From 02a36d746ee7c3319b3ba3ec651e4aa61a1ff726 Mon Sep 17 00:00:00 2001 From: aasdar <3415807009@qq.com> Date: Thu, 18 Dec 2025 16:35:16 +0800 Subject: [PATCH 1/5] =?UTF-8?q?=E8=BD=A6=E8=BE=86=E6=A3=80=E6=B5=8B?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 1.添加新的README.md文件 2.提供了预训练权重,UA-DATRAC数据集和训练完成模型文件的下载地址 --- src/Yolov4_Vehicle_inspection/README.md | 62 +++++++++++++++++++++++++ 1 file changed, 62 insertions(+) create mode 100644 src/Yolov4_Vehicle_inspection/README.md diff --git a/src/Yolov4_Vehicle_inspection/README.md b/src/Yolov4_Vehicle_inspection/README.md new file mode 100644 index 0000000000..d1a82789a8 --- /dev/null +++ b/src/Yolov4_Vehicle_inspection/README.md @@ -0,0 +1,62 @@ +## Yolov4: 基于Pytorch的Yolov4算法实现车辆检测以及跟踪(深度学习) + +--- + +本项目是基于**[yolov4-pytorch](https://github.com/bubbliiiing/yolov4-pytorch)**实现的 + +## 文件下载 + +### 预训练权重下载 +链接:https://pan.baidu.com/s/1Wt7_yn6m8B9Uud80ScLkPQ 提取码:6666 +其中yolov4_voc_weights.pth是voc数据集的预训练权重,yolov4_coco_weights.pth是coco数据集的预训练权重 + +### UA-DATRAC数据集下载 + +链接:https://pan.baidu.com/s/14G_LLbilhe7adrLDVq1k-w 提取码:6666 + +### 训练完成模型文件下载 +链接:https://pan.baidu.com/s/1Ipw7KXqdUIjmUTQ6i0clhw 提取码:6666 +该文件为以voc数据集预训练权重使用UA-DATRAC数据集训练了23个epoch后的模型文件 + +## 数据处理 + +其中`transfer.py`和`move_pic.py`是针对UA-DATRAC数据集进行处理的两个文件。 + +## 训练步骤 + +1. 将VOC格式文件放至data的文件夹 +2. JPEGImages文件夹中存放数据图片 +3. Annotations文件夹中存放与图片一一对应的标签文件 +4. 调整voc2yolo4.py文件中的`trainval_percent` 、`train_percent`的值,控制训练集、测试集、验证集的比例;调整完成后运行该文件,在`data/ImageSets/Main`文件夹中生成对应数据集的图片文件名 +5. 运行`voc_annotation.py`文件,运行前将文件中的`classes`列表替换为自己数据集的class的名字 +6. 运行后自动生成`yolo_train.txt`文件,文件中每一行的格式为图片路径和图片中标注框的位置以及分类,如下所示 + `data/JPEGImages/MVI_20011__img00001.jpg 592,378,752,540,3 557,120,604,163,2 545,88,580,118,2 508,67,536,92,3 553,70,582,89,7 731,114,783,153,2 902,250,960,357,3` +7. 修改`train.py`中的路径参数与自己电脑上的路径相匹配 + + classes_path:`classes.txt`文件路径,文件格式如下: + ``` + Truck-Flatbed + MiniVan + Sedan + Suv + Truck-Pickup + Police + Truck-Box-Med + Taxi + ... + ``` + + model_path:预训练权重文件路径或者中断后继续训练时前一个保存的模型文件的路径 + + annotation_path:第6步生成的`yolo_train.txt`文件 +8. 调整`train.py`中的参数后运行文件进行训练 + + +## 预测 +1. 修改`model_path`的参数,选择合适的模型用于预测,并根据自己电脑情况选择是否启用cuda +2. 修改`predict.py`中的`mode`参数,进入不同的预测模式:`predict`表示单张图片预测,`video`表示视频检测 +3. 在predict模式下运行`predict.py`文件,输入预测图片的路径即可进行前向推理,最后通过PIL库输出处理后的图片与处理时间 +4. 在video模式下,需要预先修改video_name参数的值,将路径指向需要检测的视频,修改video_fps参数可以控制输出视频的fps;调整好后运行`predict.py`即可对视频进行预测 + +## 效果 + +![1](data/%EF%BC%91.jpg) + +![2](data/%EF%BC%92.jpg) \ No newline at end of file From 7e23c06e0e3168b998de8e1f83ee6c54d86fcdda Mon Sep 17 00:00:00 2001 From: aasdar <3415807009@qq.com> Date: Thu, 18 Dec 2025 19:42:00 +0800 Subject: [PATCH 2/5] =?UTF-8?q?=E6=B7=BB=E5=8A=A0=E6=95=B0=E6=8D=AE?= =?UTF-8?q?=E9=9B=86=E8=AE=AD=E7=BB=83=E4=BB=A3=E7=A0=81?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 1.添加了train.py文件 --- src/Yolov4_Vehicle_inspection/train.py | 332 +++++++++++++++++++++++++ 1 file changed, 332 insertions(+) create mode 100644 src/Yolov4_Vehicle_inspection/train.py diff --git a/src/Yolov4_Vehicle_inspection/train.py b/src/Yolov4_Vehicle_inspection/train.py new file mode 100644 index 0000000000..b2965c59cf --- /dev/null +++ b/src/Yolov4_Vehicle_inspection/train.py @@ -0,0 +1,332 @@ +#-------------------------------------# +# 对数据集进行训练 +#-------------------------------------# +import numpy as np +import torch +import torch.backends.cudnn as cudnn +import torch.optim as optim +from torch.utils.data import DataLoader +from tqdm import tqdm + +from nets.yolo4 import YoloBody +from nets.yolo_training import LossHistory, YOLOLoss, weights_init +from utils.dataloader import YoloDataset, yolo_dataset_collate + + +#---------------------------------------------------# +# 获得类和先验框 +#---------------------------------------------------# +def get_classes(classes_path): + '''loads the classes''' + with open(classes_path) as f: + class_names = f.readlines() + class_names = [c.strip() for c in class_names] + return class_names + +def get_anchors(anchors_path): + '''loads the anchors from a file''' + with open(anchors_path) as f: + anchors = f.readline() + anchors = [float(x) for x in anchors.split(',')] + return np.array(anchors).reshape([-1,3,2])[::-1,:,:] + +def get_lr(optimizer): + for param_group in optimizer.param_groups: + return param_group['lr'] + + +def fit_one_epoch(net,yolo_loss,epoch,epoch_size,epoch_size_val,gen,genval,Epoch,cuda): + if Tensorboard: + global train_tensorboard_step, val_tensorboard_step + total_loss = 0 + val_loss = 0 + + net.train() + print('Start Train') + with tqdm(total=epoch_size,desc=f'Epoch {epoch + 1}/{Epoch}',postfix=dict,mininterval=0.3) as pbar: + for iteration, batch in enumerate(gen): + if iteration >= epoch_size: + break + images, targets = batch[0], batch[1] + with torch.no_grad(): + if cuda: + images = torch.from_numpy(images).type(torch.FloatTensor).cuda() + targets = [torch.from_numpy(ann).type(torch.FloatTensor) for ann in targets] + else: + images = torch.from_numpy(images).type(torch.FloatTensor) + targets = [torch.from_numpy(ann).type(torch.FloatTensor) for ann in targets] + #----------------------# + # 清零梯度 + #----------------------# + optimizer.zero_grad() + #----------------------# + # 前向传播 + #----------------------# + outputs = net(images) + losses = [] + num_pos_all = 0 + #----------------------# + # 计算损失 + #----------------------# + for i in range(3): + loss_item, num_pos = yolo_loss(outputs[i], targets) + losses.append(loss_item) + num_pos_all += num_pos + + loss = sum(losses) / num_pos_all + total_loss += loss.item() + + #----------------------# + # 反向传播 + #----------------------# + loss.backward() + optimizer.step() + + if Tensorboard: + # 将loss写入tensorboard,每一步都写 + writer.add_scalar('Train_loss', loss, train_tensorboard_step) + train_tensorboard_step += 1 + + pbar.set_postfix(**{'total_loss': total_loss / (iteration + 1), + 'lr' : get_lr(optimizer)}) + pbar.update(1) + + # 将loss写入tensorboard,下面注释的是每个世代保存一次 + # if Tensorboard: + # writer.add_scalar('Train_loss', total_loss/(iteration+1), epoch) + net.eval() + print('Start Validation') + with tqdm(total=epoch_size_val, desc=f'Epoch {epoch + 1}/{Epoch}',postfix=dict,mininterval=0.3) as pbar: + for iteration, batch in enumerate(genval): + if iteration >= epoch_size_val: + break + images_val, targets_val = batch[0], batch[1] + + with torch.no_grad(): + if cuda: + images_val = torch.from_numpy(images_val).type(torch.FloatTensor).cuda() + targets_val = [torch.from_numpy(ann).type(torch.FloatTensor) for ann in targets_val] + else: + images_val = torch.from_numpy(images_val).type(torch.FloatTensor) + targets_val = [torch.from_numpy(ann).type(torch.FloatTensor) for ann in targets_val] + optimizer.zero_grad() + + outputs = net(images_val) + losses = [] + num_pos_all = 0 + for i in range(3): + loss_item, num_pos = yolo_loss(outputs[i], targets_val) + losses.append(loss_item) + num_pos_all += num_pos + loss = sum(losses) / num_pos_all + val_loss += loss.item() + + # 将loss写入tensorboard, 下面注释的是每一步都写 + # if Tensorboard: + # writer.add_scalar('Val_loss', loss, val_tensorboard_step) + # val_tensorboard_step += 1 + pbar.set_postfix(**{'total_loss': val_loss / (iteration + 1)}) + pbar.update(1) + + # 将loss写入tensorboard,每个世代保存一次 + if Tensorboard: + writer.add_scalar('Val_loss',val_loss / (epoch_size_val+1), epoch) + loss_history.append_loss(total_loss/(epoch_size+1), val_loss/(epoch_size_val+1)) + print('Finish Validation') + print('Epoch:'+ str(epoch+1) + '/' + str(Epoch)) + print('Total Loss: %.4f || Val Loss: %.4f ' % (total_loss/(epoch_size+1),val_loss/(epoch_size_val+1))) + print('Saving state, iter:', str(epoch+1)) + torch.save(model.state_dict(), 'logs/Epoch%d-Total_Loss%.4f-Val_Loss%.4f.pth'%((epoch+1),total_loss/(epoch_size+1),val_loss/(epoch_size_val+1))) + +#----------------------------------------------------# +# 检测精度mAP和pr曲线计算参考视频 +# https://www.bilibili.com/video/BV1zE411u7Vw +#----------------------------------------------------# +if __name__ == "__main__": + #-------------------------------# + # 是否使用Tensorboard + #-------------------------------# + Tensorboard = False + #-------------------------------# + # 是否使用Cuda + # 没有GPU可以设置成False + #-------------------------------# + Cuda = True + #------------------------------------------------------# + # 是否对损失进行归一化,用于改变loss的大小 + # 用于决定计算最终loss是除上batch_size还是除上正样本数量 + #------------------------------------------------------# + normalize = False + #-------------------------------# + # 输入的shape大小 + # 显存比较小可以使用416x416 + # 显存比较大可以使用608x608 + #-------------------------------# + input_shape = (416,416) + #----------------------------------------------------# + # classes和anchor的路径,非常重要 + # 训练前一定要修改classes_path,使其对应自己的数据集 + #----------------------------------------------------# + anchors_path = 'model_data/yolo_anchors.txt' + classes_path = 'model_data/voc_classes.txt' + #------------------------------------------------------# + # Yolov4的tricks应用 + # mosaic 马赛克数据增强 True or False + # 实际测试时mosaic数据增强并不稳定,所以默认为False + # Cosine_scheduler 余弦退火学习率 True or False + # label_smoothing 标签平滑 0.01以下一般 如0.01、0.005 + #------------------------------------------------------# + mosaic = False + Cosine_lr = False + smoooth_label = 0 + + #----------------------------------------------------# + # 获取classes和anchor + #----------------------------------------------------# + class_names = get_classes(classes_path) + anchors = get_anchors(anchors_path) + num_classes = len(class_names) + + #------------------------------------------------------# + # 创建yolo模型 + # 训练前一定要修改classes_path和对应的txt文件 + #------------------------------------------------------# + model = YoloBody(len(anchors[0]), num_classes) + weights_init(model) + + #------------------------------------------------------# + # 权值文件请看README,百度网盘下载 + #------------------------------------------------------# + model_path = "model_data/yolo4_weights.pth" + print('Loading weights into state dict...') + device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') + model_dict = model.state_dict() + pretrained_dict = torch.load(model_path, map_location=device) + pretrained_dict = {k: v for k, v in pretrained_dict.items() if np.shape(model_dict[k]) == np.shape(v)} + model_dict.update(pretrained_dict) + model.load_state_dict(model_dict) + print('Finished!') + # torch.save(model.state_dict(),"logs/test.pth",_use_new_zipfile_serialization=False ) + + net = model.train() + + if Cuda: + net = torch.nn.DataParallel(model) + cudnn.benchmark = True + net = net.cuda() + + yolo_loss = YOLOLoss(np.reshape(anchors,[-1,2]), num_classes, (input_shape[1], input_shape[0]), smoooth_label, Cuda, normalize) + loss_history = LossHistory("logs/") + + #----------------------------------------------------# + # 获得图片路径和标签 + #----------------------------------------------------# + annotation_path = '2007_train.txt' + #----------------------------------------------------------------------# + # 验证集的划分在train.py代码里面进行 + # 2007_test.txt和2007_val.txt里面没有内容是正常的。训练不会使用到。 + # 当前划分方式下,验证集和训练集的比例为1:9 + #----------------------------------------------------------------------# + val_split = 0.1 + with open(annotation_path) as f: + lines = f.readlines() + np.random.seed(10101) + np.random.shuffle(lines) + np.random.seed(None) + num_val = int(len(lines)*val_split) + num_train = len(lines) - num_val + + if Tensorboard: + from tensorboardX import SummaryWriter + writer = SummaryWriter(log_dir='logs',flush_secs=60) + if Cuda: + graph_inputs = torch.randn(1,3,input_shape[0],input_shape[1]).type(torch.FloatTensor).cuda() + else: + graph_inputs = torch.randn(1,3,input_shape[0],input_shape[1]).type(torch.FloatTensor) + writer.add_graph(model, graph_inputs) + train_tensorboard_step = 1 + val_tensorboard_step = 1 + + #------------------------------------------------------# + # 主干特征提取网络特征通用,冻结训练可以加快训练速度 + # 也可以在训练初期防止权值被破坏。 + # Init_Epoch为起始世代 + # Freeze_Epoch为冻结训练的世代 + # Epoch总训练世代 + # 提示OOM或者显存不足请调小Batch_size + #------------------------------------------------------# + if True: + lr = 1e-3 + Batch_size = 4 + Init_Epoch = 0 + Freeze_Epoch = 50 + + #----------------------------------------------------------------------------# + # 我在实际测试时,发现optimizer的weight_decay起到了反作用, + # 所以去除掉了weight_decay,大家也可以开起来试试,一般是weight_decay=5e-4 + #----------------------------------------------------------------------------# + optimizer = optim.Adam(net.parameters(),lr) + if Cosine_lr: + lr_scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=5, eta_min=1e-5) + else: + lr_scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=1, gamma=0.92) + + train_dataset = YoloDataset(lines[:num_train], (input_shape[0], input_shape[1]), mosaic=mosaic, is_train=True) + val_dataset = YoloDataset(lines[num_train:], (input_shape[0], input_shape[1]), mosaic=False, is_train=False) + gen = DataLoader(train_dataset, shuffle=True, batch_size=Batch_size, num_workers=4, pin_memory=True, + drop_last=True, collate_fn=yolo_dataset_collate) + gen_val = DataLoader(val_dataset, shuffle=True, batch_size=Batch_size, num_workers=4,pin_memory=True, + drop_last=True, collate_fn=yolo_dataset_collate) + + epoch_size = num_train // Batch_size + epoch_size_val = num_val // Batch_size + + if epoch_size == 0 or epoch_size_val == 0: + raise ValueError("数据集过小,无法进行训练,请扩充数据集。") + #------------------------------------# + # 冻结一定部分训练 + #------------------------------------# + for param in model.backbone.parameters(): + param.requires_grad = False + + for epoch in range(Init_Epoch,Freeze_Epoch): + fit_one_epoch(net,yolo_loss,epoch,epoch_size,epoch_size_val,gen,gen_val,Freeze_Epoch,Cuda) + lr_scheduler.step() + + if