diff --git a/src/Yolov4_Vehicle_inspection/train.py b/src/Yolov4_Vehicle_inspection/train.py index 8ec0b3fe18..fed8859680 100644 --- a/src/Yolov4_Vehicle_inspection/train.py +++ b/src/Yolov4_Vehicle_inspection/train.py @@ -1,332 +1,425 @@ -#-------------------------------------# +# -------------------------------------# # 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 tqdm import tqdm # 用于显示训练进度条的工具库 -from nets.yolo4 import YoloBody -from nets.yolo_training import LossHistory, YOLOLoss, weights_init -from utils.dataloader import YoloDataset, yolo_dataset_collate +from nets.yolo4 import YoloBody # 导入YOLOv4的网络结构定义 +from nets.yolo_training import LossHistory, YOLOLoss, weights_init # 导入训练相关的损失函数、历史记录保存等工具 +from utils.dataloader import YoloDataset, yolo_dataset_collate # 导入自定义的数据加载器和数据拼接函数 -#---------------------------------------------------# +# ---------------------------------------------------# # 获得类和先验框 -#---------------------------------------------------# +# ---------------------------------------------------# def get_classes(classes_path): '''loads the classes''' + # 读取类别名称文件,该文件每行对应一个类别名称 with open(classes_path) as f: class_names = f.readlines() + # 去除每个类别名称前后的空白字符(如换行符、空格) class_names = [c.strip() for c in class_names] return class_names + def get_anchors(anchors_path): '''loads the anchors from a file''' + # 读取先验框文件,该文件中一行存储所有先验框的宽高(格式为:w1,h1,w2,h2,...w9,h9) with open(anchors_path) as f: anchors = f.readline() + # 将读取到的字符串转换为浮点型列表 anchors = [float(x) for x in anchors.split(',')] - return np.array(anchors).reshape([-1,3,2])[::-1,:,:] + # 重塑为三维数组:[3,3,2],其中3个尺度层,每个尺度层3个先验框,每个先验框2个值(宽、高) + # [::-1,:,:] 表示对尺度层进行逆序,适配YOLOv4的大、中、小尺度输出顺序 + return np.array(anchors).reshape([-1, 3, 2])[::-1, :, :] + def get_lr(optimizer): + # 获取当前优化器的学习率(遍历参数组,返回第一个参数组的学习率) 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): + +def fit_one_epoch(net, yolo_loss, epoch, epoch_size, epoch_size_val, gen, genval, Epoch, cuda): + # 全局变量声明,用于Tensorboard的步数计数(避免每次调用函数时重置) if Tensorboard: global train_tensorboard_step, val_tensorboard_step + # 初始化训练总损失和验证总损失 total_loss = 0 val_loss = 0 + # 将网络设置为训练模式(启用批量归一化、dropout等训练专属层) net.train() print('Start Train') - with tqdm(total=epoch_size,desc=f'Epoch {epoch + 1}/{Epoch}',postfix=dict,mininterval=0.3) as pbar: + # 使用tqdm创建训练进度条,显示当前epoch信息、实时损失和学习率 + with tqdm(total=epoch_size, desc=f'Epoch {epoch + 1}/{Epoch}', postfix=dict, mininterval=0.3) as pbar: for iteration, batch in enumerate(gen): + # 若迭代次数超过当前epoch的总步数,提前终止训练循环 if iteration >= epoch_size: break + # 从数据加载器中获取当前批次的图像和标签 images, targets = batch[0], batch[1] + # torch.no_grad() 上下文管理器:禁用梯度计算,减少显存占用(此处仅用于数据转换,无需梯度) 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] + # 将numpy数组转换为torch浮点张量,并移至GPU + images = torch.from_numpy(images).type(torch.FloatTensor).cuda() + targets = [torch.from_numpy(ann).type(torch.FloatTensor).cuda() for ann in targets] else: - images = torch.from_numpy(images).type(torch.FloatTensor) + # 将numpy数组转换为torch浮点张量,保留在CPU + images = torch.from_numpy(images).type(torch.FloatTensor) targets = [torch.from_numpy(ann).type(torch.FloatTensor) for ann in targets] - #----------------------# + # ----------------------# # 清零梯度 - #----------------------# + # ----------------------# + # 每次反向传播前必须清零梯度,避免梯度累积 optimizer.zero_grad() - #----------------------# + # ----------------------# # 前向传播 - #----------------------# + # ----------------------# + # 将图像输入网络,得到3个尺度层的预测输出 outputs = net(images) losses = [] num_pos_all = 0 - #----------------------# + # ----------------------# # 计算损失 - #----------------------# + # ----------------------# + # 遍历3个尺度层的输出,分别计算每个尺度的损失 for i in range(3): loss_item, num_pos = yolo_loss(outputs[i], targets) - losses.append(loss_item) - num_pos_all += num_pos + losses.append(loss_item) # 保存每个尺度的损失 + num_pos_all += num_pos # 累计所有尺度的正样本数量 + # 总损失 = 所有尺度损失之和 / 正样本总数(归一化损失,避免批次大小影响) loss = sum(losses) / num_pos_all + # 累加当前批次的损失到训练总损失 total_loss += loss.item() - #----------------------# + # ----------------------# # 反向传播 - #----------------------# + # ----------------------# + # 根据损失计算梯度(反向传播) loss.backward() + # 根据梯度更新网络参数(优化器步骤) optimizer.step() + # 若启用Tensorboard,记录训练损失(按步骤记录) 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)}) + # 更新进度条的实时显示信息:平均训练损失和当前学习率 + pbar.set_postfix(**{'total_loss': total_loss / (iteration + 1), + 'lr': get_lr(optimizer)}) + # 进度条前进1步 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) + # 将网络设置为验证模式(关闭批量归一化、dropout等训练专属层,使用滑动平均参数) net.eval() print('Start Validation') - with tqdm(total=epoch_size_val, desc=f'Epoch {epoch + 1}/{Epoch}',postfix=dict,mininterval=0.3) as pbar: + # 使用tqdm创建验证进度条,显示当前epoch信息和实时验证损失 + 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): + # 若迭代次数超过当前epoch的验证总步数,提前终止验证循环 if iteration >= epoch_size_val: break + # 从验证数据加载器中获取当前批次的图像和标签 images_val, targets_val = batch[0], batch[1] + # torch.no_grad() 上下文管理器:验证阶段无需计算梯度,大幅减少显存占用 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] + images_val = torch.from_numpy(images_val).type(torch.FloatTensor).cuda() + targets_val = [torch.from_numpy(ann).type(torch.FloatTensor).cuda() for ann in targets_val] else: - images_val = torch.from_numpy(images_val).type(torch.FloatTensor) + images_val = torch.from_numpy(images_val).type(torch.FloatTensor) targets_val = [torch.from_numpy(ann).type(torch.FloatTensor) for ann in targets_val] + # 验证阶段无需更新参数,此处清零梯度仅为代码一致性(无实际作用) optimizer.zero_grad() + # 前向传播获取验证集预测输出 outputs = net(images_val) losses = [] num_pos_all = 0 + # 遍历3个尺度层,计算验证损失 for i in range(3): loss_item, num_pos = yolo_loss(outputs[i], targets_val) losses.append(loss_item) num_pos_all += num_pos + # 计算验证总损失(归一化) loss = sum(losses) / num_pos_all + # 累加当前批次验证损失到总验证损失 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 + # 更新验证进度条的实时显示信息:平均验证损失 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 + + # 若启用Tensorboard,按epoch记录验证损失(便于观察长期趋势) 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)) + writer.add_scalar('Val_loss', val_loss / (epoch_size_val + 1), epoch) + # 将当前epoch的训练损失和验证损失保存到损失历史记录中(用于绘制损失曲线) + 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))) + # 打印当前epoch的训练信息 + 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))) + # 保存当前epoch的网络权重 + 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 - #-------------------------------# + # -------------------------------# + Tensorboard = False # 关闭Tensorboard可视化(如需启用,改为True) + # -------------------------------# # 是否使用Cuda # 没有GPU可以设置成False - #-------------------------------# - Cuda = True - #------------------------------------------------------# + # -------------------------------# + Cuda = True # 启用GPU加速训练(无GPU需改为False) + # ------------------------------------------------------# # 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 - #-------------------------------# + # ------------------------------------------------------# + normalize = False # 不使用批次大小归一化损失,使用正样本数量归一化 + # -------------------------------# # 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 = (320,320) - #----------------------------------------------------# + # -------------------------------# + input_shape = (320, 320) # 网络输入图像尺寸为320x320(宽x高,需为32的整数倍) + # ----------------------------------------------------# # 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' - #------------------------------------------------------# + # ----------------------------------------------------# + anchors_path = 'model_data/yolo_anchors.txt' # 先验框配置文件路径 + classes_path = 'model_data/voc_classes.txt' # 类别名称配置文件路径 + # ------------------------------------------------------# # Yolov4的tricks应用 - # mosaic 马赛克数据增强 True or False + # 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 + # ------------------------------------------------------# + mosaic = False # 