import cv2 import tqdm import wandb import torch import numpy as np import albumentations from block.val_get import val_get from block.model_ema import model_ema from block.lr_get import adam, lr_adjust ''' 处理训练过程 ''' def train_get(args, data_dict, model_dict, loss): # 加载模型 model = model_dict['model'].to(args.device, non_blocking=args.latch) print(model) # 学习率 optimizer = adam(args.regularization, args.r_value, model.parameters(), lr=args.lr_start, betas=(0.937, 0.999)) optimizer.load_state_dict(model_dict['optimizer_state_dict']) if model_dict['optimizer_state_dict'] else None step_epoch = len(data_dict['train']) // args.batch // args.device_number * args.device_number # 每轮的步数 print(len(data_dict['train']) // args.batch) print(step_epoch) optimizer_adjust = lr_adjust(args, step_epoch, model_dict['epoch_finished']) # 学习率调整函数 optimizer = optimizer_adjust(optimizer) # 学习率初始化 # 使用平均指数移动(EMA)调整参数(不能将ema放到args中，否则会导致模型保存出错) ema = model_ema(model) if args.ema else None if args.ema: ema.updates = model_dict['ema_updates'] # 数据集 train_dataset = torch_dataset(args, 'train', data_dict['train'], data_dict['class']) train_sampler = torch.utils.data.distributed.DistributedSampler(train_dataset) if args.distributed else None train_shuffle = False if args.distributed else True # 分布式设置sampler后shuffle要为False train_dataloader = torch.utils.data.DataLoader(train_dataset, batch_size=args.batch, shuffle=train_shuffle, drop_last=True, pin_memory=args.latch, num_workers=args.num_worker, sampler=train_sampler) # 验证集不对图像进行处理 val_dataset = torch_dataset(args, 'test', data_dict['test'], data_dict['class']) val_sampler = None # 分布式时数据合在主GPU上进行验证 val_batch = args.batch // args.device_number # 分布式验证时batch要减少为一个GPU的量 val_dataloader = torch.utils.data.DataLoader(val_dataset, batch_size=val_batch, shuffle=False, drop_last=False, pin_memory=args.latch, num_workers=args.num_worker, sampler=val_sampler) # 分布式初始化 model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank], output_device=args.local_rank) if args.distributed else model # wandb if args.wandb and args.local_rank == 0: wandb_image_list = [] # 记录所有的wandb_image最后一起添加(最多添加args.wandb_image_num张) epoch_base = model_dict['epoch_finished'] + 1 # 新的一轮要+1 for epoch in range(epoch_base, args.epoch + 1): # 训练 print(f'\n-----------------------第{epoch}轮-----------------------') if args.local_rank == 0 else None model.train() train_loss = 0 # 记录损失 if args.local_rank == 0: # tqdm tqdm_show = tqdm.tqdm(total=step_epoch) for index, (image_batch, true_batch) in enumerate(train_dataloader): if args.wandb and args.local_rank == 0 and len(wandb_image_list) < args.wandb_image_num: wandb_image_batch = (image_batch * 255).cpu().numpy().astype(np.uint8).transpose(0, 2, 3, 1) image_batch = image_batch.to(args.device, non_blocking=args.latch) true_batch = true_batch.to(args.device, non_blocking=args.latch) if args.amp: with torch.cuda.amp.autocast(): pred_batch = model(image_batch) loss_batch = loss(pred_batch, true_batch) args.amp.scale(loss_batch).backward() args.amp.step(optimizer) args.amp.update() optimizer.zero_grad() else: pred_batch = model(image_batch) loss_batch = loss(pred_batch, true_batch) loss_batch.backward() optimizer.step() optimizer.zero_grad() # 调整参数，ema.updates会自动+1 ema.update(model) if args.ema else None # 记录损失 train_loss += loss_batch.item() # 调整学习率 optimizer = optimizer_adjust(optimizer) # tqdm if args.local_rank == 0: tqdm_show.set_postfix({'train_loss': loss_batch.item(), 'lr': optimizer.param_groups[0]['lr']}) # 添加显示 tqdm_show.update(args.device_number) # 更新进度条 # wandb if args.wandb and args.local_rank == 0 and epoch == 0 and len(wandb_image_list) < args.wandb_image_num: cls = true_batch.cpu().numpy().tolist() for i in range(len(wandb_image_batch)): # 遍历每一张图片 image = wandb_image_batch[i] text = ['{:.0f}'.format(_) for _ in cls[i]] text = text[0] if len(text) == 1 else '--'.join(text) image = np.ascontiguousarray(image) # 将数组的内存变为连续存储(cv2画图的要求) cv2.putText(image, text, (30, 30), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 0), 2) wandb_image = wandb.Image(image) wandb_image_list.append(wandb_image) if len(wandb_image_list) == args.wandb_image_num: break # tqdm if args.local_rank == 0: tqdm_show.close() # 计算平均损失 train_loss /= index + 1 if args.local_rank == 0: print(f'\n| 训练 | train_loss:{train_loss:.4f} | lr:{optimizer.param_groups[0]["lr"]:.6f} |\n') # 清理显存空间 del image_batch, true_batch, pred_batch, loss_batch torch.cuda.empty_cache() # 验证 if args.local_rank == 0: # 分布式时只验证一次 # val_loss, accuracy, precision, recall, m_ap = val_get(args, val_dataloader, model, loss, ema, # len(data_dict['test'])) val_loss, accuracy = val_get(args, val_dataloader, model, loss, ema, len(data_dict['test'])) # 保存 if args.local_rank == 0: # 分布式时只保存一次 model_dict['model'] = model.module if args.distributed else model model_dict['epoch_finished'] = epoch model_dict['optimizer_state_dict'] = optimizer.state_dict() model_dict['ema_updates'] = ema.updates if args.ema else model_dict['ema_updates'] model_dict['class'] = data_dict['class'] model_dict['train_loss'] = train_loss model_dict['val_loss'] = val_loss model_dict['val_accuracy'] = accuracy # model_dict['val_precision'] = precision # model_dict['val_recall'] = recall # model_dict['val_m_ap'] = m_ap torch.save(model_dict, args.save_path_last if not args.prune else 'prune_last.pt') # 保存最后一次训练的模型 # if m_ap > 0.5 and m_ap > model_dict['standard']: # model_dict['standard'] = m_ap # save_path = args.save_path if not args.prune else args.prune_save # torch.save(model_dict, save_path) # 保存最佳模型 # print(f'| 保存最佳模型:{save_path} | val_m_ap:{m_ap:.4f} |') if accuracy > 0.5 and accuracy > model_dict['standard']: model_dict['standard'] = accuracy save_path = args.save_path if not args.prune else args.prune_save torch.save(model_dict, save_path) # 保存最佳模型 print(f'| 保存最佳模型:{save_path} | accuracy:{(100 * accuracy):.4f}% |') # wandb if args.wandb: wandb_log = {} if epoch == 0: wandb_log.update({f'image/train_image': wandb_image_list}) # wandb_log.update({'metric/train_loss': train_loss, # 'metric/val_loss': val_loss, # 'metric/val_m_ap': m_ap, # 'metric/val_accuracy': accuracy, # 'metric/val_precision': precision, # 'metric/val_recall': recall}) wandb_log.update({'metric/train_loss': train_loss, 'metric/val_loss': val_loss, 'metric/val_accuracy': accuracy}) args.wandb_run.log(wandb_log) torch.distributed.barrier() if args.distributed else None # 分布式时每轮训练后让所有GPU进行同步，快的GPU会在此等待 class torch_dataset(torch.utils.data.Dataset): def __init__(self, args, tag, data, class_name): self.tag = tag self.data = data self.class_name = class_name self.noise_probability = args.noise self.noise = albumentations.Compose([ albumentations.GaussianBlur(blur_limit=(5, 5), p=0.2), albumentations.GaussNoise(var_limit=(10.0, 30.0), p=0.2),], ) self.transform = albumentations.Compose([ albumentations.LongestMaxSize(args.input_size), albumentations.PadIfNeeded(min_height=args.input_size, min_width=args.input_size, border_mode=cv2.BORDER_CONSTANT, value=(128, 128, 128))]) self.rgb_mean = (0.406, 0.456, 0.485) self.rgb_std = (0.225, 0.224, 0.229) def __len__(self): return len(self.data) def __getitem__(self, index): image = cv2.imread(self.data[index][0]) # 读取图片 if self.tag == 'train' and torch.rand(1) < self.noise_probability: # 使用数据加噪 image = self.noise(image=image)['image'] image = self.transform(image=image)['image'] # 缩放和填充图片 image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB) # 转为RGB通道 image = self._image_deal(image) # 归一化、转换为tensor、调维度 label = torch.tensor(self.data[index][1], dtype=torch.float32) # 转换为tensor return image, label def _image_deal(self, image): # 归一化、转换为tensor、调维度 image = torch.tensor(image / 255, dtype=torch.float32).permute(2, 0, 1) return image