True: + lr = 1e-4 + Batch_size = 2 + Freeze_Epoch = 50 + Unfreeze_Epoch = 100 + + #----------------------------------------------------------------------------# + # 我在实际测试时,发现optimizer的weight_decay起到了反作用, + # 所以去除掉了weight_decay,大家也可以开起来试试,一般是weight_decay=5e-4 + #----------------------------------------------------------------------------# + optimizer = optim.Adam(net.parameters(),lr) + if Cosine_lr: + lr_scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=5, eta_min=1e-5) + else: + lr_scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=1, gamma=0.92) + + train_dataset = YoloDataset(lines[:num_train], (input_shape[0], input_shape[1]), mosaic=mosaic, is_train=True) + val_dataset = YoloDataset(lines[num_train:], (input_shape[0], input_shape[1]), mosaic=False, is_train=False) + gen = DataLoader(train_dataset, shuffle=True, batch_size=Batch_size, num_workers=4, pin_memory=True, + drop_last=True, collate_fn=yolo_dataset_collate) + gen_val = DataLoader(val_dataset, shuffle=True, batch_size=Batch_size, num_workers=4,pin_memory=True, + drop_last=True, collate_fn=yolo_dataset_collate) + + epoch_size = num_train // Batch_size + epoch_size_val = num_val // Batch_size + + if epoch_size == 0 or epoch_size_val == 0: + raise ValueError("数据集过小,无法进行训练,请扩充数据集。") + #------------------------------------# + # 解冻后训练 + #------------------------------------# + for param in model.backbone.parameters(): + param.requires_grad = True + + for epoch in range(Freeze_Epoch,Unfreeze_Epoch): + fit_one_epoch(net,yolo_loss,epoch,epoch_size,epoch_size_val,gen,gen_val,Unfreeze_Epoch,Cuda) + lr_scheduler.step() From c81a417a11984603f18d4b865516f5492b84e3ea Mon Sep 17 00:00:00 2001 From: aasdar <3415807009@qq.com> Date: Thu, 18 Dec 2025 19:56:27 +0800 Subject: [PATCH 3/5] =?UTF-8?q?=E6=B7=BB=E5=8A=A0VOCdevkit=E6=96=87?= =?UTF-8?q?=E4=BB=B6=E5=A4=B9?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 1.Annotations存放标签文件 2.ImageSets存放训练索引文件 3.JPEGImages存放图片文件 --- .../VOCdevkit/VOC2007/Annotations/README.md | 1 + .../VOCdevkit/VOC2007/ImageSets/README.md | 1 + .../VOCdevkit/VOC2007/JPEGImages/README.md | 1 + 3 files changed, 3 insertions(+) create mode 100644 src/Yolov4_Vehicle_inspection/VOCdevkit/VOC2007/Annotations/README.md create mode 100644 src/Yolov4_Vehicle_inspection/VOCdevkit/VOC2007/ImageSets/README.md create mode 100644 src/Yolov4_Vehicle_inspection/VOCdevkit/VOC2007/JPEGImages/README.md diff --git a/src/Yolov4_Vehicle_inspection/VOCdevkit/VOC2007/Annotations/README.md b/src/Yolov4_Vehicle_inspection/VOCdevkit/VOC2007/Annotations/README.md new file mode 100644 index 0000000000..d73a1916c9 --- /dev/null +++ b/src/Yolov4_Vehicle_inspection/VOCdevkit/VOC2007/Annotations/README.md @@ -0,0 +1 @@ +存放标签文件 \ No newline at end of file diff --git a/src/Yolov4_Vehicle_inspection/VOCdevkit/VOC2007/ImageSets/README.md b/src/Yolov4_Vehicle_inspection/VOCdevkit/VOC2007/ImageSets/README.md new file mode 100644 index 0000000000..2f64921067 --- /dev/null +++ b/src/Yolov4_Vehicle_inspection/VOCdevkit/VOC2007/ImageSets/README.md @@ -0,0 +1 @@ +存放训练索引文件 \ No newline at end of file diff --git a/src/Yolov4_Vehicle_inspection/VOCdevkit/VOC2007/JPEGImages/README.md b/src/Yolov4_Vehicle_inspection/VOCdevkit/VOC2007/JPEGImages/README.md new file mode 100644 index 0000000000..58f32ea1aa --- /dev/null +++ b/src/Yolov4_Vehicle_inspection/VOCdevkit/VOC2007/JPEGImages/README.md @@ -0,0 +1 @@ +存放图片文件 \ No newline at end of file From 150d737e2811f41ad50737bc9b3e59eec1d76f67 Mon Sep 17 00:00:00 2001 From: aasdar <3415807009@qq.com> Date: Thu, 18 Dec 2025 20:09:13 +0800 Subject: [PATCH 4/5] Delete src/Yolov4_Vehicle_inspection/train.py --- src/Yolov4_Vehicle_inspection/train.py | 332 ------------------------- 1 file changed, 332 deletions(-) delete mode 100644 src/Yolov4_Vehicle_inspection/train.py diff --git a/src/Yolov4_Vehicle_inspection/train.py b/src/Yolov4_Vehicle_inspection/train.py deleted file mode 100644 index b2965c59cf..0000000000 --- a/src/Yolov4_Vehicle_inspection/train.py +++ /dev/null @@ -1,332 +0,0 @@ -#-------------------------------------# -# 对数据集进行训练 -#-------------------------------------# -import numpy as np -import torch -import torch.backends.cudnn as cudnn -import torch.optim as optim -from torch.utils.data import DataLoader -from tqdm import tqdm - -from nets.yolo4 import YoloBody -from nets.yolo_training import LossHistory, YOLOLoss, weights_init -from utils.dataloader import YoloDataset, yolo_dataset_collate - - -#---------------------------------------------------# -# 