【RT-DETR有效改进】 | 主干篇 | RevColV1可逆列网络（特征解耦助力小目标检测）

前言

大家好，这里是RT-DETR有效涨点专栏。

本专栏的内容为根据ultralytics版本的RT-DETR进行改进，内容持续更新，每周更新文章数量3-10篇。

专栏以ResNet18、ResNet50为基础修改版本，同时修改内容也支持ResNet32、ResNet101和PPHGNet版本，其中ResNet为RT-DETR官方版本1：1移植过来的，参数量基本保持一致(误差很小很小)，不同于ultralytics仓库版本的ResNet官方版本，同时ultralytics仓库的一些参数是和RT-DETR相冲的所以我也是会教大家调好一些参数和代码，真正意义上的跑ultralytics的和RT-DETR官方版本的无区别。

👑欢迎大家订阅本专栏，一起学习RT-DETR👑

一、本文介绍

本文给大家带来的是主干网络RevColV1，翻译过来就是可逆列网络去发表于ICLR2022，其是一种新型的神经网络设计(和以前的网络结构的传播方式不太一样)，由多个子网络（列）通过多级可逆连接组成。这种设计允许在前向传播过程中特征解耦，保持总信息无压缩或丢弃。其非常适合数据集庞大的目标检测任务，数据集数量越多其效果性能越好，亲测在包含1000个图片的数据集上其涨点效果就非常明显了，大家可以多动手尝试，其RevColV2的论文同时已经发布如果代码开源我也会第一时间给大家上传。

推荐指数：⭐⭐⭐⭐⭐

涨点效果：⭐⭐⭐⭐⭐

专栏链接：RT-DETR剑指论文专栏，持续复现各种顶会内容——论文收割机RT-DETR

一、本文介绍

二、RevColV1的框架原理

2.1 RevColV1的基本原理

2.1.1 可逆连接设计

2.1.2 特征解耦

2.2 RevColV1的表现

三、RevColV1的核心代码

四、手把手教你添加RevColV1网络结构

4.1 修改一

4.2 修改二

4.3 修改三

4.4 修改四

4.5 修改五

4.6 修改六

4.7 修改七

4.8 修改八

4.9 RT-DETR不能打印计算量问题的解决

4.10 可选修改

五、RevColV1的yaml文件

5.1 yaml文件

5.2 运行文件

5.3 成功训练截图

六、全文总结

二、RevColV1的框架原理

官方论文地址： 官方论文地址

官方代码地址： 官方代码地址

2.1 RevColV1的基本原理

RevCol的主要原理和思想是利用可逆连接来设计网络结构，允许信息在网络的不同分支（列）间自由流动而不丢失。这种多列结构在前向传播过程中逐渐解耦特征，并保持全部信息，而不是进行压缩或舍弃。这样的设计提高了网络在图像分类、对象检测和语义分割等计算机视觉任务中的表现，尤其是在参数量大和数据集大时。

RevCol的创新点我将其总结为以下几点：

1. 可逆连接设计：通过多个子网络（列）间的可逆连接，保证信息在前向传播过程中不丢失。
2. 特征解耦：在每个列中，特征逐渐被解耦，保持总信息而非压缩或舍弃。
3. 适用于大型数据集和参数：在大型数据集和高参数预算下表现出色。
4. 跨模型应用：可作为宏架构方式，应用于变换器或其他神经网络，改善计算机视觉和NLP任务的性能。

简单总结：RevCol通过其独特的多列结构和可逆连接设计，使得网络能够在处理信息时保持完整性，提高特征处理的效率。这种架构在数据丰富且复杂的情况下尤为有效，且可灵活应用于不同类型的神经网络模型中。

其中的创新点第四点不用叙述了，网络结构可以应用于我们的YOLOv8就是最好的印证。

这是论文中的图片1，展示了传统单列网络（a）与RevCol（b）的信息传播对比。在图(a)中，信息通过一个接一个的层线性传播，每层处理后传递给下一层直至输出。而在图(b)中，RevCol通过多个并行列（Col 1 到 Col N）处理信息，其中可逆连接（蓝色曲线）允许信息在列间传递，保持低级别和语义级别的信息传播。这种结构有助于整个网络维持更丰富的信息，并且每个列都能从其他列中学习到信息，增强了特征的表达和网络的学习能力(但是这种做法导致模型的参数量非常巨大，而且训练速度缓慢计算量比较大)。

2.1.1 可逆连接设计

在RevCol中的可逆连接设计允许多个子网络（称为列）之间进行信息的双向流动。这意味着在前向传播的过程中，每一列都能接收到前一列的信息，并将自己的处理结果传递给下一列，同时能够保留传递过程中的所有信息。这种设计避免了在传统的深度网络中常见的信息丢失问题，特别是在网络层次较深时。因此，RevCol可以在深层网络中维持丰富的特征表示，从而提高了模型对数据的表示能力和学习效率。

