PVT：Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions

---

Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions

多用途的backbone，CLS、DET、SEG…
通过在transfomer blocks之间进行Shrink操作，来获得多尺度的特征图
通过spatial-reduction attention来降低计算量

一、Feature Pyramid for Transformer

通过每一层前的reshape和patch embedding来进行Shrink操作
将transformer的输出reshape成三维
在patch embedding中，先通过卷积降低HW并增加C，再reshape成二维

1.代码

PyramidVisionTransformer：将transformer的输出reshape成三维，此时的HW小于上一层，C大于上一层，参数来自于PatchEmbed

 def forward_features(self, x):
     outs = []

     B = x.shape[0]

     for i in range(self.num_stages):
         patch_embed = getattr(self, f"patch_embed{i + 1}")  # patch embedding中进行Shrink
         pos_embed = getattr(self, f"pos_embed{i + 1}")
         pos_drop = getattr(self, f"pos_drop{i + 1}")
         block = getattr(self, f"block{i + 1}")
         x, (H, W) = patch_embed(x) # 从patch_embed获得这一层的HW
         if i == self.num_stages - 1:
             pos_embed = self._get_pos_embed(pos_embed[:, 1:], patch_embed, H, W)
         else:
             pos_embed = self._get_pos_embed(pos_embed, patch_embed, H, W)

         x = pos_drop(x + pos_embed)
         for blk in block:
             x = blk(x, H, W)  # Attention中对k和v做SR
         x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()  # 将transformer的输出reshape成三维 patch_size[4,2,2,2] embed_dims:[64, 128, 256, 512] 在atten之后，reshape成三维，重新进行patch embedding等操作
         outs.append(x)

     return outs

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

     if self.F4:
         x = x[3:4]

     return x

PatchEmbed：先通过卷积降低HW并增加C，再reshape成二维

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

    def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768):
        super().__init__()
        img_size = to_2tuple(img_size)
        patch_size = to_2tuple(patch_size)

        self.img_size = img_size
        self.patch_size = patch_size
        assert img_size[0] % patch_size[0] == 0 and img_size[1] % patch_size[1] == 0, \
            f"img_size {img_size} should be divided by patch_size {patch_size}."
        self.H, self.W = img_size[0] // patch_size[0], img_size[1] // patch_size[1]
        self.num_patches = self.H * self.W
        self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size) # kernel_size=stride=patch_size  将H和W缩小patch_size倍  C变为embed_dim
        self.norm = nn.LayerNorm(embed_dim)

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

        x = self.proj(x).flatten(2).transpose(1, 2)  # 再卷积降低HW增加C 再reshape成二维 B, D, H/P, W/P -> B, D, N(HW/PP) -> B, N, D
        x = self.norm(x)
        H, W = H // self.patch_size[0], W // self.patch_size[1]

        return x, (H, W)

二、 Transformer Encoder

Spatial-Reduction Attention：先将输入reshape成三维，再通过卷积降低HW，再将HW reshape成一个维度

1.代码

Attention

class Attention(nn.Module):
    def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0., sr_ratio=1):
        super().__init__()
        assert dim % num_heads == 0, f"dim {dim} should be divided by num_heads {num_heads}."

        self.dim = dim
        self.num_heads = num_heads
        head_dim = dim // num_heads
        self.scale = qk_scale or head_dim ** -0.5

        self.q = nn.Linear(dim, dim, bias=qkv_bias)
        self.kv = nn.Linear(dim, dim * 2, bias=qkv_bias)
        self.attn_drop = nn.Dropout(attn_drop)
        self.proj = nn.Linear(dim, dim)
        self.proj_drop = nn.Dropout(proj_drop)

        self.sr_ratio = sr_ratio
        if sr_ratio > 1:
            self.sr = nn.Conv2d(dim, dim, kernel_size=sr_ratio, stride=sr_ratio)   # Spatial Reduction：kernel_size=stride=sr_ratio 将x_的HW降低sr_ratio倍，C不变  降低计算量和显存占用
            self.norm = nn.LayerNorm(dim)

    def forward(self, x, H, W):
        B, N, C = x.shape
        q = self.q(x).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)  # q不做处理

        if self.sr_ratio > 1:
            x_ = x.permute(0, 2, 1).reshape(B, C, H, W)  # 先reshape,将N分为H和W
            x_ = self.sr(x_).reshape(B, C, -1).permute(0, 2, 1)  # 对k和v做nn.Conv2d  从而SR
            x_ = self.norm(x_)
            kv = self.kv(x_).reshape(B, -1, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)  # 再将H和Wreshape为N
        else:
            kv = self.kv(x).reshape(B, -1, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
        k, v = kv[0], kv[1]

        attn = (q @ k.transpose(-2, -1)) * self.scale
        attn = attn.softmax(dim=-1)
        attn = self.attn_drop(attn)

        x = (attn @ v).transpose(1, 2).reshape(B, N, C)
        x = self.proj(x)
        x = self.proj_drop(x)

        return x

三、Ablation Study

1.Pyramid Structure

ViT采用patch size为32时，生成的特征图分辨率太低，导致效果不好
ViT采用的patch size过小时会导致显存占用过高
PVT可以在浅层和深层分别处理高分辨率和低分辨率的特征图

2.Deeper vs. Wider

Wider：将PVT-Small中hidden dimensions {C1, C2, C3, C4}扩宽1.4倍
Deeper：增加Transformer中的block数目

3.Computation Overhead

PVT的计算量高于ResNet50，低于PVT-Small