获得类和先验框 -#---------------------------------------------------# -def get_classes(classes_path): - '''loads the classes''' - with open(classes_path) as f: - class_names = f.readlines() - class_names = [c.strip() for c in class_names] - return class_names - -def get_anchors(anchors_path): - '''loads the anchors from a file''' - with open(anchors_path) as f: - anchors = f.readline() - anchors = [float(x) for x in anchors.split(',')] - return np.array(anchors).reshape([-1,3,2])[::-1,:,:] - -def get_lr(optimizer): - for param_group in optimizer.param_groups: - return param_group['lr'] - - -def fit_one_epoch(net,yolo_loss,epoch,epoch_size,epoch_size_val,gen,genval,Epoch,cuda): - if Tensorboard: - global train_tensorboard_step, val_tensorboard_step - total_loss = 0 - val_loss = 0 - - net.train() - print('Start Train') - with tqdm(total=epoch_size,desc=f'Epoch {epoch + 1}/{Epoch}',postfix=dict,mininterval=0.3) as pbar: - for iteration, batch in enumerate(gen): - if iteration >= epoch_size: - break - images, targets = batch[0], batch[1] - with torch.no_grad(): - if cuda: - images = torch.from_numpy(images).type(torch.FloatTensor).cuda() - targets = [torch.from_numpy(ann).type(torch.FloatTensor) for ann in targets] - else: - images = torch.from_numpy(images).type(torch.FloatTensor) - targets = [torch.from_numpy(ann).type(torch.FloatTensor) for ann in targets] - #----------------------# - # 清零梯度 - #----------------------# - optimizer.zero_grad() - #----------------------# - # 前向传播 - #----------------------# - outputs = net(images) - losses = [] - num_pos_all = 0 - #----------------------# - # 计算损失 - #----------------------# - for i in range(3): - loss_item, num_pos = yolo_loss(outputs[i], targets) - losses.append(loss_item) - num_pos_all += num_pos - - loss = sum(losses) / num_pos_all - total_loss += loss.item() - - #----------------------# - # 反向传播 - #----------------------# - loss.backward() - optimizer.step() - - if Tensorboard: - # 将loss写入tensorboard,每一步都写 - writer.add_scalar('Train_loss', loss, train_tensorboard_step) - train_tensorboard_step += 1 - - pbar.set_postfix(**{'total_loss': total_loss / (iteration + 1), - 'lr' : get_lr(optimizer)}) - pbar.update(1) - - # 将loss写入tensorboard,下面注释的是每个世代保存一次 - # if Tensorboard: - # writer.add_scalar('Train_loss', total_loss/(iteration+1), epoch) - net.eval() - print('Start Validation') - with tqdm(total=epoch_size_val, desc=f'Epoch {epoch + 1}/{Epoch}',postfix=dict,mininterval=0.3) as pbar: - for iteration, batch in enumerate(genval): - if iteration >= epoch_size_val: - break - images_val, targets_val = batch[0], batch[1] - - with torch.no_grad(): - if cuda: - images_val = torch.from_numpy(images_val).type(torch.FloatTensor).cuda() - targets_val = [torch.from_numpy(ann).type(torch.FloatTensor) for ann in targets_val] - else: - images_val = torch.from_numpy(images_val).type(torch.FloatTensor) - targets_val = [torch.from_numpy(ann).type(torch.FloatTensor) for ann in targets_val] - optimizer.zero_grad() - - outputs = net(images_val) - losses = [] - num_pos_all = 0 - for i in range(3): - loss_item, num_pos = yolo_loss(outputs[i], targets_val) - losses.append(loss_item) - num_pos_all += num_pos - loss = sum(losses) / num_pos_all - val_loss += loss.item() - - # 将loss写入tensorboard, 下面注释的是每一步都写 - # if Tensorboard: - # writer.add_scalar('Val_loss', loss, val_tensorboard_step) - # val_tensorboard_step += 1 - pbar.set_postfix(**{'total_loss': val_loss / (iteration + 1)}) - pbar.update(1) - - # 将loss写入tensorboard,每个世代保存一次 - if Tensorboard: - writer.add_scalar('Val_loss',val_loss / (epoch_size_val+1), epoch) - loss_history.append_loss(total_loss/(epoch_size+1), val_loss/(epoch_size_val+1)) - print('Finish Validation') - print('Epoch:'+ str(epoch+1) + '/' + str(Epoch)) - print('Total Loss: %.4f || Val Loss: %.4f ' % (total_loss/(epoch_size+1),val_loss/(epoch_size_val+1))) - print('Saving state, iter:', str(epoch+1)) - torch.save(model.state_dict(), 'logs/Epoch%d-Total_Loss%.4f-Val_Loss%.4f.pth'%((epoch+1),total_loss/(epoch_size+1),val_loss/(epoch_size_val+1))) - -#----------------------------------------------------# -# 检测精度mAP和pr曲线计算参考视频 -# https://www.bilibili.com/video/BV1zE411u7Vw -#----------------------------------------------------# -if __name__ == "__main__": - #-------------------------------# - # 是否使用Tensorboard - #-------------------------------# - Tensorboard = False - #-------------------------------# - # 是否使用Cuda - # 没有GPU可以设置成False - #-------------------------------# - Cuda = True - #------------------------------------------------------# - # 是否对损失进行归一化,用于改变loss的大小 - # 用于决定计算最终loss是除上batch_size还是除上正样本数量 - #------------------------------------------------------# - normalize = False - #-------------------------------# - # 输入的shape大小 - # 显存比较小可以使用416x416 - # 显存比较大可以使用608x608 - #-------------------------------# - input_shape = (416,416) - #----------------------------------------------------# - # classes和anchor的路径,非常重要 - # 训练前一定要修改classes_path,使其对应自己的数据集 - #----------------------------------------------------# - anchors_path = 'model_data/yolo_anchors.txt' - classes_path = 'model_data/voc_classes.txt' - #------------------------------------------------------# - # Yolov4的tricks应用 - # mosaic 马赛克数据增强 True or False - # 实际测试时mosaic数据增强并不稳定,所以默认为False - # Cosine_scheduler 余弦退火学习率 True or False - # label_smoothing 标签平滑 0.01以下一般 如0.01、0.005 - #------------------------------------------------------# - mosaic = False - Cosine_lr = False - smoooth_label = 0 - - #----------------------------------------------------# - # 获取classes和anchor - #----------------------------------------------------# - class_names = get_classes(classes_path) - anchors = get_anchors(anchors_path) - num_classes = len(class_names) - - #------------------------------------------------------# - # 创建yolo模型 - # 训练前一定要修改classes_path和对应的txt文件 - #------------------------------------------------------# - model = YoloBody(len(anchors[0]), num_classes) - weights_init(model) - - #------------------------------------------------------# - # 权值文件请看README,百度网盘下载 - #------------------------------------------------------# - model_path = "model_data/yolo4_weights.pth" - print('Loading weights into state dict...') - device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') - model_dict = model.state_dict() - pretrained_dict = torch.load(model_path, map_location=device) - pretrained_dict = {k: v for k, v in pretrained_dict.items() if np.shape(model_dict[k]) == np.shape(v)} - model_dict.update(pretrained_dict) - model.load_state_dict(model_dict) - print('Finished!') - # torch.save(model.state_dict(),"logs/test.pth",_use_new_zipfile_serialization=False ) - - net = model.train() - - if Cuda: - net = torch.nn.DataParallel(model) - cudnn.benchmark = True - net = net.cuda() - - yolo_loss = YOLOLoss(np.reshape(anchors,[-1,2]), num_classes, (input_shape[1], input_shape[0]), smoooth_label, Cuda, normalize) - loss_history = LossHistory("logs/") - - #----------------------------------------------------# - # 获得图片路径和标签 - #----------------------------------------------------# - annotation_path = '2007_train.txt' - #----------------------------------------------------------------------# - # 验证集的划分在train.py代码里面进行 - # 2007_test.txt和2007_val.txt里面没有内容是正常的。训练不会使用到。 - # 当前划分方式下,验证集和训练集的比例为1:9 - #----------------------------------------------------------------------# - val_split = 0.1 - with open(annotation_path) as f: - lines = f.readlines() - np.random.seed(10101) - np.random.shuffle(lines) - np.random.seed(None) - num_val = int(len(lines)*val_split) - num_train = len(lines) - num_val - - if Tensorboard: - from tensorboardX import SummaryWriter - writer = SummaryWriter(log_dir='logs',flush_secs=60) - if Cuda: - graph_inputs = torch.randn(1,3,input_shape[0],input_shape[1]).type(torch.FloatTensor).cuda() - else: - graph_inputs = torch.randn(1,3,input_shape[0],input_shape[1]).type(torch.FloatTensor) - writer.add_graph(model, graph_inputs) - train_tensorboard_step = 1 - val_tensorboard_step = 1 - - #------------------------------------------------------# - # 主干特征提取网络特征通用,冻结训练可以加快训练速度 - # 也可以在训练初期防止权值被破坏。 - # Init_Epoch为起始世代 - # Freeze_Epoch为冻结训练的世代 - # Epoch总训练世代 - # 提示OOM或者显存不足请调小Batch_size - #------------------------------------------------------# - if True: - lr = 1e-3 - Batch_size = 4 - Init_Epoch = 0 - Freeze_Epoch = 50 - - #----------------------------------------------------------------------------# - # 我在实际测试时,发现optimizer的weight_decay起到了反作用, - # 所以去除掉了weight_decay,大家也可以开起来试试,一般是weight_decay=5e-4 - #----------------------------------------------------------------------------# - optimizer = optim.Adam(net.parameters(),lr) - if Cosine_lr: - lr_scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=5, eta_min=1e-5) - else: - lr_scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=1, gamma=0.92) - - train_dataset = YoloDataset(lines[:num_train], (input_shape[0], input_shape[1]), mosaic=mosaic, is_train=True) - val_dataset = YoloDataset(lines[num_train:], (input_shape[0], input_shape[1]), mosaic=False, is_train=False) - gen = DataLoader(train_dataset, shuffle=True, batch_size=Batch_size, num_workers=4, pin_memory=True, - drop_last=True, collate_fn=yolo_dataset_collate) - gen_val = DataLoader(val_dataset, shuffle=True, batch_size=Batch_size, num_workers=4,pin_memory=True, - drop_last=True, collate_fn=yolo_dataset_collate) - - epoch_size = num_train // Batch_size - epoch_size_val = num_val // Batch_size - - if epoch_size == 0 or epoch_size_val == 0: - raise ValueError("数据集过小,无法进行训练,请扩充数据集。") - #------------------------------------# - # 冻结一定部分训练 - #------------------------------------# - for param in model.backbone.parameters(): - param.requires_grad = False - - for epoch in range(Init_Epoch,Freeze_Epoch): - fit_one_epoch(net,yolo_loss,epoch,epoch_size,epoch_size_val,gen,gen_val,Freeze_Epoch,Cuda) - lr_scheduler.step() - - if True: - lr = 1e-4 - Batch_size = 2 - Freeze_Epoch = 50 - Unfreeze_Epoch = 100 - - #----------------------------------------------------------------------------# - # 我在实际测试时,发现optimizer的weight_decay起到了反作用, - # 所以去除掉了weight_decay,大家也可以开起来试试,一般是weight_decay=5e-4 - #----------------------------------------------------------------------------# - optimizer = optim.Adam(net.parameters(),lr) - if Cosine_lr: - lr_scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=5, eta_min=1e-5) - else: - lr_scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=1, gamma=0.92) - - train_dataset = YoloDataset(lines[:num_train], (input_shape[0], input_shape[1]), mosaic=mosaic, is_train=True) - val_dataset = YoloDataset(lines[num_train:], (input_shape[0], input_shape[1]), mosaic=False, is_train=False) - gen = DataLoader(train_dataset, shuffle=True, batch_size=Batch_size, num_workers=4, pin_memory=True, - drop_last=True, collate_fn=yolo_dataset_collate) - gen_val = DataLoader(val_dataset, shuffle=True, batch_size=Batch_size, num_workers=4,pin_memory=True, - drop_last=True, collate_fn=yolo_dataset_collate) - - epoch_size = num_train // Batch_size - epoch_size_val = num_val // Batch_size - - if epoch_size == 0 or epoch_size_val == 0: - raise ValueError("数据集过小,无法进行训练,请扩充数据集。") - #------------------------------------# - # 解冻后训练 - #------------------------------------# - for param in model.backbone.parameters(): - param.requires_grad = True - - for epoch in range(Freeze_Epoch,Unfreeze_Epoch): - fit_one_epoch(net,yolo_loss,epoch,epoch_size,epoch_size_val,gen,gen_val,Unfreeze_Epoch,Cuda) - lr_scheduler.step() From 7ad6454f0e073cf96ccb831842c03818562a1b92 Mon Sep 17 00:00:00 2001 From: aasdar <3415807009@qq.com> Date: Thu, 18 Dec 2025 20:29:43 +0800 Subject: [PATCH 5/5] =?UTF-8?q?=E6=B7=BB=E5=8A=A0=E6=B3=A8=E9=87=8A?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 1.给predict.py添加更详细的注释 --- src/Yolov4_Vehicle_inspection/predict.py | 122 +++++++++++++++-------- 1 file changed, 82 insertions(+), 40 deletions(-) diff --git a/src/Yolov4_Vehicle_inspection/predict.py b/src/Yolov4_Vehicle_inspection/predict.py index 833ba458de..2ac46d0042 100644 --- a/src/Yolov4_Vehicle_inspection/predict.py +++ b/src/Yolov4_Vehicle_inspection/predict.py @@ -3,23 +3,25 @@ # 将单张图片预测、摄像头检测和FPS测试功能 # 整合到了一个py文件中,通过指定mode进行模式的修改。 # ----------------------------------------------------# -import time +import time # 导入时间模块,用于计时和FPS计算 +import cv2 # 导入OpenCV库,用于视频/图像的读取、处理和保存 +import numpy as np # 导入numpy库,用于数组格式转换 +from PIL import Image # 导入PIL库的Image模块,用于图像格式转换 -import cv2 -import numpy as np -from PIL import Image - -from yolo import YOLO +from yolo import YOLO # 导入自定义的YOLO检测类 if __name__ == "__main__": + # 初始化YOLO检测器,设置输入图像尺寸为640x640 yolo = YOLO(size=640) + # -------------------------------------------------------------------------# # mode用于指定测试的模式: # 'predict'表示单张图片预测 # 'video'表示视频检测 # 'fps'表示测试fps # -------------------------------------------------------------------------# - mode = "video" + mode = "video" # 当前运行模式设置为视频检测 + # -------------------------------------------------------------------------# # video_path用于指定视频的路径,当video_path=0时表示检测摄像头 # video_save_path表示视频保存的路径,当video_save_path=""时表示不保存 @@ -27,10 +29,10 @@ # video_path、video_save_path和video_fps仅在mode='video'时有效 # 保存视频时需要ctrl+c退出才会完成完整的保存步骤,不可直接结束程序。 # -------------------------------------------------------------------------# - video_name = "333" - video_path = r'vedio\int\%s.mp4' % video_name - video_save_path = r"vedio\out\%s.mp4" % video_name - video_fps = 30 + video_name = "333" # 视频文件名称(不含后缀) + video_path = r'vedio\int\%s.mp4' % video_name # 输入视频的路径 + video_save_path = r"vedio\out\%s.mp4" % video_name # 输出视频的保存路径 + video_fps = 30 # 保存视频的帧率 if mode == "predict": ''' @@ -43,82 +45,122 @@ 5、如果想要在预测图上写额外的字,比如检测到的特定目标的数量,可以进入yolo.detect_image函数,在绘图部分对predicted_class进行判断, 比如判断if predicted_class == 'car': 即可判断当前目标是否为车,然后记录数量即可。