这张图片展示了RevCol网络的不同组成部分和信息流动方式。

图 (a) 展示了RevNet中的一个可逆单元，标识了不同时间步长的状态。
图 (b) 展示了多级可逆单元，所有输入在不同级别上进行信息交换。
图 (c) 提供了整个可逆列网络架构的概览，其中包含了简化的多级可逆单元。

整个设计允许信息在网络的不同层级和列之间自由流动，而不会丢失任何信息，这对于深层网络的学习和特征提取是非常有益的(我觉得这里有点类似于Neck部分允许层级之间相互交流信息)。

2.1.2 特征解耦

特征解耦是指在RevCol网络的每个子网络（列）中，特征通过可逆连接传递，同时独立地进行处理和学习。这样，每个列都能保持输入信息的完整性，而不会像传统的深度网络那样，在层与层之间传递时压缩或丢弃信息。随着信息在列中的前进，特征之间的关联性逐渐减弱（解耦），使得网络能够更细致地捕捉并强调重要的特征，这有助于提高模型在复杂任务上的性能和泛化能力。

这张图展示了RevCol网络的一个级别（Level l）的微观设计，以及特征融合模块（Fusion Block）的设计。在图（a）中，展示了ConvNeXt级别的标准结构，包括下采样块和残差块。图（b）中的RevCol级别包含了融合模块、残差块和可逆操作。这里的特征解耦是通过融合模块实现的，该模块接收相邻级别的特征图 $X_{t-1}$ , $X_{t-m+1}$ 作为输入，并将它们融合以生成新的特征表示。这样，不同级别的特征在融合过程中被解耦，每个级别维持其信息而不压缩或舍弃。图（c）详细描述了融合模块的内部结构，它通过上采样和下采样操作处理不同分辨率的特征图，然后将它们线性叠加，形成为ConvNeXt块提供的特征。这种设计让特征在不同分辨率间流动时进行有效融合。

2.2 RevColV1的表现

这张图片展示了伴随着FLOPs的增长TOP1的准确率情况，可以看出RevColV1伴随着FLOPs的增加效果逐渐明显。

三、RevColV1的核心代码

下面的代码是RevColV1的全部代码，其中包含多个版本，但是大家需要注意这个模型训练非常耗时，参数量非常大，但是其特点就是参数量越大效果越好。其使用方式看章节四。

# --------------------------------------------------------
# Reversible Column Networks
# Copyright (c) 2022 Megvii Inc.
# Licensed under The Apache License 2.0 [see LICENSE for details]
# Written by Yuxuan Cai
# --------------------------------------------------------
from typing import Tuple, Any, List
from timm.models.layers import trunc_normal_
import torch
import torch.nn as nn
import torch.nn.functional as F
from timm.models.layers import DropPath


class UpSampleConvnext(nn.Module):
    def __init__(self, ratio, inchannel, outchannel):
        super().__init__()
        self.ratio = ratio
        self.channel_reschedule = nn.Sequential(
            # LayerNorm(inchannel, eps=1e-6, data_format="channels_last"),
            nn.Linear(inchannel, outchannel),
            LayerNorm(outchannel, eps=1e-6, data_format="channels_last"))
        self.upsample = nn.Upsample(scale_factor=2 ** ratio, mode='nearest')

    def forward(self, x):
        x = x.permute(0, 2, 3, 1)
        x = self.channel_reschedule(x)
        x = x = x.permute(0, 3, 1, 2)

        return self.upsample(x)


class LayerNorm(nn.Module):
    r""" LayerNorm that supports two data formats: channels_last (default) or channels_first.
    The ordering of the dimensions in the inputs. channels_last corresponds to inputs with
    shape (batch_size, height, width, channels) while channels_first corresponds to inputs
    with shape (batch_size, channels, height, width).
    """

    def __init__(self, normalized_shape, eps=1e-6, data_format="channels_first", elementwise_affine=True):
        super().__init__()
        self.elementwise_affine = elementwise_affine
        if elementwise_affine:
            self.weight = nn.Parameter(torch.ones(normalized_shape))
            self.bias = nn.Parameter(torch.zeros(normalized_shape))
        self.eps = eps
        self.data_format = data_format
        if self.data_format not in ["channels_last", "channels_first"]:
            raise NotImplementedError
        self.normalized_shape = (normalized_shape,)

    def forward(self, x):
        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)
            if self.elementwise_affine:
                x = self.weight[:, None, None] * x + self.bias[:, None, None]
            return x


class ConvNextBlock(nn.Module):
    r""" ConvNeXt Block. There are two equivalent implementations:
    (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
    We use (2) as we find it slightly faster in PyTorch

