involution：大家一起内卷起来吧
热心网友
80
转载
本文介绍用Layer方式搭建involution算子，以此魔改ResNet打造RedNet模型，已加入【Paddle-Image-Models】项目，含转换后的最新预训练参数，精度基本对齐。还展示了算子和模型的搭建代码、测试情况及精度验证结果，RedNet性能和效率优于ResNet等模型。
引入
真·内卷无处不在，现在神经网络也能内卷了这次项目就用 Layer 的方式搭建一下 involution 算子，并且使用这个算子参照论文所述魔改一下 ResNet 打造一个新模型 RedNet当然这个模型也已经添加到了 【Paddle-Image-Models】 项目中了，包含转换之后的最新预训练参数，精度基本对齐好让大家能够尽快在神经网络里面内卷起来相关资料
论文：【Involution: Inverting the Inherence of Convolution for Visual Recognition】
代码：【d-li14/involution】
论文概要
提出了一种新的神经网络算子（operator或op）称为 involution，它比 convolution 更轻量更高效，形式上比 self-attention 更加简洁，可以用在各种视觉任务的模型上取得精度和效率的双重提升。通过 involution 的结构设计，我们能够以统一的视角来理解经典的卷积操作和近来流行的自注意力操作。算子和模型搭建
导入必要的包
In [1]import paddleimport paddle.nn as nnfrom paddle.vision.models import resnet登录后复制
involution（内卷）
针对输入 feature map 的一个坐标点上的特征向量：先通过 (FC-BN-ReLU-FC) 和 reshape (channel-to-space) 变换展开成 kernel 的形状从而得到这个坐标点上对应的 involution kernel再和输入 feature map 上这个坐标点邻域的特征向量进行 Multiply-Add 得到最终输出的 feature mapinvolution 示意图如下：
In [2]class involution(nn.Layer): def __init__(self, channels, kernel_size, stride): super(involution, self).__init__() self.kernel_size = kernel_size self.stride = stride self.channels = channels reduction_ratio = 4 self.group_channels = 16 self.groups = self.channels // self.group_channels self.conv1 = nn.Sequential( ('conv', nn.Conv2D( in_channels=channels, out_channels=channels // reduction_ratio, kernel_size=1, bias_attr=False )), ('bn', nn.BatchNorm2D(channels // reduction_ratio)), ('activate', nn.ReLU()) ) self.conv2 = nn.Sequential( ('conv', nn.Conv2D( in_channels=channels // reduction_ratio, out_channels=kernel_size**2 * self.groups, kernel_size=1, stride=1)) ) if stride > 1: self.avgpool = nn.AvgPool2D(stride, stride) def forward(self, x): weight = self.conv2(self.conv1( x if self.stride == 1 else self.avgpool(x))) b, c, h, w = weight.shape weight = weight.reshape(( b, self.groups, self.kernel_size**2, h, w)).unsqueeze(2) out = nn.functional.unfold( x, self.kernel_size, strides=self.stride, paddings=(self.kernel_size-1)//2, dilations=1) out = out.reshape( (b, self.groups, self.group_channels, self.kernel_size**2, h, w)) out = (weight * out).sum(axis=3).reshape((b, self.channels, h, w)) return out登录后复制 算子测试
In [3]inv = involution(128, 7, 1)paddle.summary(inv, (1, 128, 64, 64))out = inv(paddle.randn((1, 128, 64, 64)))print(out.shape)登录后复制
