool1d(1)
        self.fc = nn.Linear(self.num_features,self.num_classes)defforward(self, x):
        x = self.patch_embedding(x)for stage in self.stages:
            x = stage(x)
        x = self.norm(x)
        x = x.permute(0,2,1)# [B, embed_dim, num_windows]
        x = self.avgpool(x)#  [B, embed_dim, 1]
        x = x.flatten(1)
        x = self.fc(x)return x

最终main函数就将之前的整合进入 SwinStage中了

defmain():
    t = torch.randn([4,3,224,224])# patch_embedding = PatchEmbedding(patch_size=4, embed_dim=96)# swin_block_w_msa = SwinBlock(dim=96, input_reslution=[56,56], num_heads=4, window_size=7,shift_size=0)# swin_block_sw_msa = SwinBlock(dim=96, input_reslution=[56,56], num_heads=4, window_size=7,shift_size=7//2)## patch_merging = PatchMerging(input_resolution=[56,56], dim=96)## print('image shape = [4, 3, 224, 224]')# out = patch_embedding(t)  #[4, 56, 56, 96]# print('patch_embedding out shape= ',out.shape)# out = swin_block_w_msa(out)# out = swin_block_sw_msa(out)# print('swinBlock out shape= ',out.shape)# out = patch_merging(out)# print('patch_merging out shape= ',out.shape)
    model = Swin()print(model)
    out = model(t)print(out.shape)