    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.
    """

    def __init__(self, in_channel, hidden_dim, out_channel, kernel_size=3, layer_scale_init_value=1e-6, drop_path=0.0):
        super().__init__()
        self.dwconv = nn.Conv2d(in_channel, in_channel, kernel_size=kernel_size, padding=(kernel_size - 1) // 2,
                                groups=in_channel)  # depthwise conv
        self.norm = nn.LayerNorm(in_channel, eps=1e-6)
        self.pwconv1 = nn.Linear(in_channel, hidden_dim)  # pointwise/1x1 convs, implemented with linear layers
        self.act = nn.GELU()
        self.pwconv2 = nn.Linear(hidden_dim, out_channel)
        self.gamma = nn.Parameter(layer_scale_init_value * torch.ones((out_channel)),
                                  requires_grad=True) if layer_scale_init_value > 0 else None
        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()

    def forward(self, x):
        input = x
        x = self.dwconv(x)
        x = x.permute(0, 2, 3, 1)  # (N, C, H, W) -> (N, H, W, C)
        x = self.norm(x)
        x = self.pwconv1(x)
        x = self.act(x)
        # print(f"x min: {x.min()}, x max: {x.max()}, input min: {input.min()}, input max: {input.max()}, x mean: {x.mean()}, x var: {x.var()}, ratio: {torch.sum(x>8)/x.numel()}")
        x = self.pwconv2(x)
        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)

        x = input + self.drop_path(x)
        return x


class Decoder(nn.Module):
    def __init__(self, depth=[2, 2, 2, 2], dim=[112, 72, 40, 24], block_type=None, kernel_size=3) -> None:
        super().__init__()
        self.depth = depth
        self.dim = dim
        self.block_type = block_type
        self._build_decode_layer(dim, depth, kernel_size)
        self.projback = nn.Sequential(
            nn.Conv2d(
                in_channels=dim[-1],
                out_channels=4 ** 2 * 3, kernel_size=1),
            nn.PixelShuffle(4),
        )

    def _build_decode_layer(self, dim, depth, kernel_size):
        normal_layers = nn.ModuleList()
        upsample_layers = nn.ModuleList()
        proj_layers = nn.ModuleList()

        norm_layer = LayerNorm

        for i in range(1, len(dim)):
            module = [self.block_type(dim[i], dim[i], dim[i], kernel_size) for _ in range(depth[i])]
            normal_layers.append(nn.Sequential(*module))
            upsample_layers.append(nn.Upsample(scale_factor=2, mode='bilinear', align_corners=True))
            proj_layers.append(nn.Sequential(
                nn.Conv2d(dim[i - 1], dim[i], 1, 1),
                norm_layer(dim[i]),
                nn.GELU()
            ))
        self.normal_layers = normal_layers
        self.upsample_layers = upsample_layers
        self.proj_layers = proj_layers

    def _forward_stage(self, stage, x):
        x = self.proj_layers[stage](x)
        x = self.upsample_layers[stage](x)
        return self.normal_layers[stage](x)

    def forward(self, c3):
        x = self._forward_stage(0, c3)  # 14
        x = self._forward_stage(1, x)  # 28
        x = self._forward_stage(2, x)  # 56
        x = self.projback(x)
        return x


class SimDecoder(nn.Module):
    def __init__(self, in_channel, encoder_stride) -> None:
        super().__init__()
        self.projback = nn.Sequential(
            LayerNorm(in_channel),
            nn.Conv2d(
                in_channels=in_channel,
                out_channels=encoder_stride ** 2 * 3, kernel_size=1),
            nn.PixelShuffle(encoder_stride),
        )

    def forward(self, c3):
        return self.projback(c3)

def get_gpu_states(fwd_gpu_devices) -> Tuple[List[int], List[torch.Tensor]]:
    # This will not error out if "arg" is a CPU tensor or a non-tensor type because
    # the conditionals short-circuit.
    fwd_gpu_states = []
    for device in fwd_gpu_devices:
        with torch.cuda.device(device):
            fwd_gpu_states.append(torch.cuda.get_rng_state())

    return fwd_gpu_states

def get_gpu_device(*args):

    fwd_gpu_devices = list(set(arg.get_device() for arg in args
                               if isinstance(arg, torch.Tensor) and arg.is_cuda))
    return fwd_gpu_devices

def set_device_states(fwd_cpu_state, devices, states) -> None:
    torch.set_rng_state(fwd_cpu_state)
    for device, state in zip(devices, states):
        with torch.cuda.device(device):
            torch.cuda.set_rng_state(state)

def detach_and_grad(inputs: Tuple[Any, ...]) -> Tuple[torch.Tensor, ...]:
    if isinstance(inputs, tuple):
        out = []
        for inp in inputs:
            if not isinstance(inp, torch.Tensor):
                out.append(inp)
                continue