关闭马赛克数据增强(训练不稳定,如需提升泛化能力可改为True) + Cosine_lr = False # 关闭余弦退火学习率调度,使用步长衰减 + smoooth_label = 0 # 关闭标签平滑(值大于0时启用,如0.01) - #----------------------------------------------------# + # ----------------------------------------------------# # 获取classes和anchor - #----------------------------------------------------# - class_names = get_classes(classes_path) - anchors = get_anchors(anchors_path) - num_classes = len(class_names) + # ----------------------------------------------------# + class_names = get_classes(classes_path) # 加载类别名称列表 + anchors = get_anchors(anchors_path) # 加载先验框数组 + num_classes = len(class_names) # 获取类别数量 - #------------------------------------------------------# + # ------------------------------------------------------# # 创建yolo模型 # 训练前一定要修改classes_path和对应的txt文件 - #------------------------------------------------------# + # ------------------------------------------------------# + # 初始化YOLOv4网络结构:参数1为每个尺度层的先验框数量,参数2为类别数量 model = YoloBody(len(anchors[0]), num_classes) + # 对网络权重进行初始化(使用自定义的权重初始化策略) weights_init(model) - #------------------------------------------------------# + # ------------------------------------------------------# # 权值文件请看README,百度网盘下载 - #------------------------------------------------------# - model_path = "model_data/yolo4_weights.pth" + # ------------------------------------------------------# + model_path = "model_data/yolo4_weights.pth" # 预训练权重文件路径 print('Loading weights into state dict...') + # 自动检测设备(有GPU用GPU,无GPU用CPU) device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') + # 获取当前模型的权重字典 model_dict = model.state_dict() + # 加载预训练权重,并映射到当前设备 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)} + # 筛选预训练权重中与当前模型形状匹配的层(避免类别数量不一致导致的报错) + pretrained_dict = {k: v for k, v in pretrained_dict.items() if np.shape(model_dict[k]) == np.shape(v)} + # 用预训练权重更新当前模型的权重字典 model_dict.update(pretrained_dict) + # 将更新后的权重字典加载到模型中 model.load_state_dict(model_dict) print('Finished!') # torch.save(model.state_dict(),"logs/test.pth",_use_new_zipfile_serialization=False ) + # 将模型设置为训练模式(启用训练专属层) 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) + # net = torch.nn.DataParallel(model) # 多GPU训练(注释掉,使用单GPU训练) + cudnn.benchmark = True # 启用cudnn基准测试,加速GPU训练 + net = net.cuda() # 将模型移至GPU + + # 初始化YOLOv4损失函数:参数依次为先验框、类别数、输入尺寸、标签平滑值、是否使用GPU、是否归一化 + yolo_loss = YOLOLoss(np.reshape(anchors, [-1, 2]), num_classes, (input_shape[1], input_shape[0]), smoooth_label, + Cuda, normalize) + # 初始化损失历史记录器:保存训练/验证损失到logs目录,并绘制损失曲线 loss_history = LossHistory("logs/") - #----------------------------------------------------# + # ----------------------------------------------------# # 获得图片路径和标签 - #----------------------------------------------------# - annotation_path = '2007_train.txt' - #----------------------------------------------------------------------# + # ----------------------------------------------------# + annotation_path = '2007_train.txt' # 训练数据集标注文件路径(每行对应一张图片的路径和标签) + # ----------------------------------------------------------------------# # 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,encoding="utf-8") as f: + # ----------------------------------------------------------------------# + val_split = 0.1 # 验证集占总数据集的比例(10%) + # 读取标注文件中的所有行(每张图片的标注信息) + with open(annotation_path, encoding="utf-8") as f: lines = f.readlines() + # 设置随机种子,保证数据集划分的可重复性 np.random.seed(10101) + # 打乱标注数据的顺序 np.random.shuffle(lines) + # 重置随机种子,恢复默认随机状态 np.random.seed(None) - num_val = int(len(lines)*val_split) + # 计算验证集的样本数量 + num_val = int(len(lines) * val_split) + # 计算训练集的样本数量 num_train = len(lines) - num_val + # 若启用Tensorboard,初始化SummaryWriter if Tensorboard: from tensorboardX import SummaryWriter - writer = SummaryWriter(log_dir='logs',flush_secs=60) + + # 创建日志写入器:日志保存到logs目录,每60秒刷新一次 + 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() + # 创建随机张量作为网络输入,用于绘制网络计算图 + 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) + graph_inputs = torch.randn(1, 3, input_shape[0], input_shape[1]).type(torch.FloatTensor) + # 将网络计算图写入Tensorboard writer.add_graph(model, graph_inputs) - train_tensorboard_step = 1 - val_tensorboard_step = 1 + # 初始化训练和验证的步数计数器 + 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 - #------------------------------------------------------# + # ------------------------------------------------------# + # 第一阶段:冻结主干网络训练(仅训练头部网络,加快训练速度,防止预训练权重被破坏) if True: - lr = 1e-3 - Batch_size = 4 - Init_Epoch = 0 - Freeze_Epoch = 50 - - #----------------------------------------------------------------------------# + lr = 1e-3 # 冻结阶段学习率(1e-3) + Batch_size = 4 # 冻结阶段批次大小(显存不足可减小) + Init_Epoch = 0 # 起始训练世代 + Freeze_Epoch = 50 # 冻结训练的结束世代(共50个epoch) + + # ----------------------------------------------------------------------------# # 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) + # ----------------------------------------------------------------------------# + # 初始化Adam优化器(仅优化当前可训练的网络参数) + optimizer = optim.Adam(net.parameters(), lr) + # 根据配置选择学习率调度器 if Cosine_lr: + # 余弦退火调度器:学习率随epoch呈余弦曲线下降 lr_scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=5, eta_min=1e-5) else: + # 步长衰减调度器:每经过1个epoch,学习率乘以0.92 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 - + # 创建训练数据集对象:传入训练集标注、输入尺寸、是否使用mosaic增强、是否为训练模式 + train_dataset = YoloDataset(lines[:num_train], (input_shape[0], input_shape[1]), mosaic=mosaic, is_train=True) + # 创建验证数据集对象:验证集不使用mosaic增强 + val_dataset = YoloDataset(lines[num_train:], (input_shape[0], input_shape[1]), mosaic=False, is_train=False) + # 创建训练数据加载器:打乱数据、指定批次大小、开启多线程加载、丢弃最后不完整批次 + gen = DataLoader(train_dataset, shuffle=True, batch_size=Batch_size, num_workers=4, pin_memory=True, + drop_last=True, collate_fn=yolo_dataset_collate) + # 创建验证数据加载器:配置与训练集一致 + 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) + + # 计算每个epoch的训练步数(训练样本数 // 批次大小) + epoch_size = num_train // Batch_size + # 计算每个epoch的验证步数(验证样本数 // 批次大小) + epoch_size_val = num_val // Batch_size + + # 检查数据集大小:若训练/验证步数为0,说明数据集过小,无法训练 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 - #------------------------------------# + # ------------------------------------# + # 冻结主干网络的所有参数(设置requires_grad=False,禁止梯度更新) 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) + # 遍历冻结训练的每个epoch,执行训练和验证 + 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) + # 更新学习率(根据调度器调整) lr_scheduler.step() + # 第二阶段:解冻主干网络训练(训练整个网络,微调权重,提升模型精度) if True: - lr = 1e-4 - Batch_size = 2 - Freeze_Epoch = 50 - Unfreeze_Epoch = 100 + lr = 1e-4 # 解冻阶段学习率(比冻结阶段小10倍,避免权重震荡) + Batch_size = 2 # 解冻阶段批次大小(显存占用更高,需减小) + Freeze_Epoch = 50 # 解冻训练的起始世代 + Unfreeze_Epoch = 100 # 解冻训练的结束世代(总训练100个epoch) - #----------------------------------------------------------------------------# + # ----------------------------------------------------------------------------# # 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) + # ----------------------------------------------------------------------------# + # 初始化Adam优化器(优化所有网络参数) + optimizer = optim.Adam(net.parameters(), lr) + # 根据配置选择学习率调度器 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) + # 重新创建训练/验证数据集(配置与冻结阶段一致) + 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) + # 重新创建训练/验证数据加载器(批次大小减小) + 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) + + # 重新计算每个epoch的训练/验证步数 + epoch_size = num_train // Batch_size + epoch_size_val = num_val // Batch_size - epoch_size = num_train // Batch_size - epoch_size_val = num_val // Batch_size - + # 检查数据集大小 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 - #------------------------------------# + # ------------------------------------# + # 解冻主干网络的所有参数(设置requires_grad=True,允许梯度更新) 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() + # 遍历解冻训练的每个epoch,执行训练和验证 + 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) + # 更新学习率 + lr_scheduler.step() \ No newline at end of file