def voc_test(model, device, input_size): evaluator = VOCAPIEvaluator(data_root=VOC_ROOT, img_size=input_size, device=device, transform=BaseTransform(input_size), labelmap=VOC_CLASSES, display=True) # VOC evaluation evaluator.evaluate(model)
def train(): args = parse_args() path_to_save = os.path.join(args.save_folder, args.dataset, args.version) os.makedirs(path_to_save, exist_ok=True) # cuda if args.cuda: print('use cuda') cudnn.benchmark = True device = torch.device("cuda") else: device = torch.device("cpu") # mosaic augmentation if args.mosaic: print('use Mosaic Augmentation ...') # multi-scale if args.multi_scale: print('use the multi-scale trick ...') train_size = [640, 640] val_size = [512, 512] else: train_size = [512, 512] val_size = [512, 512] cfg = train_cfg # dataset and evaluator print("Setting Arguments.. : ", args) print("----------------------------------------------------------") print('Loading the dataset...') if args.dataset == 'voc': data_dir = VOC_ROOT num_classes = 20 dataset = VOCDetection(root=data_dir, img_size=train_size[0], transform=SSDAugmentation(train_size), mosaic=args.mosaic ) evaluator = VOCAPIEvaluator(data_root=data_dir, img_size=val_size, device=device, transform=BaseTransform(val_size), labelmap=VOC_CLASSES ) elif args.dataset == 'coco': data_dir = coco_root num_classes = 80 dataset = COCODataset( data_dir=data_dir, img_size=train_size[0], transform=SSDAugmentation(train_size), debug=args.debug, mosaic=args.mosaic ) evaluator = COCOAPIEvaluator( data_dir=data_dir, img_size=val_size, device=device, transform=BaseTransform(val_size) ) else: print('unknow dataset !! Only support voc and coco !!') exit(0) print('Training model on:', dataset.name) print('The dataset size:', len(dataset)) print("----------------------------------------------------------") # dataloader dataloader = torch.utils.data.DataLoader( dataset, batch_size=args.batch_size, shuffle=True, collate_fn=detection_collate, num_workers=args.num_workers, pin_memory=True ) # build model if args.version == 'centernet': from models.centernet import CenterNet net = CenterNet(device, input_size=train_size, num_classes=num_classes, trainable=True) print('Let us train centernet on the %s dataset ......' % (args.dataset)) else: print('Unknown version !!!') exit() model = net model.to(device).train() # use tfboard if args.tfboard: print('use tensorboard') from torch.utils.tensorboard import SummaryWriter c_time = time.strftime('%Y-%m-%d %H:%M:%S',time.localtime(time.time())) log_path = os.path.join('log/coco/', args.version, c_time) os.makedirs(log_path, exist_ok=True) writer = SummaryWriter(log_path) # keep training if args.resume is not None: print('keep training model: %s' % (args.resume)) model.load_state_dict(torch.load(args.resume, map_location=device)) # optimizer setup base_lr = args.lr tmp_lr = base_lr optimizer = optim.SGD(model.parameters(), lr=args.lr, momentum=args.momentum, weight_decay=args.weight_decay ) max_epoch = cfg['max_epoch'] epoch_size = len(dataset) // args.batch_size # start training loop t0 = time.time() for epoch in range(args.start_epoch, max_epoch): # use cos lr if args.cos and epoch > 20 and epoch <= max_epoch - 20: # use cos lr tmp_lr = 0.00001 + 0.5*(base_lr-0.00001)*(1+math.cos(math.pi*(epoch-20)*1./ (max_epoch-20))) set_lr(optimizer, tmp_lr) elif args.cos and epoch > max_epoch - 20: tmp_lr = 0.00001 set_lr(optimizer, tmp_lr) # use step lr else: if epoch in cfg['lr_epoch']: tmp_lr = tmp_lr * 0.1 set_lr(optimizer, tmp_lr) for iter_i, (images, targets) in enumerate(dataloader): # WarmUp strategy for learning rate if not args.no_warm_up: if epoch < args.wp_epoch: tmp_lr = base_lr * pow((iter_i+epoch*epoch_size)*1. / (args.wp_epoch*epoch_size), 4) # tmp_lr = 1e-6 + (base_lr-1e-6) * (iter_i+epoch*epoch_size) / (epoch_size * (args.wp_epoch)) set_lr(optimizer, tmp_lr) elif epoch == args.wp_epoch and iter_i == 0: tmp_lr = base_lr set_lr(optimizer, tmp_lr) # to device images = images.to(device) # multi-scale trick