【图像分类】2022-CMT CVPR

【图像分类】2022-CMT CVPR

论文题目：CMT: Convolutional Neural Networks Meet Vision Transformers

论文链接：https://arxiv.org/abs/2107.06263

论文代码：https://github.com/ggjy/cmt.pytorch

发表时间：2021年7月

引用：Guo J, Han K, Wu H, et al. Cmt: Convolutional neural networks meet vision transformers[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2022: 12175-12185.

引用数：67

1. 简介

1.1 摘要

Vision Transformer 已成功应用于图像识别任务，因为它们能够捕获图像中的远程依赖关系。然而，Transformer 和现有的卷积神经网络 (CNN) 在性能和计算成本上仍然存在差距。在本文中，我们的目标是解决这个问题并开发一个网络，该网络不仅可以胜过传统的 Transformer，还可以胜过高性能卷积模型。

我们提出了一种新的基于 Transformer 的混合网络，利用变压器来捕获远程依赖关系，并利用 CNN 对局部特征进行建模。此外，我们对其进行缩放以获得一系列模型，称为 CMT，与以前的基于卷积和 Transformer 的模型相比，获得了更好的准确性和效率。

特别是，我们的 CMT-S 在 ImageNet 上实现了 83.5% 的 top-1 准确率，而在 FLOP 上分别比现有的 DeiT 和 EfficientNet 小 14 倍和 2 倍。所提出的 CMT-S 在 CIFAR10 (99.2%)、CIFAR100 (91.7%)、Flowers (98.7%) 和其他具有挑战性的视觉数据集如 COCO (44.3% mAP) 上也能很好地推广，而且计算成本要低得多。

2. 网络

2.1 整体架构

• 首先，输入 Image 进入 CMT Stem，CMT Stem 架构是一个 3×3 卷积、步幅为 2 和一个输出通道为 32 的茎架构来减小输入图像的大小，后接的是另外两个步幅为 1 的 3×3 卷积以获得更好的局部 信息
• 然后，$2\ast 2$ Conv stride=2 接 CMT Block*3，重复 4 次后 + 全局平均池化 + 全连接 + softmax 的1000 路分类

2.2 CMT Block

2.3 Lightweight Multi-head Self-attention

原始的self-attention模块中，输入 X 被线性变换为 query、key、value 再进行计算，运算成本高

此模块主要功能就是使用深度卷积计算代替了 key 和 value 的计算，从而减轻了计算开销，具体计算过程，可以看一下原文进行参考
$$\mathrm{LightAttn}(\mathbf{Q},\mathbf{K},\mathbf{V})=\mathrm{Softmax}\left(\frac{\mathbf{Q}{\mathbf{K}}^{\prime T}}{\sqrt{d_k}}+\mathbf{B}\right){\mathbf{V}}^{\prime }$$

3. 代码

# 2022.06.28-Changed for building CMT
#            Huawei Technologies Co., Ltd. 
# Author: Jianyuan Guo (jyguo@pku.edu.cn)

import math
import logging
from functools import partial
from collections import OrderedDict

import torch
import torch.nn as nn
import torch.nn.functional as F

from timm.data import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD
from timm.models.helpers import load_pretrained
from timm.models.layers import DropPath, to_2tuple, trunc_normal_
from timm.models.resnet import resnet26d, resnet50d
from timm.models.registry import register_model

_logger = logging.getLogger(__name__)


def _cfg(url='', **kwargs):
    return {
        'url': url,
        'num_classes': 1000, 'input_size': (3, 224, 224), 'pool_size': None,
        'crop_pct': .9, 'interpolation': 'bicubic',
        'mean': IMAGENET_DEFAULT_MEAN, 'std': IMAGENET_DEFAULT_STD,
        'first_conv': 'patch_embed.proj', 'classifier': 'head',
        **kwargs
    }


# A memory-efficient implementation of Swish function
class SwishImplementation(torch.autograd.Function):
    @staticmethod
    def forward(ctx, i):
        result = i * torch.sigmoid(i)
        ctx.save_for_backward(i)
        return result

    @staticmethod
    def backward(ctx, grad_output):
        i = ctx.saved_tensors[0]
        sigmoid_i = torch.sigmoid(i)
        return grad_output * (sigmoid_i * (1 + i * (1 - sigmoid_i)))


class MemoryEfficientSwish(nn.Module):
    def forward(self, x):
        return SwishImplementation.apply(x)


