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