if iter_i % 10 == 0 and iter_i > 0 and args.multi_scale: # randomly choose a new size size = random.randint(10, 19) * 32 train_size = [size, size] model.set_grid(train_size) if args.multi_scale: # interpolate images = torch.nn.functional.interpolate(images, size=train_size, mode='bilinear', align_corners=False) # make train label targets = [label.tolist() for label in targets] targets = tools.gt_creator(train_size, net.stride, args.num_classes, targets) targets = torch.tensor(targets).float().to(device) # forward and loss cls_loss, txty_loss, twth_loss, total_loss = model(images, target=targets) # backprop total_loss.backward() optimizer.step() optimizer.zero_grad() if iter_i % 10 == 0: if args.tfboard: # viz loss writer.add_scalar('class loss', cls_loss.item(), iter_i + epoch * epoch_size) writer.add_scalar('txty loss', txty_loss.item(), iter_i + epoch * epoch_size) writer.add_scalar('twth loss', twth_loss.item(), iter_i + epoch * epoch_size) writer.add_scalar('total loss', total_loss.item(), iter_i + epoch * epoch_size) t1 = time.time() print('[Epoch %d/%d][Iter %d/%d][lr %.6f]' '[Loss: cls %.2f || txty %.2f || twth %.2f ||total %.2f || size %d || time: %.2f]' % (epoch+1, max_epoch, iter_i, epoch_size, tmp_lr, cls_loss.item(), txty_loss.item(), twth_loss.item(), total_loss.item(), train_size[0], t1-t0), flush=True) t0 = time.time() # evaluation if (epoch) % args.eval_epoch == 0: model.trainable = False model.set_grid(val_size) model.eval() # evaluate evaluator.evaluate(model) # convert to training mode. model.trainable = True model.set_grid(train_size) model.train() # save model if (epoch + 1) % 10 == 0: print('Saving state, epoch:', epoch + 1) torch.save(model.state_dict(), os.path.join(path_to_save, args.version + '_' + repr(epoch + 1) + '.pth') )
def train(): args = parse_args() path_to_save = os.path.join(args.save_folder, args.dataset, args.version) # path_to_save = os.path.join('weights', 'voc', 'yolov') os.makedirs(path_to_save, exist_ok=True) # use hi-res backbone if args.high_resolution: print('use hi-res backbone') hr = True else: hr = False # cuda if args.cuda: print('use cuda') cudnn.benchmark = True device = torch.device("cuda", 2) else: device = torch.device("cpu") # multi-scale if args.multi_scale: print('use the multi-scale trick ...') train_size = [640, 640] val_size = [416, 416] else: train_size = [416, 416] val_size = [416, 416] cfg = train_cfg # dataset and evaluator print("Setting Arguments.. : ", args) print("----------------------------------------------------------") print('Loading the dataset...') if args.dataset == 'voc': VOC_ROOT = '/home/liyi219/python/new-YOLOv1_PyTorch-master/VOCdevkit' data_dir = VOC_ROOT num_classes = 20 dataset = VOCDetection(root=data_dir, img_size=train_size[0], transform=SSDAugmentation(train_size)) evaluator = VOCAPIEvaluator(data_root=data_dir, img_size=val_size, device=device, transform=BaseTransform(val_size), labelmap=VOC_CLASSES) # elif args.dataset == 'coco': # data_dir = coco_root # num_classes = 80 # dataset = COCODataset( # data_dir=data_dir, # img_size=train_size[0], # transform=SSDAugmentation(train_size), # debug=args.debug # ) # # evaluator = COCOAPIEvaluator( # data_dir=data_dir, # img_size=val_size, # device=device, # transform=BaseTransform(val_size) # ) else: print('unknow dataset !! Only support voc and coco !!') exit(0) print('Training model on:', dataset.name) print('The dataset size:', len(dataset)) print("----------------------------------------------------------") # dataloader dataloader = torch.utils.data.DataLoader(dataset, batch_size=args.batch_size, shuffle=True, collate_fn=detection_collate, num_workers=args.num_workers, pin_memory=True) """ dataloader = torch.utils.data.DataLoader( dataset, batch_size=32, shuffle=True, collate_fn=detection_collate, num_workers=8, pin_memory=True ) """ # build model if args.version == 'yolo': from models.yolo import myYOLO yolo_net = myYOLO(device, input_size=train_size, num_classes=num_classes, trainable=True) print('Let