--------------------------------------------------------------------------- Layer (type) Input Shape Output Shape Param # =========================================================================== Conv2D-1 [[1, 128, 64, 64]] [1, 32, 64, 64] 4,096 BatchNorm2D-1 [[1, 32, 64, 64]] [1, 32, 64, 64] 128 ReLU-1 [[1, 32, 64, 64]] [1, 32, 64, 64] 0 Conv2D-2 [[1, 32, 64, 64]] [1, 392, 64, 64] 12,936 ===========================================================================Total params: 17,160Trainable params: 17,032Non-trainable params: 128---------------------------------------------------------------------------Input size (MB): 2.00Forward/backward pass size (MB): 15.25Params size (MB): 0.07Estimated Total Size (MB): 17.32---------------------------------------------------------------------------[1, 128, 64, 64]登录后复制
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/nn/layer/norm.py:648: UserWarning: When training, we now always track global mean and variance. "When training, we now always track global mean and variance.")登录后复制
RedNet
使用 involution 替换 ResNet BottleneckBlock 中的 3x3 convolution 得到了一族新的骨干网络 RedNet性能和效率优于 ResNet 和其他 self-attention 做 op 的 SOTA 模型模型具体信息如下：In [4]class BottleneckBlock(resnet.BottleneckBlock): def __init__(self, inplanes, planes, stride=1, downsample=None, groups=1, base_width=64, dilation=1, norm_layer=None): super(BottleneckBlock, self).__init__(inplanes, planes, stride, downsample, groups, base_width, dilation, norm_layer) width = int(planes * (base_width / 64.)) * groups self.conv2 = involution(width, 7, stride) class RedNet(resnet.ResNet): def __init__(self, block, depth, num_classes=1000, with_pool=True): super(RedNet, self).__init__(block=block, depth=50, num_classes=num_classes, with_pool=with_pool) layer_cfg = { 26: [1, 2, 4, 1], 38: [2, 3, 5, 2], 50: [3, 4, 6, 3], 101: [3, 4, 23, 3], 152: [3, 8, 36, 3] } layers = layer_cfg[depth] self.conv1 = None self.bn1 = None self.relu = None self.inplanes = 64 self.stem = nn.Sequential( nn.Sequential( ('conv', nn.Conv2D( in_channels=3, out_channels=self.inplanes // 2, kernel_size=3, stride=2, padding=1, bias_attr=False )), ('bn', nn.BatchNorm2D(self.inplanes // 2)), ('activate', nn.ReLU()) ), involution(self.inplanes // 2, 3, 1), nn.BatchNorm2D(self.inplanes // 2), nn.ReLU(), nn.Sequential( ('conv', nn.Conv2D( in_channels=self.inplanes // 2, out_channels=self.inplanes, kernel_size=3, stride=1, padding=1, bias_attr=False )), ('bn', nn.BatchNorm2D(self.inplanes)), ('activate', nn.ReLU()) ) ) self.layer1 = self._make_layer(block, 64, layers[0]) self.layer2 = self._make_layer(block, 128, layers[1], stride=2) self.layer3 = self._make_layer(block, 256, layers[2], stride=2) self.layer4 = self._make_layer(block, 512, layers[3], stride=2) def forward(self, x): x = self.stem(x) x = self.maxpool(x) x = self.layer1(x) x = self.layer2(x) x = self.layer3(x) x = self.layer4(x) if self.with_pool: x = self.avgpool(x) if self.num_classes > 0: x = paddle.flatten(x, 1) x = self.fc(x) return x登录后复制 模型测试