            x = inp.detach()
            x.requires_grad = True
            out.append(x)
        return tuple(out)
    else:
        raise RuntimeError(
            "Only tuple of tensors is supported. Got Unsupported input type: ", type(inputs).__name__)

def get_cpu_and_gpu_states(gpu_devices):
    return torch.get_rng_state(), get_gpu_states(gpu_devices)

class ReverseFunction(torch.autograd.Function):
    @staticmethod
    def forward(ctx, run_functions, alpha, *args):
        l0, l1, l2, l3 = run_functions
        alpha0, alpha1, alpha2, alpha3 = alpha
        ctx.run_functions = run_functions
        ctx.alpha = alpha
        ctx.preserve_rng_state = True

        ctx.gpu_autocast_kwargs = {"enabled": torch.is_autocast_enabled(),
                                   "dtype": torch.get_autocast_gpu_dtype(),
                                   "cache_enabled": torch.is_autocast_cache_enabled()}
        ctx.cpu_autocast_kwargs = {"enabled": torch.is_autocast_cpu_enabled(),
                                   "dtype": torch.get_autocast_cpu_dtype(),
                                   "cache_enabled": torch.is_autocast_cache_enabled()}

        assert len(args) == 5
        [x, c0, c1, c2, c3] = args
        if type(c0) == int:
            ctx.first_col = True
        else:
            ctx.first_col = False
        with torch.no_grad():
            gpu_devices = get_gpu_device(*args)
            ctx.gpu_devices = gpu_devices
            ctx.cpu_states_0, ctx.gpu_states_0 = get_cpu_and_gpu_states(gpu_devices)
            c0 = l0(x, c1) + c0 * alpha0
            ctx.cpu_states_1, ctx.gpu_states_1 = get_cpu_and_gpu_states(gpu_devices)
            c1 = l1(c0, c2) + c1 * alpha1
            ctx.cpu_states_2, ctx.gpu_states_2 = get_cpu_and_gpu_states(gpu_devices)
            c2 = l2(c1, c3) + c2 * alpha2
            ctx.cpu_states_3, ctx.gpu_states_3 = get_cpu_and_gpu_states(gpu_devices)
            c3 = l3(c2, None) + c3 * alpha3
        ctx.save_for_backward(x, c0, c1, c2, c3)
        return x, c0, c1, c2, c3

    @staticmethod
    def backward(ctx, *grad_outputs):
        x, c0, c1, c2, c3 = ctx.saved_tensors
        l0, l1, l2, l3 = ctx.run_functions
        alpha0, alpha1, alpha2, alpha3 = ctx.alpha
        gx_right, g0_right, g1_right, g2_right, g3_right = grad_outputs
        (x, c0, c1, c2, c3) = detach_and_grad((x, c0, c1, c2, c3))

        with torch.enable_grad(), \
                torch.random.fork_rng(devices=ctx.gpu_devices, enabled=ctx.preserve_rng_state), \
                torch.cuda.amp.autocast(**ctx.gpu_autocast_kwargs), \
                torch.cpu.amp.autocast(**ctx.cpu_autocast_kwargs):

            g3_up = g3_right
            g3_left = g3_up * alpha3  ##shortcut
            set_device_states(ctx.cpu_states_3, ctx.gpu_devices, ctx.gpu_states_3)
            oup3 = l3(c2, None)
            torch.autograd.backward(oup3, g3_up, retain_graph=True)
            with torch.no_grad():
                c3_left = (1 / alpha3) * (c3 - oup3)  ## feature reverse
            g2_up = g2_right + c2.grad
            g2_left = g2_up * alpha2  ##shortcut

            (c3_left,) = detach_and_grad((c3_left,))
            set_device_states(ctx.cpu_states_2, ctx.gpu_devices, ctx.gpu_states_2)
            oup2 = l2(c1, c3_left)
            torch.autograd.backward(oup2, g2_up, retain_graph=True)
            c3_left.requires_grad = False
            cout3 = c3_left * alpha3  ##alpha3 update
            torch.autograd.backward(cout3, g3_up)

            with torch.no_grad():
                c2_left = (1 / alpha2) * (c2 - oup2)  ## feature reverse
            g3_left = g3_left + c3_left.grad if c3_left.grad is not None else g3_left
            g1_up = g1_right + c1.grad
            g1_left = g1_up * alpha1  ##shortcut

            (c2_left,) = detach_and_grad((c2_left,))
            set_device_states(ctx.cpu_states_1, ctx.gpu_devices, ctx.gpu_states_1)
            oup1 = l1(c0, c2_left)
            torch.autograd.backward(oup1, g1_up, retain_graph=True)
            c2_left.requires_grad = False
            cout2 = c2_left * alpha2  ##alpha2 update
            torch.autograd.backward(cout2, g2_up)

            with torch.no_grad():
                c1_left = (1 / alpha1) * (c1 - oup1)  ## feature reverse
            g0_up = g0_right + c0.grad
            g0_left = g0_up * alpha0  ##shortcut
            g2_left = g2_left + c2_left.grad if c2_left.grad is not None else g2_left  ## Fusion