输出结果如下

模型结构

Swin((patch_embedding): PatchEmbedding((patch_embed): Conv2d(3,96, kernel_size=(4,4), stride=(4,4))(norm): LayerNorm((96,), eps=1e-05, elementwise_affine=True))(stages): ModuleList((0): SwinStage((blocks): ModuleList((0): SwinBlock((attn_norm): LayerNorm((96,), eps=1e-05, elementwise_affine=True)(attn): WindowAttention((softmax): Softmax(dim=-1)(qkv): Linear(in_features=96, out_features=288, bias=True)(proj): Linear(in_features=96, out_features=96, bias=True))(mlp_norm): LayerNorm((96,), eps=1e-05, elementwise_affine=True)(mlp): Mlp((fc1): Linear(in_features=96, out_features=384, bias=True)(fc2): Linear(in_features=384, out_features=96, bias=True)(act): GELU(approximate='none')(dropout): Dropout(p=0.0, inplace=False)))(1): SwinBlock((attn_norm): LayerNorm((96,), eps=1e-05, elementwise_affine=True)(attn): WindowAttention((softmax): Softmax(dim=-1)(qkv): Linear(in_features=96, out_features=288, bias=True)(proj): Linear(in_features=96, out_features=96, bias=True))(mlp_norm): LayerNorm((96,), eps=1e-05, elementwise_affine=True)(mlp): Mlp((fc1): Linear(in_features=96, out_features=384, bias=True)(fc2): Linear(in_features=384, out_features=96, bias=True)(act): GELU(approximate='none')(dropout): Dropout(p=0.0, inplace=False))))(patch_merging): PatchMerging((reduction): Linear(in_features=384, out_features=192, bias=True)(norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)))(1): SwinStage((blocks): ModuleList((0): SwinBlock((attn_norm): LayerNorm((192,), eps=1e-05, elementwise_affine=True)(attn): WindowAttention((softmax): Softmax(dim=-1)(qkv): Linear(in_features=192, out_features=576, bias=True)(proj): Linear(in_features=192, out_features=192, bias=True))(mlp_norm): LayerNorm((192,), eps=1e-05, elementwise_affine=True)(mlp): Mlp((fc1): Linear(in_features=192, out_features=768, bias=True)(fc2): Linear(in_features=768, out_features=192, bias=True)(act): GELU(approximate='none')(dropout): Dropout(p=0.0, inplace=False)))(1): SwinBlock((attn_norm): LayerNorm((192,), eps=1e-05, elementwise_affine=True)(attn): WindowAttention((softmax): Softmax(dim=-1)(qkv): Linear(in_features=192, out_features=576, bias=True)(proj): Linear(in_features=192, out_features=192, bias=True))(mlp_norm): LayerNorm((192,), eps=1e-05, elementwise_affine=True)(mlp): Mlp((fc1): Linear(in_features=192, out_features=768, bias=True)(fc2): Linear(in_features=768, out_features=192, bias=True)(act): GELU(approximate='none')(dropout): Dropout(p=0.0, inplace=False))))(patch_merging): PatchMerging((reduction): Linear(in_features=768, out_features=384, bias=True)(norm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)))(2): SwinStage((blocks): ModuleList((0): SwinBlock((attn_norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)(attn): WindowAttention((softmax): Softmax(dim=-1)(qkv): Linear(in_features=384, out_features=1152, bias=True)(proj): Linear(in_features=384, out_features=384, bias=True))(mlp_norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)(mlp): Mlp((fc1): Linear(in_features=384, out_features=1536, bias=True)(fc2): Linear(in_features=1536, out_features=384, bias=True)(act): GELU(approximate='none')(dropout): Dropout(p=0.0, inplace=False)))(1): SwinBlock((attn_norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)(attn): WindowAttention((softmax): Softmax(dim=-1)(qkv): Linear(in_features=384, out_features=1152, bias=True)(proj): Linear(in_features=384, out_features=384, bias=True))(mlp_norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)(mlp): Mlp((fc1): Linear(in_features=384, out_features=1536, bias=True)(fc2): Linear(in_features=1536, out_features=384, bias=True)(act): GELU(approximate='none')(dropout): Dropout(p=0.0, inplace=False)))(2): SwinBlock((attn_norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)(attn): WindowAttention((softmax): Softmax(dim=-1)(qkv): Linear(in_features=384, out_features=1152, bias=True)(proj): Linear(in_features=384, out_features=384, bias=True))(mlp_norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)(mlp): Mlp((fc1): Linear(in_features=384, out_features=1536, bias=True)(fc2): Linear(in_features=1536, out_features=384, bias=True)(act): GELU(approximate='none')(dropout): Dropout(p=0.0, inplace=False)))(3): SwinBlock((attn_norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)(attn): WindowAttention((softmax): Softmax(dim=-1)(qkv): Linear(in_features=384, out_features=1152, bias=True)(proj): Linear(in_features=384, out_features=384, bias=True))(mlp_norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)(mlp): Mlp((fc1): Linear(in_features=384, out_features=1536, bias=True)(fc2): Linear(in_features=1536, out_features=384, bias=True)(act): GELU(approximate='none')(dropout): Dropout(p=0.0, inplace=False)))(4): SwinBlock((attn_norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)(attn): WindowAttention((softmax): Softmax(dim=-1)(qkv): Linear(in_features=384, out_features=1152, bias=True)(proj): Linear(in_features=384, out_features=384, bias=True))(mlp_norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)(mlp): Mlp((fc1): Linear(in_features=384, out_features=1536, bias=True)(fc2): Linear(in_features=1536, out_features=384, bias=True)(act): GELU(approximate='none')(dropout): Dropout(p=0.0, inplace=False)))(5): SwinBlock((attn_norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)(attn): WindowAttention((softmax): Softmax(dim=-1)(qkv): Linear(in_features=384, out_features=1152, bias=True)(proj): Linear(in_features=384, out_features=384, bias=True))(mlp_norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)(mlp): Mlp((fc1): Linear(in_features=384, out_features=1536, bias=True)(fc2): Linear(in_features=1536, out_features=384, bias=True)(act): GELU(approximate='none')(dropout): Dropout(p=0.0, inplace=False))))(patch_merging): PatchMerging((reduction): Linear(in_features=1536, out_features=768, bias=True)(norm): LayerNorm((1536,), eps=1e-05, elementwise_affine=True)))(3): SwinStage((blocks): ModuleList((0): SwinBlock((attn_norm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)(attn): WindowAttention((softmax): Softmax(dim=-1)(qkv): Linear(in_features=768, out_features=2304, bias=True)(proj): Linear(in_features=768, out_features=768, bias=True))(mlp_norm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)(mlp): Mlp((fc1): Linear(in_features=768, out_features=3072, bias=True)(fc2): Linear(in_features=3072, out_features=768, bias=True)(act): GELU(approximate='none')(dropout): Dropout(p=0.0, inplace=False)))(1): SwinBlock((attn_norm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)(attn): WindowAttention((softmax): Softmax(dim=-1)(qkv): Linear(in_features=768, out_features=2304, bias=True)(proj): Linear(in_features=768, out_features=768, bias=True))(mlp_norm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)(mlp): Mlp((fc1): Linear(in_features=768, out_features=3072, bias=True)(fc2): Linear(in_features=3072, out_features=768, bias=True)(act): GELU(approximate='none')(dropout): Dropout(p=0.0, inplace=False))))(patch_merging): Indentity()))(norm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)(avgpool): AdaptiveAvgPool1d(output_size=1)(fc): Linear(in_features=768, out_features=1000, bias=True))

3.2 数据加载

那我们正常情况需要从数据集中加载数据，那数据如何取呢？

首先看一下python的 for 循环在做什么

比如我想for循环遍历，那我应该怎么写MyInterable方法呢？

my_iterable = MyInterable()for d in my_iterable:print(d)

我们需要实现getitem类，来返回__getitem__对象，然后__getitem__中通过next方法来取值，如果想手动选择索引，在MyInterable中实现__getitem__方法

Dataset返回的是单个的样本，而Dataloader是返回一个batch的数据

查看下图，我们走一遍流程，for循环可以从dataloader中取数据，那它本质上调用了__iter__()方法，它其中其实就是调用了DataLoaderIter中__iter__()方法，其中有一个__next__()方法来取数据，next中又调用了fetch，fetch又从Sampler拿到indices，再去fetch data，之后再从Dataset中__getitem__()来数据。从 Dataset中取到的数据是list，需要通过collate_fn()帮我们把list变为tensor