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 1/4] =?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 2/4] =?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 3/4] =?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 4/4] 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()