            (c1_left,) = detach_and_grad((c1_left,))
            set_device_states(ctx.cpu_states_0, ctx.gpu_devices, ctx.gpu_states_0)
            oup0 = l0(x, c1_left)
            torch.autograd.backward(oup0, g0_up, retain_graph=True)
            c1_left.requires_grad = False
            cout1 = c1_left * alpha1  ##alpha1 update
            torch.autograd.backward(cout1, g1_up)

            with torch.no_grad():
                c0_left = (1 / alpha0) * (c0 - oup0)  ## feature reverse
            gx_up = x.grad  ## Fusion
            g1_left = g1_left + c1_left.grad if c1_left.grad is not None else g1_left  ## Fusion
            c0_left.requires_grad = False
            cout0 = c0_left * alpha0  ##alpha0 update
            torch.autograd.backward(cout0, g0_up)

        if ctx.first_col:
            return None, None, gx_up, None, None, None, None
        else:
            return None, None, gx_up, g0_left, g1_left, g2_left, g3_left



class Fusion(nn.Module):
    def __init__(self, level, channels, first_col) -> None:
        super().__init__()

        self.level = level
        self.first_col = first_col
        self.down = nn.Sequential(
            nn.Conv2d(channels[level - 1], channels[level], kernel_size=2, stride=2),
            LayerNorm(channels[level], eps=1e-6, data_format="channels_first"),
        ) if level in [1, 2, 3] else nn.Identity()
        if not first_col:
            self.up = UpSampleConvnext(1, channels[level + 1], channels[level]) if level in [0, 1, 2] else nn.Identity()

    def forward(self, *args):

        c_down, c_up = args

        if self.first_col:
            x = self.down(c_down)
            return x

        if self.level == 3:
            x = self.down(c_down)
        else:
            x = self.up(c_up) + self.down(c_down)
        return x


class Level(nn.Module):
    def __init__(self, level, channels, layers, kernel_size, first_col, dp_rate=0.0) -> None:
        super().__init__()
        countlayer = sum(layers[:level])
        expansion = 4
        self.fusion = Fusion(level, channels, first_col)
        modules = [ConvNextBlock(channels[level], expansion * channels[level], channels[level], kernel_size=kernel_size,
                                 layer_scale_init_value=1e-6, drop_path=dp_rate[countlayer + i]) for i in
                   range(layers[level])]
        self.blocks = nn.Sequential(*modules)

    def forward(self, *args):
        x = self.fusion(*args)
        x = self.blocks(x)
        return x


class SubNet(nn.Module):
    def __init__(self, channels, layers, kernel_size, first_col, dp_rates, save_memory) -> None:
        super().__init__()
        shortcut_scale_init_value = 0.5
        self.save_memory = save_memory
        self.alpha0 = nn.Parameter(shortcut_scale_init_value * torch.ones((1, channels[0], 1, 1)),
                                   requires_grad=True) if shortcut_scale_init_value > 0 else None
        self.alpha1 = nn.Parameter(shortcut_scale_init_value * torch.ones((1, channels[1], 1, 1)),
                                   requires_grad=True) if shortcut_scale_init_value > 0 else None
        self.alpha2 = nn.Parameter(shortcut_scale_init_value * torch.ones((1, channels[2], 1, 1)),
                                   requires_grad=True) if shortcut_scale_init_value > 0 else None
        self.alpha3 = nn.Parameter(shortcut_scale_init_value * torch.ones((1, channels[3], 1, 1)),
                                   requires_grad=True) if shortcut_scale_init_value > 0 else None

        self.level0 = Level(0, channels, layers, kernel_size, first_col, dp_rates)

        self.level1 = Level(1, channels, layers, kernel_size, first_col, dp_rates)

        self.level2 = Level(2, channels, layers, kernel_size, first_col, dp_rates)

        self.level3 = Level(3, channels, layers, kernel_size, first_col, dp_rates)

    def _forward_nonreverse(self, *args):
        x, c0, c1, c2, c3 = args

        c0 = (self.alpha0) * c0 + self.level0(x, c1)
        c1 = (self.alpha1) * c1 + self.level1(c0, c2)
        c2 = (self.alpha2) * c2 + self.level2(c1, c3)
        c3 = (self.alpha3) * c3 + self.level3(c2, None)
        return c0, c1, c2, c3

    def _forward_reverse(self, *args):

        local_funs = [self.level0, self.level1, self.level2, self.level3]
        alpha = [self.alpha0, self.alpha1, self.alpha2, self.alpha3]
        _, c0, c1, c2, c3 = ReverseFunction.apply(
            local_funs, alpha, *args)