us train yolo on the %s dataset ......' % (args.dataset)) else: print('We only support YOLO !!!') exit() model = yolo_net model.to(device).train() # use tfboard if args.tfboard: print('use tensorboard') from torch.utils.tensorboard import SummaryWriter c_time = time.strftime('%Y-%m-%d %H:%M:%S', time.localtime(time.time())) log_path = os.path.join('log/coco/', args.version, c_time) # log_path = os.path.join('log/coco/', 'yolo', c_time) os.makedirs(log_path, exist_ok=True) writer = SummaryWriter(log_path) # keep training if args.resume is not None: print('keep training model: %s' % (args.resume)) model.load_state_dict(torch.load(args.resume, map_location=device)) # model.load_state_dict(torch.load('keep training', map_location=device)) # optimizer setup base_lr = args.lr tmp_lr = base_lr optimizer = optim.SGD(model.parameters(), lr=args.lr, momentum=args.momentum, weight_decay=args.weight_decay) """ optimizer = optim.SGD(model.parameters(), lr=1e-3, momentum=0.9, weight_decay=5e-4 ) """ max_epoch = cfg['max_epoch'] epoch_size = len(dataset) // args.batch_size # epoch_size = len(dataset) // 32 # start training loop t0 = time.time() for epoch in range(args.start_epoch, max_epoch): # use cos lr if args.cos and epoch > 20 and epoch <= max_epoch - 20: # use cos lr tmp_lr = 0.00001 + 0.5 * (base_lr - 0.00001) * ( 1 + math.cos(math.pi * (epoch - 20) * 1. / (max_epoch - 20))) set_lr(optimizer, tmp_lr) elif args.cos and epoch > max_epoch - 20: tmp_lr = 0.00001 set_lr(optimizer, tmp_lr) # use step lr else: if epoch in cfg['lr_epoch']: tmp_lr = tmp_lr * 0.1 set_lr(optimizer, tmp_lr) for iter_i, (images, targets) in enumerate(dataloader): # WarmUp strategy for learning rate if not args.no_warm_up: if epoch < args.wp_epoch: tmp_lr = base_lr * pow((iter_i + epoch * epoch_size) * 1. / (args.wp_epoch * epoch_size), 4) # tmp_lr = base_lr * pow((iter_i + epoch * epoch_size) * 1. / (2 * epoch_size), 4) # tmp_lr = 1e-6 + (base_lr-1e-6) * (iter_i+epoch*epoch_size) / (epoch_size * (args.wp_epoch)) set_lr(optimizer, tmp_lr) elif epoch == args.wp_epoch and iter_i == 0: tmp_lr = base_lr set_lr(optimizer, tmp_lr) # to device images = images.to(device) # multi-scale trick if iter_i % 10 == 0 and iter_i > 0 and args.multi_scale: # randomly choose a new size size = random.randint(10, 19) * 32 train_size = [size, size] model.set_grid(train_size) if args.multi_scale: # interpolate images = torch.nn.functional.interpolate(images, size=train_size, mode='bilinear', align_corners=False) # make train label # 原本的targets里面只有bbox坐标和类别参数,用gt_creator函数加上置信度和权重参数 targets = [label.tolist() for label in targets] targets = tools.gt_creator(input_size=train_size, stride=yolo_net.stride, label_lists=targets) targets = torch.tensor(targets).float().to(device) # forward and loss conf_loss, cls_loss, txtytwth_loss, total_loss = model( images, target=targets) # backprop total_loss.backward() optimizer.step() optimizer.zero_grad() # display if iter_i % 10 == 0: if args.tfboard: # viz loss writer.add_scalar('object loss', conf_loss.item(), iter_i + epoch * epoch_size) writer.add_scalar('class loss', cls_loss.item(), iter_i + epoch * epoch_size) writer.add_scalar('local loss', txtytwth_loss.item(), iter_i + epoch * epoch_size) t1 = time.time() print( '[Epoch %d/%d][Iter %d/%d][lr %.6f]' '[Loss: obj %.2f || cls %.2f || bbox %.2f || total %.2f || size %d || time: %.2f]' % (epoch + 1, max_epoch, iter_i, epoch_size, tmp_lr, conf_loss.item(), cls_loss.item(), txtytwth_loss.item(), total_loss.item(), train_size[0], t1 - t0), flush=True) t0 = time.time() # evaluation if (epoch + 1) % args.eval_epoch == 0: model.trainable = False model.set_grid(val_size) model.eval() # evaluate evaluator.evaluate(model) # convert to training mode. model.trainable = True model.set_grid(train_size) model.train() # save model if (epoch + 1) % 10 == 0: print('Saving state, epoch:', epoch + 1) torch.save( model.state_dict(), os.path.join(path_to_save, args.version + '_' + repr(epoch + 1) + '.pth'))