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.conv1 = nn.Sequential(
            nn.Conv2d(in_features, hidden_features, 1, 1, 0, bias=True),
            nn.GELU(),
            nn.BatchNorm2d(hidden_features, eps=1e-5),
        )
        self.proj = nn.Conv2d(hidden_features, hidden_features, 3, 1, 1, groups=hidden_features)
        self.proj_act = nn.GELU()
        self.proj_bn = nn.BatchNorm2d(hidden_features, eps=1e-5)
        self.conv2 = nn.Sequential(
            nn.Conv2d(hidden_features, out_features, 1, 1, 0, bias=True),
            nn.BatchNorm2d(out_features, eps=1e-5),
        )
        self.drop = nn.Dropout(drop)

    def forward(self, x, H, W):
        B, N, C = x.shape
        x = x.permute(0, 2, 1).reshape(B, C, H, W)
        x = self.conv1(x)
        x = self.drop(x)
        x = self.proj(x) + x
        x = self.proj_act(x)
        x = self.proj_bn(x)
        x = self.conv2(x)
        x = x.flatten(2).permute(0, 2, 1)
        x = self.drop(x)
        return x


class Attention(nn.Module):
    def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None,
                 attn_drop=0., proj_drop=0., qk_ratio=1, sr_ratio=1):
        super().__init__()
        self.num_heads = num_heads
        head_dim = dim // num_heads
        self.scale = qk_scale or head_dim ** -0.5
        self.qk_dim = dim // qk_ratio

        self.q = nn.Linear(dim, self.qk_dim, bias=qkv_bias)
        self.k = nn.Linear(dim, self.qk_dim, bias=qkv_bias)
        self.v = nn.Linear(dim, dim, 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
        # Exactly same as PVTv1
        if self.sr_ratio > 1:
            self.sr = nn.Sequential(
                nn.Conv2d(dim, dim, kernel_size=sr_ratio, stride=sr_ratio, groups=dim, bias=True),
                nn.BatchNorm2d(dim, eps=1e-5),
            )

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

        if self.sr_ratio > 1:
            x_ = x.permute(0, 2, 1).reshape(B, C, H, W)
            x_ = self.sr(x_).reshape(B, C, -1).permute(0, 2, 1)
            k = self.k(x_).reshape(B, -1, self.num_heads, self.qk_dim // self.num_heads).permute(0, 2, 1, 3)
            v = self.v(x_).reshape(B, -1, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
        else:
            k = self.k(x).reshape(B, N, self.num_heads, self.qk_dim // self.num_heads).permute(0, 2, 1, 3)
            v = self.v(x).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)

        attn = (q @ k.transpose(-2, -1)) * self.scale + relative_pos
        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


class Block(nn.Module):
    def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,
                 drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, qk_ratio=1, sr_ratio=1):
        super().__init__()
        self.norm1 = norm_layer(dim)
        self.attn = Attention(
            dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale,
            attn_drop=attn_drop, proj_drop=drop, qk_ratio=qk_ratio, sr_ratio=sr_ratio)
        # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
        self.norm2 = norm_layer(dim)
        mlp_hidden_dim = int(dim * mlp_ratio)
        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
        self.proj = nn.Conv2d(dim, dim, 3, 1, 1, groups=dim)

    def forward(self, x, H, W, relative_pos):
        B, N, C = x.shape
        cnn_feat = x.permute(0, 2, 1).reshape(B, C, H, W)
        x = self.proj(cnn_feat) + cnn_feat
        x = x.flatten(2).permute(0, 2, 1)
        x = x + self.drop_path(self.attn(self.norm1(x), H, W, relative_pos))
        x = x + self.drop_path(self.mlp(self.norm2(x), H, W))
        return x


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)
        num_patches = (img_size[1] // patch_size[1]) * (img_size[0] // patch_size[0])

        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.img_size = img_size
        self.patch_size = patch_size
        self.num_patches = num_patches

        self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)
        self.norm = nn.LayerNorm(embed_dim)

    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]})."
        x = self.proj(x).flatten(2).transpose(1, 2)
        x = self.norm(x)