        return c0, c1, c2, c3

    def forward(self, *args):

        self._clamp_abs(self.alpha0.data, 1e-3)
        self._clamp_abs(self.alpha1.data, 1e-3)
        self._clamp_abs(self.alpha2.data, 1e-3)
        self._clamp_abs(self.alpha3.data, 1e-3)

        if self.save_memory:
            return self._forward_reverse(*args)
        else:
            return self._forward_nonreverse(*args)

    def _clamp_abs(self, data, value):
        with torch.no_grad():
            sign = data.sign()
            data.abs_().clamp_(value)
            data *= sign


class Classifier(nn.Module):
    def __init__(self, in_channels, num_classes):
        super().__init__()

        self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
        self.classifier = nn.Sequential(
            nn.LayerNorm(in_channels, eps=1e-6),  # final norm layer
            nn.Linear(in_channels, num_classes),
        )

    def forward(self, x):
        x = self.avgpool(x)
        x = x.view(x.size(0), -1)
        x = self.classifier(x)
        return x


class FullNet(nn.Module):
    def __init__(self, channels=[32, 64, 96, 128], layers=[2, 3, 6, 3], num_subnet=5, kernel_size=3, drop_path=0.0,
                 save_memory=True, inter_supv=True) -> None:
        super().__init__()
        self.num_subnet = num_subnet
        self.inter_supv = inter_supv
        self.channels = channels
        self.layers = layers

        self.stem = nn.Sequential(
            nn.Conv2d(3, channels[0], kernel_size=4, stride=4),
            LayerNorm(channels[0], eps=1e-6, data_format="channels_first")
        )

        dp_rate = [x.item() for x in torch.linspace(0, drop_path, sum(layers))]
        for i in range(num_subnet):
            first_col = True if i == 0 else False
            self.add_module(f'subnet{str(i)}', SubNet(
                channels, layers, kernel_size, first_col, dp_rates=dp_rate, save_memory=save_memory))

        self.apply(self._init_weights)
        self.width_list = [i.size(1) for i in self.forward(torch.randn(1, 3, 640, 640))]
    def forward(self, x):

        c0, c1, c2, c3 = 0, 0, 0, 0
        x = self.stem(x)
        for i in range(self.num_subnet):
            c0, c1, c2, c3 = getattr(self, f'subnet{str(i)}')(x, c0, c1, c2, c3)
        return [c0, c1, c2, c3]

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


##-------------------------------------- Tiny -----------------------------------------

def revcol_tiny(save_memory=True, inter_supv=True, drop_path=0.1, kernel_size=3):
    channels = [64, 128, 256, 512]
    layers = [2, 2, 4, 2]
    num_subnet = 4
    return FullNet(channels, layers, num_subnet, drop_path=drop_path, save_memory=save_memory, inter_supv=inter_supv,
                   kernel_size=kernel_size)


##-------------------------------------- Small -----------------------------------------

def revcol_small(save_memory=True, inter_supv=True, drop_path=0.3, kernel_size=3):
    channels = [64, 128, 256, 512]
    layers = [2, 2, 4, 2]
    num_subnet = 8
    return FullNet(channels, layers, num_subnet, drop_path=drop_path, save_memory=save_memory, inter_supv=inter_supv,
                   kernel_size=kernel_size)


##-------------------------------------- Base -----------------------------------------

def revcol_base(save_memory=True, inter_supv=True, drop_path=0.4, kernel_size=3, head_init_scale=None):
    channels = [72, 144, 288, 576]
    layers = [1, 1, 3, 2]
    num_subnet = 16
    return FullNet(channels, layers, num_subnet, drop_path=drop_path, save_memory=save_memory, inter_supv=inter_supv, kernel_size=kernel_size)


##-------------------------------------- Large -----------------------------------------

def revcol_large(save_memory=True, inter_supv=True, drop_path=0.5, kernel_size=3, head_init_scale=None):
    channels = [128, 256, 512, 1024]
    layers = [1, 2, 6, 2]
    num_subnet = 8
    return FullNet(channels, layers, num_subnet, drop_path=drop_path, save_memory=save_memory, inter_supv=inter_supv, kernel_size=kernel_size)


##--------------------------------------Extra-Large -----------------------------------------
def revcol_xlarge(save_memory=True, inter_supv=True, drop_path=0.5, kernel_size=3, head_init_scale=None):
    channels = [224, 448, 896, 1792]
    layers = [1, 2, 6, 2]
    num_subnet = 8
    return FullNet(channels, layers, num_subnet, drop_path=drop_path, save_memory=save_memory, inter_supv=inter_supv, kernel_size=kernel_size)

# model = revcol_xlarge(True)
# # 示例输入
# input = torch.randn(64, 3, 224, 224)
# output = model(input)
#
# print(len(output))#torch.Size([3, 64, 224, 224])

