【CVPR2022 ConvNeXt网络结构详解】 A ConvNet for the 2020s

FAIR ConvNext主体网络结构代码详解，提供多个模型版本以满足不同应用场景的使用。

paper：https://arxiv.org/pdf/2201.03545.pdf

Source Code：GitHub - facebookresearch/ConvNeXt: Code release for ConvNeXt model

url中提供ImageNet 1K以及22K的权重链接，大家合理使用即可。具体性能表现见如上述github链接中所示。

# Copyright (c) Meta Platforms, Inc. and affiliates.

# All rights reserved.

# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.

import torch
import torch.nn as nn
import torch.nn.functional as F
from timm.models.layers import trunc_normal_, DropPath
from timm.models.registry import register_model

#---------------------------------------------------------------------------------#
#   LayerNorm 支持两种形式channels_last (default) or channels_first
#   channels_last   对应具有形状的输入(batch_size, height, width, channels) 
#   channels_first  对应具有形状的输入(batch_size, channels, height, width)
#   我们这里默认为channels_last
#---------------------------------------------------------------------------------#
class LayerNorm(nn.Module):
    def __init__(self, normalized_shape, eps=1e-6, data_format="channels_last"):
        super().__init__()

        self.weight           = nn.Parameter(torch.ones(normalized_shape))
        self.bias             = nn.Parameter(torch.zeros(normalized_shape))
        self.eps              = eps
        self.data_format      = data_format
        self.normalized_shape = (normalized_shape, )

        if self.data_format not in ["channels_last", "channels_first"]:
            raise NotImplementedError
    
    def forward(self, x):
        #-----------------------------------------------#
        # 当为默认形式的时候我们直接调用torch自带的layer_norm
        #-----------------------------------------------#
        if self.data_format == "channels_last":
            return F.layer_norm(x, self.normalized_shape, self.weight, self.bias, self.eps)

        elif self.data_format == "channels_first":
            u = x.mean(1, keepdim=True)
            s = (x - u).pow(2).mean(1, keepdim=True)
            x = (x - u) / torch.sqrt(s + self.eps)
            x = self.weight[:, None, None] * x + self.bias[:, None, None]
            return x

#--------------------------------------------------------------------------------------------------------------#
#   ConvNeXt Block有两种等效的实现:
#   (1) DwConv -> LayerNorm (channels_first) -> 1x1 Conv -> GELU -> 1x1 Conv; all in (N, C, H, W)
#   (2) DwConv -> Permute to (N, H, W, C); LayerNorm (channels_last) -> Linear -> GELU -> Linear; Permute back
#   代码中使用（2），因为这个在PyTorch中稍微快一点
#   args:
#   dim (int): Number of input channels.
#   drop_path (float): Stochastic depth rate. Default: 0.0
#   layer_scale_init_value (float): Init value for Layer Scale. Default: 1e-6.
#--------------------------------------------------------------------------------------------------------------#
class Block(nn.Module):
    def __init__(self, dim, drop_path=0., layer_scale_init_value=1e-6):
        super().__init__()
        #--------------------------#
        #   7x7的逐层卷积
        #--------------------------#
        self.dwconv     = nn.Conv2d(dim, dim, kernel_size=7, padding=3, groups=dim)
        self.norm       = LayerNorm(dim, eps=1e-6)
        #--------------------------#
        #   利用全连接层代替1x1卷积
        #--------------------------#
        self.pwconv1    = nn.Linear(dim, 4 * dim)
        self.act        = nn.GELU()
        #--------------------------#
        #   利用全连接层代替1x1卷积
        #--------------------------#
        self.pwconv2    = nn.Linear(4 * dim, dim)
        #--------------------------#
        #   加入缩放系数
        #--------------------------#
        self.gamma      = nn.Parameter(layer_scale_init_value * torch.ones((dim)), requires_grad=True) if layer_scale_init_value > 0 else None
        #--------------------------#
        #   加入Drop_path正则化
        #--------------------------#
        self.drop_path  = DropPath(drop_path) if drop_path > 0. else nn.Identity()