        H, W = H // self.patch_size[0], W // self.patch_size[1]
        return x, (H, W)


class CMT(nn.Module):
    def __init__(self, img_size=224, in_chans=3, num_classes=1000, embed_dims=[46, 92, 184, 368], stem_channel=16,
                 fc_dim=1280,
                 num_heads=[1, 2, 4, 8], mlp_ratios=[3.6, 3.6, 3.6, 3.6], qkv_bias=True, qk_scale=None,
                 representation_size=None,
                 drop_rate=0., attn_drop_rate=0., drop_path_rate=0., hybrid_backbone=None, norm_layer=None,
                 depths=[2, 2, 10, 2], qk_ratio=1, sr_ratios=[8, 4, 2, 1], dp=0.1):
        super().__init__()
        self.num_classes = num_classes
        self.num_features = self.embed_dim = embed_dims[-1]
        norm_layer = norm_layer or partial(nn.LayerNorm, eps=1e-6)

        self.stem_conv1 = nn.Conv2d(3, stem_channel, kernel_size=3, stride=2, padding=1, bias=True)
        self.stem_relu1 = nn.GELU()
        self.stem_norm1 = nn.BatchNorm2d(stem_channel, eps=1e-5)

        self.stem_conv2 = nn.Conv2d(stem_channel, stem_channel, kernel_size=3, stride=1, padding=1, bias=True)
        self.stem_relu2 = nn.GELU()
        self.stem_norm2 = nn.BatchNorm2d(stem_channel, eps=1e-5)

        self.stem_conv3 = nn.Conv2d(stem_channel, stem_channel, kernel_size=3, stride=1, padding=1, bias=True)
        self.stem_relu3 = nn.GELU()
        self.stem_norm3 = nn.BatchNorm2d(stem_channel, eps=1e-5)

        self.patch_embed_a = PatchEmbed(
            img_size=img_size // 2, patch_size=2, in_chans=stem_channel, embed_dim=embed_dims[0])
        self.patch_embed_b = PatchEmbed(
            img_size=img_size // 4, patch_size=2, in_chans=embed_dims[0], embed_dim=embed_dims[1])
        self.patch_embed_c = PatchEmbed(
            img_size=img_size // 8, patch_size=2, in_chans=embed_dims[1], embed_dim=embed_dims[2])
        self.patch_embed_d = PatchEmbed(
            img_size=img_size // 16, patch_size=2, in_chans=embed_dims[2], embed_dim=embed_dims[3])

        self.relative_pos_a = nn.Parameter(torch.randn(
            num_heads[0], self.patch_embed_a.num_patches,
            self.patch_embed_a.num_patches // sr_ratios[0] // sr_ratios[0]))
        self.relative_pos_b = nn.Parameter(torch.randn(
            num_heads[1], self.patch_embed_b.num_patches,
            self.patch_embed_b.num_patches // sr_ratios[1] // sr_ratios[1]))
        self.relative_pos_c = nn.Parameter(torch.randn(
            num_heads[2], self.patch_embed_c.num_patches,
            self.patch_embed_c.num_patches // sr_ratios[2] // sr_ratios[2]))
        self.relative_pos_d = nn.Parameter(torch.randn(
            num_heads[3], self.patch_embed_d.num_patches,
            self.patch_embed_d.num_patches // sr_ratios[3] // sr_ratios[3]))

        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))]  # stochastic depth decay rule
        cur = 0
        self.blocks_a = nn.ModuleList([
            Block(
                dim=embed_dims[0], num_heads=num_heads[0], mlp_ratio=mlp_ratios[0], qkv_bias=qkv_bias,
                qk_scale=qk_scale, drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i],
                norm_layer=norm_layer, qk_ratio=qk_ratio, sr_ratio=sr_ratios[0])
            for i in range(depths[0])])
        cur += depths[0]
        self.blocks_b = nn.ModuleList([
            Block(
                dim=embed_dims[1], num_heads=num_heads[1], mlp_ratio=mlp_ratios[1], qkv_bias=qkv_bias,
                qk_scale=qk_scale, drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i],
                norm_layer=norm_layer, qk_ratio=qk_ratio, sr_ratio=sr_ratios[1])
            for i in range(depths[1])])
        cur += depths[1]
        self.blocks_c = nn.ModuleList([
            Block(
                dim=embed_dims[2], num_heads=num_heads[2], mlp_ratio=mlp_ratios[2], qkv_bias=qkv_bias,
                qk_scale=qk_scale, drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i],
                norm_layer=norm_layer, qk_ratio=qk_ratio, sr_ratio=sr_ratios[2])
            for i in range(depths[2])])
        cur += depths[2]
        self.blocks_d = nn.ModuleList([
            Block(
                dim=embed_dims[3], num_heads=num_heads[3], mlp_ratio=mlp_ratios[3], qkv_bias=qkv_bias,
                qk_scale=qk_scale, drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i],
                norm_layer=norm_layer, qk_ratio=qk_ratio, sr_ratio=sr_ratios[3])
            for i in range(depths[3])])