四、手把手教你添加RevColV1网络结构

下面教大家如何修改该网络结构，主干网络结构的修改步骤比较复杂，我也会将task.py文件上传到CSDN的文件中，大家如果自己修改不正确，可以尝试用我的task.py文件替换你的，然后只需要修改其中的第1、2、3、5步即可。

⭐修改过程中大家一定要仔细⭐

4.1 修改一

首先我门中到如下“ultralytics/nn”的目录，我们在这个目录下在创建一个新的目录，名字为'Addmodules'（此文件之后就用于存放我们的所有改进机制），之后我们在创建的目录内创建一个新的py文件复制粘贴进去，可以根据文章改进机制来起，这里大家根据自己的习惯命名即可。

4.2 修改二

第二步我们在我们创建的目录内创建一个新的py文件名字为'__init__.py'（只需要创建一个即可），然后在其内部导入我们本文的改进机制即可，其余代码均为未发大家没有不用理会！。

4.3 修改三

第三步我门中到如下文件'ultralytics/nn/tasks.py'然后在开头导入我们的所有改进机制（如果你用了我多个改进机制，这一步只需要修改一次即可）。

4.4 修改四

添加如下两行代码！！！

4.5 修改五

找到七百多行大概把具体看图片，按照图片来修改就行，添加红框内的部分，注意没有()只是函数名（此处我的文件里已经添加很多了后期都会发出来，大家没有的不用理会即可）。

        elif m in {自行添加对应的模型即可，下面都是一样的}:
            m = m(*args)
            c2 = m.width_list  # 返回通道列表
            backbone = True

4.6 修改六

用下面的代码替换红框内的内容。

if isinstance(c2, list):
    m_ = m
    m_.backbone = True
else:
    m_ = nn.Sequential(*(m(*args) for _ in range(n))) if n > 1 else m(*args)  # module
    t = str(m)[8:-2].replace('__main__.', '')  # module type
m.np = sum(x.numel() for x in m_.parameters())  # number params
m_.i, m_.f, m_.type = i + 4 if backbone else i, f, t  # attach index, 'from' index, type
if verbose:
    LOGGER.info(f'{i:>3}{str(f):>20}{n_:>3}{m.np:10.0f}  {t:<45}{str(args):<30}')  # print
save.extend(
    x % (i + 4 if backbone else i) for x in ([f] if isinstance(f, int) else f) if x != -1)  # append to savelist
layers.append(m_)
if i == 0:
    ch = []
if isinstance(c2, list):
    ch.extend(c2)
    if len(c2) != 5:
        ch.insert(0, 0)
else:
    ch.append(c2)

4.7 修改七

修改七这里非常要注意，不是文件开头YOLOv8的那predict，是400+行的RTDETR的predict！！！初始模型如下，用我给的代码替换即可！！！

代码如下->

 def predict(self, x, profile=False, visualize=False, batch=None, augment=False, embed=None):
        """
        Perform a forward pass through the model.

        Args:
            x (torch.Tensor): The input tensor.
            profile (bool, optional): If True, profile the computation time for each layer. Defaults to False.
            visualize (bool, optional): If True, save feature maps for visualization. Defaults to False.
            batch (dict, optional): Ground truth data for evaluation. Defaults to None.
            augment (bool, optional): If True, perform data augmentation during inference. Defaults to False.
            embed (list, optional): A list of feature vectors/embeddings to return.

        Returns:
            (torch.Tensor): Model's output tensor.
        """
        y, dt, embeddings = [], [], []  # outputs
        for m in self.model[:-1]:  # except the head part
            if m.f != -1:  # if not from previous layer
                x = y[m.f] if isinstance(m.f, int) else [x if j == -1 else y[j] for j in m.f]  # from earlier layers
            if profile:
                self._profile_one_layer(m, x, dt)
            if hasattr(m, 'backbone'):
                x = m(x)
                if len(x) != 5:  # 0 - 5
                    x.insert(0, None)
                for index, i in enumerate(x):
                    if index in self.save:
                        y.append(i)
                    else:
                        y.append(None)
                x = x[-1]  # 最后一个输出传给下一层
            else:
                x = m(x)  # run
                y.append(x if m.i in self.save else None)  # save output
            if visualize:
                feature_visualization(x, m.type, m.i, save_dir=visualize)
            if embed and m.i in embed:
                embeddings.append(nn.functional.adaptive_avg_pool2d(x, (1, 1)).squeeze(-1).squeeze(-1))  # flatten
                if m.i == max(embed):
                    return torch.unbind(torch.cat(embeddings, 1), dim=0)
        head = self.model[-1]
        x = head([y[j] for j in head.f], batch)  # head inference
        return x