利用draw.text即可写字。 ''' + # 循环读取用户输入的图片路径,实现多次单张图片预测 while True: - img = input('Input image filename:') - start = time.time() + img = input('Input image filename:') # 提示用户输入图片文件路径 + start = time.time() # 记录图片检测开始时间 try: + # 尝试打开指定路径的图片文件 image = Image.open(img) except: + # 打开失败时提示错误并重新输入 print('Open Error! Try again!') continue else: + # 图片打开成功,调用YOLO的检测函数进行目标检测 r_image = yolo.detect_image(image) + # 打印单张图片检测耗时 print("time : ", time.time() - start) + # 显示检测后的图片 r_image.show() elif mode == "video": + # 初始化视频读取器,读取指定路径的视频文件 capture = cv2.VideoCapture(video_path) - start = time.time() + start = time.time() # 记录视频检测总开始时间 + + # 如果设置了视频保存路径,则初始化视频写入器 if video_save_path != "": + # 设置视频编码格式为mp4v(兼容mp4格式) fourcc = cv2.VideoWriter_fourcc(*'mp4v') + # 获取原视频的分辨率(宽度和高度) size = (int(capture.get(cv2.CAP_PROP_FRAME_WIDTH)), int(capture.get(cv2.CAP_PROP_FRAME_HEIGHT))) + # 创建视频写入器,指定保存路径、编码格式、帧率和分辨率 out = cv2.VideoWriter(video_save_path, fourcc, video_fps, size) - fps = 0.0 - time_cnt = 0 - cars = [] - car_not_vis = [] + # 初始化FPS相关参数 + fps = 0.0 # 实时帧率(滑动平均) + time_cnt = 0 # 帧计数,记录已处理的帧数 + cars = [] # 存储当前可见的车辆目标信息 + car_not_vis = [] # 存储不可见的车辆目标信息 + + # 循环读取视频帧,直到视频结束或用户退出 while (True): - t1 = time.time() - # 读取某一帧 - ref, frame = capture.read() + t1 = time.time() # 记录当前帧处理开始时间 + # 读取视频的一帧,ref为读取成功标志,frame为帧数据 + ref, frame = capture.read() - if not ref: # 防止视频读取完/帧为空导致后续报错 + # 若读取失败(如视频结束、帧为空),跳出循环 + if not ref: break - - # 调整对比度(alpha)和亮度(beta),可根据实际视频微调: + # ====================== 图像预处理:对比度和亮度调整 ====================== # alpha>1增强对比度,alpha<1降低对比度;beta>0增亮,beta<0调暗 - alpha = 1.2 # 对比度增强20%(可改1.1~1.5) - beta = 8 # 亮度提升8(可改5~15) + alpha = 1.2 # 对比度增强20%(可根据实际视频调整1.1~1.5) + beta = 8 # 亮度提升8(可根据实际视频调整5~15) + # 调整帧的对比度和亮度,提升检测效果 frame = cv2.convertScaleAbs(frame, alpha=alpha, beta=beta) - # 格式转变,BGRtoRGB - frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB) # 原try-except可删除,因为已提前判断ref - - # 转变成Image + # ====================== 格式转换:适配YOLO检测要求 ====================== + # OpenCV读取的帧是BGR格式,转换为RGB格式(YOLO检测要求RGB) + frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB) + # 将numpy数组格式的帧转换为PIL的Image格式(YOLO检测函数入参要求) frame = Image.fromarray(np.uint8(frame)) - # 进行检测 + + # ====================== 目标检测:调用YOLO检测函数 ====================== + # 传入帧数据和车辆跟踪列表,执行目标检测并返回标注后的帧 frame = np.array(yolo.detect_image(frame, cars, car_not_vis)) - # RGBtoBGR满足opencv显示格式 + + # ====================== 格式还原:适配OpenCV显示/保存 ====================== + # 将RGB格式转换回BGR格式(OpenCV显示/保存要求BGR) frame = cv2.cvtColor(frame, cv2.COLOR_RGB2BGR) + # ====================== FPS计算与信息标注 ====================== + # 滑动平均计算实时FPS:(上一帧FPS + 当前帧FPS) / 2 fps = (fps + (1. / (time.time() - t1))) / 2 + # 打印实时帧号、FPS和总耗时(\r实现单行覆盖打印) print("\rframe_id = %d\tfps = %.2f\ttime = %.2f" % (time_cnt, fps, time.time() - start), end="") - time_cnt += 1 + time_cnt += 1 # 帧计数+1 + + # 在帧上标注实时FPS(位置:左上角,字体:默认,大小:1,颜色:绿色,线宽:2) frame = cv2.putText(frame, "fps = %.2f" % (fps), (0, 40), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 0), 2) + # 在帧上标注当前检测到的车辆数量(位置:FPS下方,样式同上) frame = cv2.putText(frame, "cars_num = %d" % (len(cars)), (0, 80), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 0), 2) + # 可选:显示视频窗口(注释掉以提升处理速度) # cv2.imshow("video",frame) + + # 等待1ms,检测键盘输入(ESC键退出) c = cv2.waitKey(1) & 0xff + + # 如果设置了保存路径,将标注后的帧写入视频文件 if video_save_path != "": out.write(frame) - if c == 27: # 按ESC退出 + # 按ESC键(ASCII码27)退出视频检测循环 + if c == 27: break - # print("time: ", time.time() - start) + + # 换行,避免打印信息重叠 print() - capture.release() - if video_save_path != "": # 防止未保存视频时out未定义 + + # ====================== 资源释放 ====================== + capture.release() # 释放视频读取器资源 + # 若初始化了视频写入器,释放其资源 + if video_save_path != "": out.release() - cv2.destroyAllWindows() + cv2.destroyAllWindows() # 关闭所有OpenCV创建的窗口 elif mode == "fps": - test_interval = 100 - img = Image.open('img/street.jpg') + # FPS测试模式:计算模型处理单张图片的平均耗时和帧率 + test_interval = 100 # 测试次数(连续检测100次取平均) + img = Image.open('img/street.jpg') # 读取测试用图片 + # 调用YOLO的FPS测试函数,返回单次检测平均耗时(秒) tact_time = yolo.get_FPS(img, test_interval) + # 打印测试结果:单次耗时、帧率(1/耗时)、批量大小为1 print(str(tact_time) + ' seconds, ' + str(1 / tact_time) + 'FPS, @batch_size 1') else: + # 模式指定错误时抛出断言异常,提示正确的模式选项 raise AssertionError("Please specify the correct mode: 'predict', 'video' or 'fps'.") \ No newline at end of file