        # Representation layer
        if representation_size:
            self.num_features = representation_size
            self.pre_logits = nn.Sequential(OrderedDict([
                ('fc', nn.Linear(self.embed_dim, representation_size)),
                ('act', nn.Tanh())
            ]))
        else:
            self.pre_logits = nn.Identity()

        # Classifier head
        self._fc = nn.Conv2d(embed_dims[-1], fc_dim, kernel_size=1)
        self._bn = nn.BatchNorm2d(fc_dim, eps=1e-5)
        self._swish = MemoryEfficientSwish()
        self._avg_pooling = nn.AdaptiveAvgPool2d(1)
        self._drop = nn.Dropout(dp)
        self.head = nn.Linear(fc_dim, num_classes) if num_classes > 0 else nn.Identity()
        self.apply(self._init_weights)

    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.Conv2d):
            nn.init.kaiming_normal_(m.weight, mode='fan_out')
            if isinstance(m, nn.Conv2d) 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)
        elif isinstance(m, nn.BatchNorm2d):
            nn.init.constant_(m.bias, 0)
            nn.init.constant_(m.weight, 1.0)

    def update_temperature(self):
        for m in self.modules():
            if isinstance(m, Attention):
                m.update_temperature()

    @torch.jit.ignore
    def no_weight_decay(self):
        return {'pos_embed', 'cls_token'}

    def get_classifier(self):
        return self.head

    def reset_classifier(self, num_classes, global_pool=''):
        self.num_classes = num_classes
        self.head = nn.Linear(self.embed_dims[-1], num_classes) if num_classes > 0 else nn.Identity()

    def forward_features(self, x):
        B = x.shape[0]
        x = self.stem_conv1(x)
        x = self.stem_relu1(x)
        x = self.stem_norm1(x)

        x = self.stem_conv2(x)
        x = self.stem_relu2(x)
        x = self.stem_norm2(x)

        x = self.stem_conv3(x)
        x = self.stem_relu3(x)
        x = self.stem_norm3(x)

        x, (H, W) = self.patch_embed_a(x)
        for i, blk in enumerate(self.blocks_a):
            x = blk(x, H, W, self.relative_pos_a)

        x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
        x, (H, W) = self.patch_embed_b(x)
        for i, blk in enumerate(self.blocks_b):
            x = blk(x, H, W, self.relative_pos_b)

        x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
        x, (H, W) = self.patch_embed_c(x)
        for i, blk in enumerate(self.blocks_c):
            x = blk(x, H, W, self.relative_pos_c)

        x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
        x, (H, W) = self.patch_embed_d(x)
        for i, blk in enumerate(self.blocks_d):
            x = blk(x, H, W, self.relative_pos_d)

        B, N, C = x.shape
        x = self._fc(x.permute(0, 2, 1).reshape(B, C, H, W))
        x = self._bn(x)
        x = self._swish(x)
        x = self._avg_pooling(x).flatten(start_dim=1)
        x = self._drop(x)
        x = self.pre_logits(x)
        return x

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


def resize_pos_embed(posemb, posemb_new):
    # Rescale the grid of position embeddings when loading from state_dict. Adapted from
    # https://github.com/google-research/vision_transformer/blob/00883dd691c63a6830751563748663526e811cee/vit_jax/checkpoint.py#L224
    _logger.info('Resized position embedding: %s to %s', posemb.shape, posemb_new.shape)
    ntok_new = posemb_new.shape[1]
    if True:
        posemb_tok, posemb_grid = posemb[:, :1], posemb[0, 1:]
        ntok_new -= 1
    else:
        posemb_tok, posemb_grid = posemb[:, :0], posemb[0]
    gs_old = int(math.sqrt(len(posemb_grid)))
    gs_new = int(math.sqrt(ntok_new))
    _logger.info('Position embedding grid-size from %s to %s', gs_old, gs_new)
    posemb_grid = posemb_grid.reshape(1, gs_old, gs_old, -1).permute(0, 3, 1, 2)
    posemb_grid = F.interpolate(posemb_grid, size=(gs_new, gs_new), mode='bilinear')
    posemb_grid = posemb_grid.permute(0, 2, 3, 1).reshape(1, gs_new * gs_new, -1)
    posemb = torch.cat([posemb_tok, posemb_grid], dim=1)
    return posemb