4.8 修改八

我们将下面的s用640替换即可，这一步也是部分的主干可以不修改，但有的不修改就会报错，所以我们还是修改一下。

4.9 RT-DETR不能打印计算量问题的解决

计算的GFLOPs计算异常不打印，所以需要额外修改一处，我们找到如下文件'ultralytics/utils/torch_utils.py'文件内有如下的代码按照如下的图片进行修改，大家看好函数就行，其中红框的640可能和你的不一样，然后用我给的代码替换掉整个代码即可。

def get_flops(model, imgsz=640):
    """Return a YOLO model's FLOPs."""
    try:
        model = de_parallel(model)
        p = next(model.parameters())
        # stride = max(int(model.stride.max()), 32) if hasattr(model, 'stride') else 32  # max stride
        stride = 640
        im = torch.empty((1, 3, stride, stride), device=p.device)  # input image in BCHW format
        flops = thop.profile(deepcopy(model), inputs=[im], verbose=False)[0] / 1E9 * 2 if thop else 0  # stride GFLOPs
        imgsz = imgsz if isinstance(imgsz, list) else [imgsz, imgsz]  # expand if int/float
        return flops * imgsz[0] / stride * imgsz[1] / stride  # 640x640 GFLOPs
    except Exception:
        return 0

4.10 可选修改

有些读者的数据集部分图片比较特殊，在验证的时候会导致形状不匹配的报错，如果大家在验证的时候报错形状不匹配的错误可以固定验证集的图片尺寸，方法如下 ->

找到下面这个文件ultralytics/models/yolo/detect/train.py然后其中有一个类是DetectionTrainer class中的build_dataset函数中的一个参数rect=mode == 'val'改为rect=False

五、RevColV1的yaml文件

5.1 yaml文件

大家复制下面的yaml文件，然后通过我给大家的运行代码运行即可，RT-DETR的调参部分需要后面的文章给大家讲，现在目前免费给大家看这一部分不开放。

# Ultralytics YOLO 🚀, AGPL-3.0 license
# RT-DETR-l object detection model with P3-P5 outputs. For details see https://docs.ultralytics.com/models/rtdetr

# Parameters
nc: 80  # number of classes
scales: # model compound scaling constants, i.e. 'model=yolov8n-cls.yaml' will call yolov8-cls.yaml with scale 'n'
  # [depth, width, max_channels]
  l: [1.00, 1.00, 1024]

backbone:
  # [from, repeats, module, args]
  - [-1, 1, revcol_small, []]  # 4

head:
  - [-1, 1, Conv, [256, 1, 1, None, 1, 1, False]]  # 5 input_proj.2
  - [-1, 1, AIFI, [1024, 8]] # 6
  - [-1, 1, Conv, [256, 1, 1]]  # 7, Y5, lateral_convs.0

  - [-1, 1, nn.Upsample, [None, 2, 'nearest']] # 8
  - [3, 1, Conv, [256, 1, 1, None, 1, 1, False]]  # 9 input_proj.1
  - [[-2, -1], 1, Concat, [1]] # 10
  - [-1, 3, RepC3, [256, 0.5]]  # 11, fpn_blocks.0
  - [-1, 1, Conv, [256, 1, 1]]   # 12, Y4, lateral_convs.1

  - [-1, 1, nn.Upsample, [None, 2, 'nearest']] # 13
  - [2, 1, Conv, [256, 1, 1, None, 1, 1, False]]  # 14 input_proj.0
  - [[-2, -1], 1, Concat, [1]]  # 15 cat backbone P4
  - [-1, 3, RepC3, [256, 0.5]]    # X3 (16), fpn_blocks.1

  - [-1, 1, Conv, [256, 3, 2]]   # 17, downsample_convs.0
  - [[-1, 12], 1, Concat, [1]]  # 18 cat Y4
  - [-1, 3, RepC3, [256, 0.5]]    # F4 (19), pan_blocks.0

  - [-1, 1, Conv, [256, 3, 2]]   # 20, downsample_convs.1
  - [[-1, 7], 1, Concat, [1]]  # 21 cat Y5
  - [-1, 3, RepC3, [256, 0.5]]    # F5 (22), pan_blocks.1

  - [[16, 19, 22], 1, RTDETRDecoder, [nc, 256, 300, 4, 8, 3]]  # Detect(P3, P4, P5)

5.2 运行文件

大家可以创建一个train.py文件将下面的代码粘贴进去然后替换你的文件运行即可开始训练。

import warnings
from ultralytics import RTDETR
warnings.filterwarnings('ignore')

if __name__ == '__main__':
    model = RTDETR('替换你想要运行的yaml文件')
    # model.load('') # 可以加载你的版本预训练权重
    model.train(data=r'替换你的数据集地址即可',
                cache=False,
                imgsz=640,
                epochs=72,
                batch=4,
                workers=0,
                device='0',
                project='runs/RT-DETR-train',
                name='exp',
                # amp=True
                )