网络模型backbone我使用了之前搭建过的MobileNetV3，具体的搭建过程可以查看：【20】MobileNetV3
然后由于此次任务需要的是输出其中的中间特征矩阵，所以前向传播的部分代码需要修改一下：
参考代码如下：
backbone.py
import torch import torch.nn as nn import torchvision num_class = 5 # 定义h-swith激活函数 class HardSwish(nn.Module): def __init__(self, inplace=True): super(HardSwish, self).__init__() self.relu6 = nn.ReLU6(inplace) def forward(self, x): return x*self.relu6(x+3)/6 # DW卷积 def ConvBNActivation(in_channels,out_channels,kernel_size,stride,activate): # 通过设置padding达到当stride=2时，hw减半的效果。此时不与kernel_size有关，所实现的公式为: padding=(kernel_size-1)//2 # 当kernel_size=3,padding=1时: stride=2 hw减半, stride=1 hw不变 # 当kernel_size=5,padding=2时: stride=2 hw减半, stride=1 hw不变 # 从而达到了使用 stride 来控制hw的效果， 不用去关心kernel_size的大小，控制单一变量 return nn.Sequential( nn.Conv2d(in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, stride=stride, padding=(kernel_size-1)//2, groups=in_channels), nn.BatchNorm2d(out_channels), nn.ReLU6(inplace=True) if activate == 'relu' else HardSwish() ) # PW卷积(接全连接层) def Conv1x1BN(in_channels,out_channels): return nn.Sequential( nn.Conv2d(in_channels=in_channels, out_channels=out_channels, kernel_size=1, stride=1), nn.BatchNorm2d(out_channels) ) # 普通的1x1卷积 def Conv1x1BNActivation(in_channels,out_channels,activate): return nn.Sequential( nn.Conv2d(in_channels=in_channels, out_channels=out_channels, kernel_size=1, stride=1), nn.BatchNorm2d(out_channels), nn.ReLU6(inplace=True) if activate == 'relu' else HardSwish() ) # 注意力机制(SE模块) class SqueezeAndExcite(nn.Module): def __init__(self, in_channels, out_channels, se_kernel_size, divide=4): super(SqueezeAndExcite, self).__init__() mid_channels = in_channels // divide # 维度变为原来的1/4 # 将当前的channel平均池化成1 self.pool = nn.AvgPool2d(kernel_size=se_kernel_size,stride=1) # 两个全连接层 最后输出每层channel的权值 self.SEblock = nn.Sequential( nn.Linear(in_features=in_channels, out_features=mid_channels), nn.ReLU6(inplace=True), nn.Linear(in_features=mid_channels, out_features=out_channels), HardSwish(inplace=True), ) def forward(self, x): b, c, h, w = x.size() out = self.pool(x) # 不管当前的 h,w 为多少, 全部池化为1 out = out.view(b, -1) # 打平处理，与全连接层相连 # 获取注意力机制后的权重 out = self.SEblock(out) # out是每层channel的权重，需要扩维才能与原特征矩阵相乘 out = out.view(b, c, 1, 1) # 增维 return out * x class SEInvertedBottleneck(nn.Module): def __init__(self, in_channels, mid_channels, out_channels, kernel_size, stride, activate, use_se, se_kernel_size=1): super(SEInvertedBottleneck, self).__init__() self.stride = stride self.use_se = use_se self.in_channels = in_channels self.out_channels = out_channels # mid_channels = (in_channels * expansion_factor) # 普通1x1卷积升维操作 self.conv = Conv1x1BNActivation(in_channels, mid_channels,activate) # DW卷积 维度不变，但可通过stride改变尺寸 groups=in_channels self.depth_conv = ConvBNActivation(mid_channels, mid_channels, kernel_size,stride,activate) # 注意力机制的使用判断 if self.use_se: self.SEblock = SqueezeAndExcite(mid_channels, mid_channels, se_kernel_size) # PW卷积 降维操作 self.point_conv = Conv1x1BNActivation(mid_channels, out_channels,activate) # shortcut的使用判断 if self.stride == 1: self.shortcut = Conv1x1BN(in_channels, out_channels) def forward(self, x): # DW卷积 out = self.depth_conv(self.conv(x)) # 当 use_se=True 时使用注意力机制 if self.use_se: out = self.SEblock(out) # PW卷积 out = self.point_conv(out) # 残差操作 # 第一种: 只看步长，步长相同shape不一样的输入输出使用1x1卷积使其相加 # out = (out + self.shortcut(x)) if self.stride == 1 else out # 第二种: 同时满足步长与输入输出的channel, 不使用1x1卷积强行升维 out = (out + x) if self.stride == 1 and self.in_channels == self.out_channels else out return out class MobileNetV3(nn.Module): def __init__(self, num_classes=num_class): super(MobileNetV3, self).