In [5]model = RedNet(BottleneckBlock, 26)paddle.summary(model, (1, 3, 224, 224))out = model(paddle.randn((1, 3, 224, 224)))print(out.shape)登录后复制
------------------------------------------------------------------------------- Layer (type) Input Shape Output Shape Param # =============================================================================== Conv2D-88 [[1, 3, 224, 224]] [1, 32, 112, 112] 864 BatchNorm2D-71 [[1, 32, 112, 112]] [1, 32, 112, 112] 128 ReLU-35 [[1, 32, 112, 112]] [1, 32, 112, 112] 0 Conv2D-89 [[1, 32, 112, 112]] [1, 8, 112, 112] 256 BatchNorm2D-72 [[1, 8, 112, 112]] [1, 8, 112, 112] 32 ReLU-36 [[1, 8, 112, 112]] [1, 8, 112, 112] 0 Conv2D-90 [[1, 8, 112, 112]] [1, 18, 112, 112] 162 involution-18 [[1, 32, 112, 112]] [1, 32, 112, 112] 0 BatchNorm2D-73 [[1, 32, 112, 112]] [1, 32, 112, 112] 128 ReLU-37 [[1, 32, 112, 112]] [1, 32, 112, 112] 0 Conv2D-91 [[1, 32, 112, 112]] [1, 64, 112, 112] 18,432 BatchNorm2D-74 [[1, 64, 112, 112]] [1, 64, 112, 112] 256 ReLU-38 [[1, 64, 112, 112]] [1, 64, 112, 112] 0 MaxPool2D-1 [[1, 64, 112, 112]] [1, 64, 56, 56] 0 Conv2D-93 [[1, 64, 56, 56]] [1, 64, 56, 56] 4,096 BatchNorm2D-76 [[1, 64, 56, 56]] [1, 64, 56, 56] 256 ReLU-39 [[1, 256, 56, 56]] [1, 256, 56, 56] 0 Conv2D-96 [[1, 64, 56, 56]] [1, 16, 56, 56] 1,024 BatchNorm2D-79 [[1, 16, 56, 56]] [1, 16, 56, 56] 64 ReLU-40 [[1, 16, 56, 56]] [1, 16, 56, 56] 0 Conv2D-97 [[1, 16, 56, 56]] [1, 196, 56, 56] 3,332 involution-19 [[1, 64, 56, 56]] [1, 64, 56, 56] 0 BatchNorm2D-77 [[1, 64, 56, 56]] [1, 64, 56, 56] 256 Conv2D-95 [[1, 64, 56, 56]] [1, 256, 56, 56] 16,384 BatchNorm2D-78 [[1, 256, 56, 56]] [1, 256, 56, 56] 1,024 Conv2D-92 [[1, 64, 56, 56]] [1, 256, 56, 56] 16,384 BatchNorm2D-75 [[1, 256, 56, 56]] [1, 256, 56, 56] 1,024 BottleneckBlock-17 [[1, 64, 56, 56]] [1, 256, 56, 56] 0 Conv2D-99 [[1, 256, 56, 56]] [1, 128, 56, 56] 32,768 BatchNorm2D-81 [[1, 128, 56, 56]] [1, 128, 56, 56] 512 ReLU-41 [[1, 512, 28, 28]] [1, 512, 28, 28] 0 AvgPool2D-4 [[1, 128, 56, 56]] [1, 128, 28, 28] 0 Conv2D-102 [[1, 128, 28, 28]] [1, 32, 28, 28] 4,096 BatchNorm2D-84 [[1, 32, 28, 28]] [1, 32, 28, 28] 128 ReLU-42 [[1, 32, 28, 28]] [1, 32, 28, 28] 0 Conv2D-103 [[1, 32, 28, 28]] [1, 392, 28, 28] 12,936 involution-20 [[1, 128, 56, 56]] [1, 128, 28, 28] 0 BatchNorm2D-82 [[1, 128, 28, 28]] [1, 128, 28, 28] 512 Conv2D-101 [[1, 128, 28, 28]] [1, 512, 28, 28] 65,536 BatchNorm2D-83 [[1, 512, 28, 28]] [1, 512, 28, 28] 2,048 Conv2D-98 [[1, 256, 56, 56]] [1, 512, 28, 28] 131,072 BatchNorm2D-80 [[1, 