目标检测目标检测目标检测目标检测
资源内容介绍
目标检测目标检测目标检测目标检测 import torchfrom torch import nnfrom torchsummary import summaryimport matplotlib.pyplot as pltimport config############################# road cls header############################# -------------- 路面分类头部网络 -----------------#class ClassifyHead(nn.Module): """" 功能说明：路面分类 header """ def __init__(self,num_classes=4): super().__init__() # Batch norm parameters momentum = 0.01 epsilon = 1e-3 in_channels = 448 out_channels = 1280 self.head = nn.Sequential( nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=1, padding=0, bias=False), nn.BatchNorm2d(num_features=out_channels, momentum=momentum, eps=epsilon), nn.ReLU6(inplace=True), ) self.avgpool = torch.nn.AdaptiveAvgPool2d((1, 1)) dropout_rate = 0.3 if dropout_rate > 0: self.dropout = nn.Dropout(dropout_rate) else: self.dropout = None self.fc = torch.nn.Linear(out_channels, num_classes) # [1 448 7 7] def forward(self, x): # [1,1280,7,7] x = self.head(x) # [1,1280,1,1] x = self.avgpool(x) # [1,1280] x = x.view(x.size(0), -1) if self.dropout is not None: x = self.dropout(x) # [1,4] x = self.fc(x) return xif __name__ == '__main__': num_class = config.class_num model = ClassifyHead(num_class) model.to('cuda') summary(model, input_size=(448,7,7)) # input_size=(channels, H, W)############################# loss############################# 功能:分类损失计算loss_function = nn.CrossEntropyLoss()# 使用方法说明# 导入：from use_in_yolox import loss_function# 调用：loss = loss_function(outputs, labels)# 输入格式说明：outputs:（1，class_num） labels:（class_id）# 功能:输出训练过程中的路面分类指标import numpyfrom sklearn.metrics import classification_reportclass ConfusionMatrix(object): """ 混淆矩阵 """ # 初始化 def __init__(self, class_num: int): self.matrix = numpy.zeros((class_num, class_num)) self.class_num = class_num self.pred=[] self.true=[] # 统计更新 def update(self, pred, label): # p代表Predicted label、t代表True label for p, t in zip(pred, label): self.matrix[p, t] += 1 self.pred.append(p) self.true.append(t) # 输出准确度 def acc(self): acc = 0 for i in range(self.class_num): acc += self.matrix[i, i] acc = acc / numpy.sum(self.matrix) return acc # 打印结果 def report(self): report = classification_report(self.true, self.pred, digits=4) return report############################# predict############################# 输入:1*class_num的tensor# 输出:预测类别def road_classify_decode(road_cls_output): predict = torch.softmax(torch.squeeze(road_cls_output).cpu(), dim=0) predicted_road_classify_result = torch.argmax(predict).numpy() print('predicted class id: ', predicted_road_classify_result) return predicted_road_classify_result############################# 验证指标 test.py############################# 输入:model,images, labels, confusion_matrix,device# 输出:路面验证结果def road_cls_validation(model,images, labels, confusion_matrix,device): images, labels = images.to(device), labels.to(device) road_cls_input = model(images) confusion_matrix.update(road_cls_input.argmax(1).cpu().numpy(), labels.cpu().numpy()) return confusion_matrix.report()# 调用示例:# from use_in_yolox import ConfusionMatrix, road_cls_validation# confusion_matrix = ConfusionMatrix(config.class_num)# with torch.no_grad(): # confusion_matrix.__init__(config.class_num) # for images, labels in test_loader: # road_cls_valid_result = road_cls_validation(net,images, labels,confusion_matrix, device)# print(road_cls_valid_result)############################# train.py############################# 功能: 绘制训练过程的路面分类LOSS曲线图# 调用示例:# confusion_matrix = ConfusionMatrix(class_num)# for i in range(1, 总训练轮数): # 训练 # model.train() # confusion_matrix.__init__(config.class_num) # for images, labels in train_loader: # 根据输出更新混淆矩阵 # outputs = model(images) #confusion_matrix.update(outputs.argmax(1).cpu().numpy(), labels.cpu().numpy()) # 打印: print("train accurancy", confusion_matrix.acc()) # 验证 # model.eval() # confusion_matrix.__init__(config.class_num) # for images, labels in val_loader: # 根据输出更新混淆矩阵 # outputs = model(images) #confusion_matrix.update(outputs.argmax(1).cpu().numpy(), labels.cpu().numpy()) # 打印: print("val accurancy", confusion_matrix.acc())# 功能: 绘制训练过程的路面分类LOSS曲线图def draw_cls_loss(train_losses,val_losses,epochs): """" 功能说明：记录每一轮训练过程中的分类loss """ plt.figure() plt.title('Road Sufrace Classify LOSS during training') # 标题 plt.plot(epochs, train_losses, label="train loss") plt.plot(epochs, val_losses, label="val loss") plt.xlabel('epochs') plt.ylabel('loss') plt.legend() plt.grid() # TODO save_path = "./" ### 这里需要修改一下保存地址 ### plt.savefig(f'{save_path}/Road Surface Classify LOSS.png')# 调用示例:# from use_in_yolox draw_cls_loss # 开始训练... # epochs = [] # t_loss = [] # for i in range(1, 总训练轮数): # 记录epoch # epochs.append(i + 1) # 记录train_loss # t_loss.append(train_loss) # 记录val_loss # val_loss.append(valid_loss) # 调用该函数绘制路面分类曲线图 # draw_cls_loss(train_losses=t_loss,val_losses=val_loss,epochs=epochs)################################ EfficientNet-Lite Backbone###############################import torch.functional as Fimport mathdef round_filters(filters, multiplier, divisor=8, min_width=None): """Calculate and round number of filters based on width multiplier.""" if not multiplier: return filters filters *= multiplier min_width = min_width or divisor new_filters = max(min_width, int(filters + divisor / 2) // divisor * divisor) # Make sure that round down does not go down by more than 10%. if new_filters < 0.9 * filters: new_filters += divisor return int(new_filters)def round_repeats(repeats, multiplier): """Round number of filters based on depth multiplier.""" if not multiplier: return repeats return int(math.ceil(multiplier * repeats))def drop_connect(x, drop_connect_rate, training): if not training: return x keep_prob = 1.0 - drop_connect_rate batch_size = x.shape[0] random_tensor = keep_prob random_tensor += torch.rand([batch_size, 1, 1, 1], dtype=x.dtype, device=x.device) binary_mask = torch.floor(random_tensor) x = (x / keep_prob) * binary_mask return xclass MBConvBlock(nn.Module): def __init__(self, inp, final_oup, k, s, expand_ratio, se_ratio, has_se=False): super(MBConvBlock, self).__init__() self._momentum = 0.01 self._epsilon = 1e-3 self.input_filters = inp self.output_filters = final_oup self.stride = s self.expand_ratio = expand_ratio