def checkpoint_filter_fn(state_dict, model):
    """ convert patch embedding weight from manual patchify + linear proj to conv"""
    out_dict = {}
    if 'model' in state_dict:
        # For deit models
        state_dict = state_dict['model']
    for k, v in state_dict.items():
        if 'patch_embed.proj.weight' in k and len(v.shape) < 4:
            # For old models that I trained prior to conv based patchification
            O, I, H, W = model.patch_embed.proj.weight.shape
            v = v.reshape(O, -1, H, W)
        elif k == 'pos_embed' and v.shape != model.pos_embed.shape:
            # To resize pos embedding when using model at different size from pretrained weights
            v = resize_pos_embed(v, model.pos_embed)
        out_dict[k] = v
    return out_dict


def _create_cmt_model(pretrained=False, distilled=False, **kwargs):
    default_cfg = _cfg()
    default_num_classes = default_cfg['num_classes']
    default_img_size = default_cfg['input_size'][-1]

    num_classes = kwargs.pop('num_classes', default_num_classes)
    img_size = kwargs.pop('img_size', default_img_size)
    repr_size = kwargs.pop('representation_size', None)
    if repr_size is not None and num_classes != default_num_classes:
        # Remove representation layer if fine-tuning. This may not always be the desired action,
        # but I feel better than doing nothing by default for fine-tuning. Perhaps a better interface?
        _logger.warning("Removing representation layer for fine-tuning.")
        repr_size = None

    model = CMT(img_size=img_size, num_classes=num_classes, representation_size=repr_size, **kwargs)
    model.default_cfg = default_cfg

    if pretrained:
        load_pretrained(
            model, num_classes=num_classes, in_chans=kwargs.get('in_chans', 3),
            filter_fn=partial(checkpoint_filter_fn, model=model))
    return model


@register_model
def cmt_ti(pretrained=False, **kwargs):
    """
    CMT-Tiny
    """
    model_kwargs = dict(qkv_bias=True, **kwargs)
    model = _create_cmt_model(pretrained=pretrained, **model_kwargs)
    return model


@register_model
def cmt_xs(pretrained=False, **kwargs):
    """
    CMT-XS: dim x 0.9, depth x 0.8, input 192
    """
    model_kwargs = dict(
        qkv_bias=True, embed_dims=[52, 104, 208, 416], stem_channel=16, num_heads=[1, 2, 4, 8],
        depths=[3, 3, 12, 3], mlp_ratios=[3.77, 3.77, 3.77, 3.77], qk_ratio=1, sr_ratios=[8, 4, 2, 1], **kwargs)
    model = _create_cmt_model(pretrained=pretrained, **model_kwargs)
    return model


@register_model
def cmt_s(pretrained=False, **kwargs):
    """
    CMT-Small
    """
    model_kwargs = dict(
        qkv_bias=True, embed_dims=[64, 128, 256, 512], stem_channel=32, num_heads=[1, 2, 4, 8],
        depths=[3, 3, 16, 3], mlp_ratios=[4, 4, 4, 4], qk_ratio=1, sr_ratios=[8, 4, 2, 1], **kwargs)
    model = _create_cmt_model(pretrained=pretrained, **model_kwargs)
    return model


@register_model
def cmt_b(pretrained=False, **kwargs):
    """
    CMT-Base
    """
    model_kwargs = dict(
        qkv_bias=True, embed_dims=[76, 152, 304, 608], stem_channel=38, num_heads=[1, 2, 4, 8],
        depths=[4, 4, 20, 4], mlp_ratios=[4, 4, 4, 4], qk_ratio=1, sr_ratios=[8, 4, 2, 1], dp=0.3, **kwargs)
    model = _create_cmt_model(pretrained=pretrained, **model_kwargs)
    return model

if __name__ == '__main__':
    x=torch.randn(1,3,224,224)
    model=cmt_ti(num_classes=10)
    y=model(x)
    print(y.shape)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476

参考资料

CVPR22 ｜CMT：CNN和Transformer的高效结合（开源） - 知乎 (zhihu.com)

Transformer（十五）CMT - 知乎 (zhihu.com)