__init__() self.type = type # 224x224x3 conv2d 3 -> 16 SE=False HS s=2 self.first_conv = nn.Sequential( nn.Conv2d(in_channels=3, out_channels=16, kernel_size=3, stride=2, padding=1), nn.BatchNorm2d(16), HardSwish(inplace=True), ) # torch.Size([1, 16, 112, 112]) # MobileNetV3_Large 网络结构 # torch.Size([1, 16, 112, 112]) 16 -> 16 -> 16 SE=False RE s=1 self.block1 = SEInvertedBottleneck(in_channels=16, mid_channels=16, out_channels=16, kernel_size=3, stride=1,activate='relu', use_se=False) # torch.Size([1, 16, 112, 112]) 16 -> 64 -> 24 SE=False RE s=2 self.block2 = SEInvertedBottleneck(in_channels=16, mid_channels=64, out_channels=24, kernel_size=3, stride=2, activate='relu', use_se=False) # torch.Size([1, 24, 56, 56]) 24 -> 72 -> 24 SE=False RE s=1 self.block3 = SEInvertedBottleneck(in_channels=24, mid_channels=72, out_channels=24, kernel_size=3, stride=1, activate='relu', use_se=False) # torch.Size([1, 24, 56, 56]) 24 -> 72 -> 40 SE=True RE s=2 self.block4 = SEInvertedBottleneck(in_channels=24, mid_channels=72, out_channels=40, kernel_size=5, stride=2,activate='relu', use_se=True, se_kernel_size=28) # torch.Size([1, 40, 28, 28]) 40 -> 120 -> 40 SE=True RE s=1 self.block5 = SEInvertedBottleneck(in_channels=40, mid_channels=120, out_channels=40, kernel_size=5, stride=1,activate='relu', use_se=True, se_kernel_size=28) # torch.Size([1, 40, 28, 28]) 40 -> 120 -> 40 SE=True RE s=1 self.block6 = SEInvertedBottleneck(in_channels=40, mid_channels=120, out_channels=40, kernel_size=5, stride=1,activate='relu', use_se=True, se_kernel_size=28) # torch.Size([1, 40, 28, 28]) 40 -> 240 -> 80 SE=False HS s=1 self.block7 = SEInvertedBottleneck(in_channels=40, mid_channels=240, out_channels=80, kernel_size=3, stride=1,activate='hswish', use_se=False) # torch.Size([1, 80, 28, 28]) 80 -> 200 -> 80 SE=False HS s=1 self.block8 = SEInvertedBottleneck(in_channels=80, mid_channels=200, out_channels=80, kernel_size=3, stride=1,activate='hswish', use_se=False) # torch.Size([1, 80, 28, 28]) 80 -> 184 -> 80 SE=False HS s=2 self.block9 = SEInvertedBottleneck(in_channels=80, mid_channels=184, out_channels=80, kernel_size=3, stride=2,activate='hswish', use_se=False) # torch.Size([1, 80, 14, 14]) 80 -> 184 -> 80 SE=False HS s=1 self.block10 = SEInvertedBottleneck(in_channels=80, mid_channels=184, out_channels=80, kernel_size=3, stride=1,activate='hswish', use_se=False) # torch.Size([1, 80, 14, 14]) 80 -> 480 -> 112 SE=True HS s=1 self.block11 = SEInvertedBottleneck(in_channels=80, mid_channels=480, out_channels=112, kernel_size=3, stride=1,activate='hswish', use_se=True, se_kernel_size=14) # torch.Size([1, 112, 14, 14]) 112 -> 672 -> 112 SE=True HS s=1 self.block12 = SEInvertedBottleneck(in_channels=112, mid_channels=672, out_channels=112, kernel_size=3, stride=1,activate='hswish', use_se=True, se_kernel_size=14) # torch.Size([1, 112, 14, 14]) 112 -> 672 -> 160 SE=True HS s=2 self.block13 = SEInvertedBottleneck(in_channels=112, mid_channels=672, out_channels=160, kernel_size=5, stride=2,activate='hswish', use_se=True,se_kernel_size=7) # torch.Size([1, 160, 7, 7]) 160 -> 960 -> 160 SE=True HS s=1 self.block14 = SEInvertedBottleneck(in_channels=160, mid_channels=960, out_channels=160, kernel_size=5, stride=1,activate='hswish', use_se=True,se_kernel_size=7) # torch.Size([1, 160, 7, 7]) 160 -> 960 -> 160 SE=True HS s=1 self.block15 = SEInvertedBottleneck(in_channels=160, mid_channels=960, out_channels=160, kernel_size=5, stride=1,activate='hswish', use_se=True,se_kernel_size=7) # torch.Size([1, 160, 7, 7]) # 相比MobileNetV2，尾部结构改变，,变得更加的高效 self.large_last_stage = nn.Sequential( nn.Conv2d(in_channels=160, out_channels=960, kernel_size=1, stride=1), nn.BatchNorm2d(960), HardSwish(inplace=True), nn.AvgPool2d(kernel_size=7, stride=1), nn.Conv2d(in_channels=960, out_channels=1280, kernel_size=1, stride=1), HardSwish(inplace=True), ) self.classifier = nn.Linear(in_features=1280,out_features=num_classes) self.init_params() # 初始化权重 def init_params(self): for m in self.modules(): if isinstance(m, nn.Conv2d): nn.init.kaiming_normal_(m.weight) nn.init.constant_(m.bias, 0) elif isinstance(m, nn.BatchNorm2d) or isinstance(m, nn.Linear): nn.init.constant_(m.weight, 1) nn.init.constant_(m.bias, 0) def forward(self, x): outputs = [] x = self.first_conv(x) # torch.Size([1, 16, 112, 112]) outputs.append(x) x = self.block1(x) outputs.append(x) x = self.block2(x) outputs.append(x) x = self.block3(x) outputs.append(x) x = self.block4(x) outputs.append(x) x = self.block5(x) outputs.append(x) x = self.block6(x) outputs.append(x) x = self.block7(x) outputs.append(x) x = self.block8(x) outputs.append(x) x = self.block9(x) outputs.append(x) x = self.block10(x) outputs.append(x) x = self.block11(x) outputs.append(x) x = self.block12(x) outputs.append(x) x = self.block13(x) outputs.append(x) x = self.block14(x) outputs.append(x) x = self.block15(x) # torch.Size([1, 160, 7, 7]) outputs.append(x) x = self.large_last_stage(x) # torch.Size([1, 1280, 1, 1]) # outputs.append(x) x = x.view(x.size(0), -1) # torch.Size([1, 1280]) x = self.classifier(x) # torch.Size([1, 5]) return outputs def MobileNetV3_large(): return MobileNetV3() if __name__ == '__main__': model = MobileNetV3_large() # print(model) input = torch.randn(1, 3, 224, 224) out = model(input) # print(out.shape) torch.save(model.state_dict(), 'MobileNetV3_Large.mdl')
主要思路：
主要思路就是经过卷积之后的特征矩阵保存在一个列表中，然后依次对其进行读取并利用plt进行显示与保存图像，而且具体是查看特征矩阵的每层channel的图像：plt.imshow(im[:, :, i], cmap=‘gray’)
analyze_feature_map.py
import torch from backbone import MobileNetV3_large import matplotlib.pyplot as plt import numpy as np import os from PIL import Image from torchvision import transforms # 图像的格式转换 data_transform = transforms.Compose( [transforms.Resize(256), transforms.CenterCrop(224), transforms.ToTensor(), transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])]) # create model model = MobileNetV3_large() # load model weights model_weight_path = 'MobileNetV3_Large.mdl' model.load_state_dict(torch.load(model_weight_path)) # print(model) # load image # img = Image.open('./Brightprint_above_1.jpg') img = Image.open('./Brightprint_side_1.jpg') # [N, C, H, W] img = data_transform(img) # expand batch dimension img = torch.unsqueeze(img, dim=0) # print(img.shape) # torch.Size([1, 3, 224, 224]) # forward out_put = model(img) print(out_put) # for feature_map in out_put: for batchidx, feature_map in enumerate(out_put): # 保存中间层的参数 with open('parameter.txt', mode='a') as fw: fw.writelines('block{}:parameter'.format(batchidx)) fw.writelines('\n\t') fw.writelines(str(feature_map)) fw.writelines('\n\t\n\t') print('batchidx = ', batchidx) # [N, C, H, W] -> [C, H, W] im = np.squeeze(feature_map.detach().numpy()) # [C, H, W] -> [H, W, C] # 需要改变才可以正常显示图像 im = np.transpose(im, [1, 2, 0]) # show top 16 feature maps plt.figure() for i in range(16): ax = plt.subplot(4, 4, i+1) # [H, W, C] cmap='gray' :设置为灰度图， [:, :, i]选择对channels进行切分 plt.imshow(im[:, :, i], cmap='gray') # 保存图像的方法 plt.savefig('block{}_outputs.jpg'.format(batchidx)) # plt.imsave(batchidx, arr, format='jpg') plt.show()
参考链接：
https://www.bilibili.com/video/BV1z7411f7za
Python中读取,显示,保存图片的方法
![论文推荐|[PR 2019]SegLink++：基于实例感知与组件组合的任意形状密集场景文本检测方法](https://ucc.alicdn.com/pic/developer-ecology/097a66ba52f04beca527bdfba5b61e8e.png?x-oss-process=image/resize,h_160,m_lfit)