512, 28, 28]] [1, 512, 28, 28] 2,048 BottleneckBlock-18 [[1, 256, 56, 56]] [1, 512, 28, 28] 0 Conv2D-104 [[1, 512, 28, 28]] [1, 128, 28, 28] 65,536 BatchNorm2D-85 [[1, 128, 28, 28]] [1, 128, 28, 28] 512 ReLU-43 [[1, 512, 28, 28]] [1, 512, 28, 28] 0 Conv2D-107 [[1, 128, 28, 28]] [1, 32, 28, 28] 4,096 BatchNorm2D-88 [[1, 32, 28, 28]] [1, 32, 28, 28] 128 ReLU-44 [[1, 32, 28, 28]] [1, 32, 28, 28] 0 Conv2D-108 [[1, 32, 28, 28]] [1, 392, 28, 28] 12,936 involution-21 [[1, 128, 28, 28]] [1, 128, 28, 28] 0 BatchNorm2D-86 [[1, 128, 28, 28]] [1, 128, 28, 28] 512 Conv2D-106 [[1, 128, 28, 28]] [1, 512, 28, 28] 65,536 BatchNorm2D-87 [[1, 512, 28, 28]] [1, 512, 28, 28] 2,048 BottleneckBlock-19 [[1, 512, 28, 28]] [1, 512, 28, 28] 0 Conv2D-110 [[1, 512, 28, 28]] [1, 256, 28, 28] 131,072 BatchNorm2D-90 [[1, 256, 28, 28]] [1, 256, 28, 28] 1,024 ReLU-45 [[1, 1024, 14, 14]] [1, 1024, 14, 14] 0 AvgPool2D-5 [[1, 256, 28, 28]] [1, 256, 14, 14] 0 Conv2D-113 [[1, 256, 14, 14]] [1, 64, 14, 14] 16,384 BatchNorm2D-93 [[1, 64, 14, 14]] [1, 64, 14, 14] 256 ReLU-46 [[1, 64, 14, 14]] [1, 64, 14, 14] 0 Conv2D-114 [[1, 64, 14, 14]] [1, 784, 14, 14] 50,960 involution-22 [[1, 256, 28, 28]] [1, 256, 14, 14] 0 BatchNorm2D-91 [[1, 256, 14, 14]] [1, 256, 14, 14] 1,024 Conv2D-112 [[1, 256, 14, 14]] [1, 1024, 14, 14] 262,144 BatchNorm2D-92 [[1, 1024, 14, 14]] [1, 1024, 14, 14] 4,096 Conv2D-109 [[1, 512, 28, 28]] [1, 1024, 14, 14] 524,288 BatchNorm2D-89 [[1, 1024, 14, 14]] [1, 1024, 14, 14] 4,096 BottleneckBlock-20 [[1, 512, 28, 28]] [1, 1024, 14, 14] 0 Conv2D-115 [[1, 1024, 14, 14]] [1, 256, 14, 14] 262,144 BatchNorm2D-94 [[1, 256, 14, 14]] [1, 256, 14, 14] 1,024 ReLU-47 [[1, 1024, 14, 14]] [1, 1024, 14, 14] 0 Conv2D-118 [[1, 256, 14, 14]] [1, 64, 14, 14] 16,384 BatchNorm2D-97 [[1, 64, 14, 14]] [1, 64, 14, 14] 256 ReLU-48 [[1, 64, 14, 14]] [1, 64, 14, 14] 0 Conv2D-119 [[1, 64, 14, 14]] [1, 784, 14, 14] 50,960 involution-23 [[1, 256, 14, 14]] [1, 256, 14, 14] 0 BatchNorm2D-95 [[1, 256, 14, 14]] [1, 256, 14, 14] 1,024 Conv2D-117 [[1, 256, 14, 14]] [1, 1024, 14, 14] 262,144 BatchNorm2D-96 [[1, 1024, 14, 14]] [1, 1024, 14, 14] 4,096 BottleneckBlock-21 [[1, 1024, 14, 14]] [1, 1024, 14, 14] 0 Conv2D-120 [[1, 1024, 14, 14]] [1, 256, 14, 14] 262,144 BatchNorm2D-98 [[1, 256, 14, 14]] [1, 256, 14, 14] 1,024 ReLU-49 [[1, 1024, 14, 14]] [1, 1024, 14, 14] 0 Conv2D-123 [[1, 256, 14, 14]] [1, 64, 14, 14] 16,384 BatchNorm2D-101 [[1, 64, 14, 14]] [1, 64, 14, 14] 256 ReLU-50 [[1, 64, 14, 14]] [1, 64, 14, 14] 0 Conv2D-124 [[1, 64, 14, 14]] [1, 784, 14, 14] 50,960 involution-24 [[1, 256, 14, 14]] [1, 256, 14, 14] 0 BatchNorm2D-99 [[1, 256, 14, 14]] [1, 256, 14, 14] 1,024 Conv2D-122 [[1, 256, 14, 14]] [1, 1024, 14, 14] 262,144 BatchNorm2D-100 [[1, 1024, 14, 14]] [1, 1024, 14, 14] 4,096 BottleneckBlock-22 [[1, 1024, 14, 14]] [1, 1024, 14, 14] 0 Conv2D-125 [[1, 1024, 14, 14]] [1, 256, 14, 14] 262,144 BatchNorm2D-102 [[1, 256, 14, 14]] [1, 256, 14, 14] 1,024 ReLU-51 [[1, 1024, 14, 14]] [1, 1024, 14, 14] 0 Conv2D-128 [[1, 256, 14, 14]] [1, 64, 14, 14] 16,384 BatchNorm2D-105 [[1, 64, 14, 14]] [1, 64, 14, 14] 256 ReLU-52 [[1, 64, 14, 14]] [1, 64, 14, 14] 0 Conv2D-129 [[1, 64, 14, 14]] [1, 784, 14, 14] 50,960 involution-25 [[1, 256, 14, 14]] [1, 256, 14, 14] 0 BatchNorm2D-103 [[1, 256, 14, 14]] [1, 256, 14, 14] 1,024 Conv2D-127 [[1, 256, 14, 14]] [1, 1024, 14, 14] 262,144 BatchNorm2D-104 [[1, 1024, 14, 14]] [1, 1024, 14, 14] 4,096 BottleneckBlock-23 [[1, 1024, 14, 14]] [1, 1024, 14, 14] 0 Conv2D-131 [[1, 1024, 14, 14]] [1, 512, 14, 14] 524,288 BatchNorm2D-107 [[1, 512, 14, 14]] [1, 512, 14, 14] 2,048 ReLU-53 [[1, 2048, 7, 7]] [1, 2048, 7, 7] 0 AvgPool2D-6 [[1, 512, 14, 14]] [1, 512, 7, 7] 0 Conv2D-134 [[1, 512, 7, 7]] [1, 128, 7, 7] 65,536 BatchNorm2D-110 [[1, 128, 7, 7]] [1, 128, 7, 7] 512 ReLU-54 [[1, 128, 7, 7]] [1, 128, 7, 7] 0 Conv2D-135 [[1, 128, 7, 7]] [1, 1568, 7, 7] 202,272 involution-26 [[1, 512, 14, 14]] [1, 512, 7, 7] 0 BatchNorm2D-108 [[1, 512, 7, 7]] [1, 512, 7, 7] 2,048 Conv2D-133 [[1, 512, 7, 7]] [1, 2048, 7, 7] 1,048,576 BatchNorm2D-109 [[1, 2048, 7, 7]] [1, 2048, 7, 7] 8,192 Conv2D-130 [[1, 1024, 14, 14]] [1, 2048, 7, 7] 2,097,152 BatchNorm2D-106 [[1, 2048, 7, 7]] [1, 2048, 7, 7] 8,192 BottleneckBlock-24 [[1, 1024, 14, 14]] [1, 2048, 7, 7] 0 AdaptiveAvgPool2D-1 [[1, 2048, 7, 7]] [1, 2048, 1, 1] 0 Linear-1 [[1, 2048]] [1, 1000] 2,049,000 ===============================================================================Total params: 9,264,318Trainable params: 9,202,014Non-trainable params: 62,304-------------------------------------------------------------------------------Input size (MB): 0.57Forward/backward pass size (MB): 188.62Params size (MB): 35.34Estimated Total Size (MB): 224.53-------------------------------------------------------------------------------[1, 1000]登录后复制
模型精度验证
使用 Paddle-Image-Models 来进行模型精度验证安装 PPIM
In [ ]!pip install ppim==1.0.1 -i https://pypi.python.org/pypi登录后复制
解压数据集
解压 ILSVRC2012 验证集In [ ]# 解压数据集!mkdir ~/data/ILSVRC2012!tar -xf ~/data/data68594/ILSVRC2012_img_val.tar -C ~/data/ILSVRC2012登录后复制
模型评估