    def forward(self, x):
        input = x
        #--------------------------#
        #   7x7的逐层卷积
        #--------------------------#
        x = self.dwconv(x)
        x = x.permute(0, 2, 3, 1) # (N, C, H, W) -> (N, H, W, C)
        x = self.norm(x)
        #--------------------------#
        #   利用全连接层代替1x1卷积
        #--------------------------#
        x = self.pwconv1(x)
        x = self.act(x)
        #--------------------------#
        #   利用全连接层代替1x1卷积
        #--------------------------#
        x = self.pwconv2(x)
        #--------------------------#
        #   加入缩放系数
        #--------------------------#
        if self.gamma is not None:
            x = self.gamma * x
        x = x.permute(0, 3, 1, 2) # (N, H, W, C) -> (N, C, H, W)
        #--------------------------#
        #   加入Drop_path正则化
        #--------------------------#
        x = input + self.drop_path(x)
        return x

#--------------------------------------------------------------------------------------------------------#
#   ConvNeXt
#     Args:
#         in_chans (int): Number of input image channels. Default: 3
#         num_classes (int): Number of classes for classification head. Default: 1000
#         depths (tuple(int)): Number of blocks at each stage. Default: [3, 3, 9, 3]
#         dims (int): Feature dimension at each stage. Default: [96, 192, 384, 768]
#         drop_path_rate (float): Stochastic depth rate. Default: 0.
#         layer_scale_init_value (float): Init value for Layer Scale. Default: 1e-6.
#         head_init_scale (float): Init scaling value for classifier weights and biases. Default: 1.
#-----------------------------------------------------------------------------------------------------=---#
class ConvNeXt(nn.Module):
    def __init__(
        self, in_chans=3, num_classes=1000, depths=[3, 3, 9, 3], dims=[96, 192, 384, 768], 
        drop_path_rate=0., layer_scale_init_value=1e-6, head_init_scale=1.,):
        super().__init__()

        self.downsample_layers = nn.ModuleList()
        #--------------------------------------------------#
        #   bs, 3, 224, 224 -> bs, 96, 56, 56
        #--------------------------------------------------#
        stem = nn.Sequential(
            nn.Conv2d(in_chans, dims[0], kernel_size=4, stride=4),
            LayerNorm(dims[0], eps=1e-6, data_format="channels_first")
        )
        self.downsample_layers.append(stem)
        
        #--------------------------------------------------#
        #   定义三次下采样的过程
        #   利用步长为2x2，卷积核大小为2x2的卷积进行下采样
        #--------------------------------------------------#
        for i in range(3):
            downsample_layer = nn.Sequential(
                LayerNorm(dims[i], eps=1e-6, data_format="channels_first"),
                nn.Conv2d(dims[i], dims[i+1], kernel_size=2, stride=2),
            )
            self.downsample_layers.append(downsample_layer)

        #--------------------------------------------------#
        #   根据深度的不同,定义不同的drop率
        #--------------------------------------------------#
        self.stages = nn.ModuleList()
        dp_rates    = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))] 
        cur         = 0
        #--------------------------------------------------#
        #   整个ConvNeXt除了Stem外,存在四个Stage
        #   每个Stage里面是多个ConvNeXt Block的堆叠
        #--------------------------------------------------#
        for i in range(4):
            stage = nn.Sequential(
                *[Block(dim=dims[i], drop_path=dp_rates[cur + j], layer_scale_init_value=layer_scale_init_value) for j in range(depths[i])]
            )
            self.stages.append(stage)
            cur += depths[i]

        self.norm = nn.LayerNorm(dims[-1], eps=1e-6)  # final norm layer
        self.head = nn.Linear(dims[-1], num_classes)

        self.apply(self._init_weights)
        self.head.weight.data.mul_(head_init_scale)
        self.head.bias.data.mul_(head_init_scale)

    def _init_weights(self, m):
        if isinstance(m, (nn.Conv2d, nn.Linear)):
            trunc_normal_(m.weight, std=.02)
            nn.init.constant_(m.bias, 0)

    def forward_features(self, x):
        for i in range(4):
            x = self.downsample_layers[i](x)
            x = self.stages[i](x)

        return self.norm(x.mean([-2, -1]))  # global average pooling, (N, C, H, W) -> (N, C)

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

        return out


model_urls = {
    "convnext_tiny_1k": "https://dl.fbaipublicfiles.com/convnext/convnext_tiny_1k_224_ema.pth",
    "convnext_small_1k": "https://dl.fbaipublicfiles.com/convnext/convnext_small_1k_224_ema.pth",
    "convnext_base_1k": "https://dl.fbaipublicfiles.com/convnext/convnext_base_1k_224_ema.pth",
    "convnext_large_1k": "https://dl.fbaipublicfiles.com/convnext/convnext_large_1k_224_ema.pth",
    "convnext_tiny_22k": "https://dl.fbaipublicfiles.com/convnext/convnext_tiny_22k_224.pth",
    "convnext_small_22k": "https://dl.fbaipublicfiles.com/convnext/convnext_small_22k_224.pth",
    "convnext_base_22k": "https://dl.fbaipublicfiles.com/convnext/convnext_base_22k_224.pth",
    "convnext_large_22k": "https://dl.fbaipublicfiles.com/convnext/convnext_large_22k_224.pth",
    "convnext_xlarge_22k": "https://dl.fbaipublicfiles.com/convnext/convnext_xlarge_22k_224.pth",
}


@register_model
def convnext_tiny(pretrained=False, in_22k=False, **kwargs):
    model = ConvNeXt(depths=[3, 3, 9, 3], dims=[96, 192, 384, 768], **kwargs)
    if pretrained:
        url = model_urls['convnext_tiny_22k'] if in_22k else model_urls['convnext_tiny_1k']
        checkpoint = torch.hub.load_state_dict_from_url(url=url, map_location="cpu", check_hash=True)
        model.load_state_dict(checkpoint["model"])
    return model


@register_model
def convnext_small(pretrained=False, in_22k=False, **kwargs):
    model = ConvNeXt(depths=[3, 3, 27, 3], dims=[96, 192, 384, 768], **kwargs)
    if pretrained:
        url = model_urls['convnext_small_22k'] if in_22k else model_urls['convnext_small_1k']
        checkpoint = torch.hub.load_state_dict_from_url(url=url, map_location="cpu")
        model.load_state_dict(checkpoint["model"])
    return model


@register_model
def convnext_base(pretrained=False, in_22k=False, **kwargs):
    model = ConvNeXt(depths=[3, 3, 27, 3], dims=[128, 256, 512, 1024], **kwargs)
    if pretrained:
        url = model_urls['convnext_base_22k'] if in_22k else model_urls['convnext_base_1k']
        checkpoint = torch.hub.load_state_dict_from_url(url=url, map_location="cpu")
        model.load_state_dict(checkpoint["model"])
    return model


@register_model
def convnext_large(pretrained=False, in_22k=False, **kwargs):
    model = ConvNeXt(depths=[3, 3, 27, 3], dims=[192, 384, 768, 1536], **kwargs)
    if pretrained:
        url = model_urls['convnext_large_22k'] if in_22k else model_urls['convnext_large_1k']
        checkpoint = torch.hub.load_state_dict_from_url(url=url, map_location="cpu")
        model.load_state_dict(checkpoint["model"])
    return model


@register_model
def convnext_xlarge(pretrained=False, in_22k=False, **kwargs):
    model = ConvNeXt(depths=[3, 3, 27, 3], dims=[256, 512, 1024, 2048], **kwargs)
    if pretrained:
        assert in_22k, "only ImageNet-22K pre-trained ConvNeXt-XL is available; please set in_22k=True"
        url = model_urls['convnext_xlarge_22k']
        checkpoint = torch.hub.load_state_dict_from_url(url=url, map_location="cpu")
        model.load_state_dict(checkpoint["model"])
    return model