使用 ILSVRC2012 验证集进行精度验证In [ ]import osimport cv2import numpy as npimport paddleimport paddle.vision.transforms as Tfrom ppim import rednet26, rednet38, rednet50, rednet101, rednet152# 构建数据集# backend cv2class ILSVRC2012(paddle.io.Dataset): def __init__(self, root, label_list, transform): self.transform = transform self.root = root self.label_list = label_list self.load_datas() def load_datas(self): self.imgs = [] self.labels = [] with open(self.label_list, 'r') as f: for line in f: img, label = line[:-1].split(' ') self.imgs.append(os.path.join(self.root, img)) self.labels.append(int(label)) def __getitem__(self, idx): label = self.labels[idx] image = self.imgs[idx] image = cv2.imread(image) image = self.transform(image) return image.astype('float32'), np.array(label).astype('int64') def __len__(self): return len(self.imgs)# 配置模型model, val_transforms = rednet26(pretrained=True)model = paddle.Model(model)model.prepare(metrics=paddle.metric.Accuracy(topk=(1, 5)))# 配置数据集val_dataset = ILSVRC2012('data/ILSVRC2012', transform=val_transforms, label_list='data/data68594/val_list.txt')# 模型验证model.evaluate(val_dataset, batch_size=16)登录后复制 {'acc_top1': 0.75956, 'acc_top5': 0.9319}登录后复制
免责声明：
游乐网为非赢利性网站，所展示的游戏/软件/文章内容均来自于互联网或第三方用户上传分享，版权归原作者所有，本站不承担相应法律责任。如您发现有涉嫌抄袭侵权的内容，请联系youleyoucom@outlook.com。
相关攻略
Trae在Python数据分析与机器学习项目中主要通过四种方式提供支持：利用Auto模式自动生成并执行端到端分析脚本；通过AgentCLI命令行自动化机器学习建模流程；对现有代码进行智能调试与优化；借助语音交互快速构建数据处理函数。这些功能覆盖了从需求描述到代码生成、模型构建及代码优化的全流程。
在Python编程中，你是否也曾编写过类似的统计代码？ 统计词频 count = {} for word in words: if word in count: count[word] += 1 else: count[word] = 1 实际上，这种高频的计数需求，完全可以通过Python内置
Trae稳定支持Python3 10至3 13版本，3 9及以下版本无法运行。Python3 14处于实验性支持阶段，核心功能可能受限。当存在多个3 10以上版本时，Trae优先选择虚拟环境中的解释器，其次为最高系统版本。此外，Trae仅兼容64位Python解释器，不支持32位架构。
在企业级数据采集与自动化运维实践中，IT团队普遍面临一个核心挑战：Python爬虫为何频繁报错，修补维护何时才能终结？随着前端技术演进与动态反爬机制的日益复杂，依赖DOM解析的传统爬虫脚本往往陷入“部署即过时，运行即异常”的困境。本文将深入解析传统爬虫代码脆弱性的根本原因，并系统介绍一种能够重塑数据
很多刚接触Docker的开发者常有一个误解：制作镜像不就是把源代码打包进去就行了吗？实际上，在企业级的标准化开发流程中，直接将源码打包进Docker镜像是非常不专业的做法。这会导致镜像体积臃肿、引入潜在安全风险，并且模糊了“构建环境”与“运行环境”的边界。本文将深入解析Java、Vue、Go、Pyt
热门专题
刀塔传奇破解版无限钻石下载大全
2025-08-05
洛克王国正式正版手游下载安装大全
2025-08-05
热门推荐
在《燕云十六声》中领悟“菩提苦海”，需沉浸探索游戏世界。主线剧情构建认知框架，战斗观察、场景细节与NPC对话皆暗藏线索。通过多元视角拼凑因果，方能深入理解游戏蕴含的宏大叙事与深邃魅力。
2026年618大促的序幕刚刚拉开，初期战报已经透露出一些耐人寻味的信号。截至5月21日，海信电视在京东平板电视累计销售竞速榜上拔得头筹，其RGB-Mini LED爆款王——海信小墨E5S Pro，更是同时拿下了天猫平板电视和抖音大家电的5 20单品销冠。 这并非偶然。奥维云网的全渠道监测数据给出了
充电桩领域的“军备竞赛”再次迎来重磅升级。5月22日，极氪汽车正式发布了其全新一代液冷超级充电桩，将单枪峰值功率一举提升至行业领先的800kW，标志着超充技术迈入新阶段。 根据官方披露的核心信息，这款超充桩主要具备四大优势：极速补能、高效节能、广泛适配与多重安全。具体而言，其单枪峰值电流高达800A
获取电弧机剑主要有五种途径：推进主线任务以解锁线索；探索遗迹、工厂等特定区域；挑战特定副本与Boss；完成提及传说武器或遗物的支线任务；参与限时活动并达成要求。玩家可根据偏好选择或组合多种方式获取该武器。
小米汽车再次为潜在车主带来惊喜福利！即日起至5月31日，用户只需提前完成预约，并到店参与任意车型的试驾体验，即可免费获赠一款1:64精致合金车模。车模款式与颜色随机发放，为试驾过程增添一份专属的收藏乐趣，诚意十足。 参与本次活动需注意以下细则：试驾必须通过官方渠道提前预约；各授权门店的车模备货数量不