def main(opts): # ===== Setup distributed ===== distributed.init_process_group(backend='nccl', init_method='env://') if opts.device is not None: device_id = opts.device else: device_id = opts.local_rank device = torch.device(device_id) rank, world_size = distributed.get_rank(), distributed.get_world_size() if opts.device is not None: torch.cuda.set_device(opts.device) else: torch.cuda.set_device(device_id) # ===== Initialize logging ===== logdir_full = f"{opts.logdir}/{opts.dataset}/{opts.name}/" if rank == 0: logger = Logger(logdir_full, rank=rank, debug=opts.debug, summary=opts.visualize) else: logger = Logger(logdir_full, rank=rank, debug=opts.debug, summary=False) logger.print(f"Device: {device}") checkpoint_path = f"checkpoints/{opts.dataset}/{opts.name}.pth" os.makedirs(f"checkpoints/{opts.dataset}", exist_ok=True) # ===== Setup random seed to reproducibility ===== torch.manual_seed(opts.random_seed) torch.cuda.manual_seed(opts.random_seed) np.random.seed(opts.random_seed) random.seed(opts.random_seed) # ===== Set up dataset ===== train_dst, val_dst = get_dataset(opts, train=True) train_loader = data.DataLoader(train_dst, batch_size=opts.batch_size, sampler=DistributedSampler( train_dst, num_replicas=world_size, rank=rank), num_workers=opts.num_workers, drop_last=True, pin_memory=True) val_loader = data.DataLoader(val_dst, batch_size=opts.batch_size, sampler=DistributedSampler( val_dst, num_replicas=world_size, rank=rank), num_workers=opts.num_workers) logger.info(f"Dataset: {opts.dataset}, Train set: {len(train_dst)}, " f"Val set: {len(val_dst)}, n_classes {opts.num_classes}") logger.info(f"Total batch size is {opts.batch_size * world_size}") # This is necessary for computing the scheduler decay opts.max_iter = opts.max_iter = opts.epochs * len(train_loader) # ===== Set up model and ckpt ===== model = Trainer(device, logger, opts) model.distribute() cur_epoch = 0 if opts.continue_ckpt: opts.ckpt = checkpoint_path if opts.ckpt is not None: assert os.path.isfile( opts.ckpt), "Error, ckpt not found. Check the correct directory" checkpoint = torch.load(opts.ckpt, map_location="cpu") cur_epoch = checkpoint["epoch"] + 1 model.load_state_dict(checkpoint["model_state"]) logger.info("[!] Model restored from %s" % opts.ckpt) del checkpoint else: logger.info("[!] Train from scratch") # ===== Train procedure ===== # print opts before starting training to log all parameters logger.add_table("Opts", vars(opts)) # uncomment if you want qualitative on val # if rank == 0 and opts.sample_num > 0: # sample_ids = np.random.choice(len(val_loader), opts.sample_num, replace=False) # sample idxs for visualization # logger.info(f"The samples id are {sample_ids}") # else: # sample_ids = None label2color = utils.Label2Color(cmap=utils.color_map( opts.dataset)) # convert labels to images denorm = utils.Denormalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225 ]) # de-normalization for original images train_metrics = StreamSegMetrics(opts.num_classes) val_metrics = StreamSegMetrics(opts.num_classes) results = {} # check if random is equal here. logger.print(torch.randint(0, 100, (1, 1))) while cur_epoch < opts.epochs and not opts.test: # ===== Train ===== start = time.time() epoch_loss = model.train(cur_epoch=cur_epoch, train_loader=train_loader, metrics=train_metrics, print_int=opts.print_interval) train_score = train_metrics.get_results() end = time.time() len_ep = int(end - start) logger.info( f"End of Epoch {cur_epoch}/{opts.epochs}, Average Loss={epoch_loss[0] + epoch_loss[1]:.4f}, " f"Class Loss={epoch_loss[0]:.4f}, Reg Loss={epoch_loss[1]}\n" f"Train_Acc={train_score['Overall Acc']:.4f}, Train_Iou={train_score['Mean IoU']:.4f} " f"\n -- time: {len_ep // 60}:{len_ep % 60} -- ") logger.info( f"I will finish in {len_ep * (opts.epochs - cur_epoch) // 60} minutes" ) logger.add_scalar("E-Loss", epoch_loss[0] + epoch_loss[1], cur_epoch) # logger.add_scalar("E-Loss-reg", epoch_loss[1], cur_epoch) # logger.add_scalar("E-Loss-cls", epoch_loss[0], cur_epoch) # ===== Validation ===== if (cur_epoch + 1) % opts.val_interval == 0: logger.info("validate on val set...") val_loss, _ = model.validate(loader=val_loader, metrics=val_metrics, ret_samples_ids=None) val_score = val_metrics.get_results() logger.print("Done validation") logger.info( f"End of Validation {cur_epoch}/{opts.epochs}, Validation Loss={val_loss}" ) log_val(logger, val_metrics, val_score, val_loss, cur_epoch) # keep the metric to print them at the end of training results["V-IoU"] = val_score['Class IoU'] results["V-Acc"] = val_score['Class Acc'] # ===== Save Model ===== if rank == 0: if not opts.debug: save_ckpt(checkpoint_path, model, cur_epoch) logger.info("[!] Checkpoint saved.") cur_epoch += 1 torch.distributed.barrier() # ==== TESTING ===== logger.info("*** Test the model on all seen classes...") # make data loader test_dst = get_dataset(opts, train=False) test_loader = data.DataLoader(test_dst, batch_size=opts.batch_size_test, sampler=DistributedSampler( test_dst, num_replicas=world_size, rank=rank), num_workers=opts.num_workers) if rank == 0 and opts.sample_num > 0: sample_ids = np.random.choice(len(test_loader), opts.sample_num, replace=False) # sample idxs for visual. logger.info(f"The samples id are {sample_ids}") else: sample_ids = None val_loss, ret_samples = model.validate(loader=test_loader, metrics=val_metrics, ret_samples_ids=sample_ids) val_score = val_metrics.get_results() conf_matrixes = val_metrics.get_conf_matrixes() logger.print("Done test on all") logger.info(f"*** End of Test on all, Total Loss={val_loss}") logger.info(val_metrics.to_str(val_score)) log_samples(logger, ret_samples, denorm, label2color, 0) logger.add_figure("Test_Confusion_Matrix_Recall", conf_matrixes['Confusion Matrix']) logger.add_figure("Test_Confusion_Matrix_Precision", conf_matrixes["Confusion Matrix Pred"]) results["T-IoU"] = val_score['Class IoU'] results["T-Acc"] = val_score['Class Acc'] results["T-Prec"] = val_score['Class Prec'] logger.add_results(results) logger.add_scalar("T_Overall_Acc", val_score['Overall Acc']) logger.add_scalar("T_MeanIoU", val_score['Mean IoU']) logger.add_scalar("T_MeanAcc", val_score['Mean Acc']) ret = val_score['Mean IoU'] logger.close() return ret
def main(opts): distributed.init_process_group(backend='nccl', init_method='env://') device_id, device = opts.local_rank, torch.device(opts.local_rank) rank, world_size = distributed.get_rank(), distributed.get_world_size() torch.cuda.set_device(device_id) # Initialize logging task_name = f"{opts.task}-{opts.dataset}" logdir_full = f"{opts.logdir}/{task_name}/{opts.name}/" if rank == 0: logger = Logger(logdir_full, rank=rank, debug=opts.debug, summary=opts.visualize, step=opts.step) else: logger = Logger(logdir_full, rank=rank, debug=opts.debug, summary=False) logger.print(f"Device: {device}") # Set up random seed torch.manual_seed(opts.random_seed) torch.cuda.manual_seed(opts.random_seed) np.random.seed(opts.random_seed) random.seed(opts.random_seed) # xxx Set up dataloader train_dst, val_dst, test_dst, n_classes = get_dataset(opts) # reset the seed, this revert changes in random seed random.seed(opts.random_seed) train_loader = data.DataLoader(train_dst, batch_size=opts.batch_size, sampler=DistributedSampler(train_dst, num_replicas=world_size, rank=rank), num_workers=opts.num_workers, drop_last=True) val_loader = data.DataLoader(val_dst, batch_size=opts.batch_size if opts.crop_val else 1, sampler=DistributedSampler(val_dst, num_replicas=world_size, rank=rank), num_workers=opts.num_workers) logger.info(f"Dataset: {opts.dataset}, Train set: {len(train_dst)}, Val set: {len(val_dst)}," f" Test set: {len(test_dst)}, n_classes {n_classes}") logger.info(f"Total batch size is {opts.batch_size * world_size}") # xxx Set up model logger.info(f"Backbone: {opts.backbone}") step_checkpoint = None model = make_model(opts, classes=tasks.get_per_task_classes(opts.dataset, opts.task, opts.step)) logger.info(f"[!] Model made with{'out' if opts.no_pretrained else ''} pre-trained") if opts.step == 0: # if step 0, we don't need to instance the model_old model_old = None else: # instance model_old model_old = make_model(opts, classes=tasks.get_per_task_classes(opts.dataset, opts.task, opts.step - 1)) if opts.fix_bn: model.fix_bn() logger.debug(model) # xxx Set up optimizer params = [] if not opts.freeze: params.append({"params": filter(lambda p: p.requires_grad, model.body.parameters()), 'weight_decay': opts.weight_decay}) params.append({"params": filter(lambda p: p.requires_grad, model.head.parameters()), 'weight_decay': opts.weight_decay}) params.append({"params": filter(lambda p: p.requires_grad, model.cls.parameters()), 'weight_decay': opts.weight_decay}) optimizer = torch.optim.SGD(params, lr=opts.lr, momentum=0.9, nesterov=True) if opts.lr_policy == 'poly': scheduler = utils.PolyLR(optimizer, max_iters=opts.epochs * len(train_loader), power=opts.lr_power) elif opts.lr_policy == 'step': scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=opts.lr_decay_step, gamma=opts.lr_decay_factor) else: raise NotImplementedError logger.debug("Optimizer:\n%s" % optimizer) if model_old is not None: [model, model_old], optimizer = amp.initialize([model.to(device), model_old.to(device)], optimizer, opt_level=opts.opt_level) model_old = DistributedDataParallel(model_old) else: model, optimizer = amp.initialize(model.to(device), optimizer, opt_level=opts.opt_level) # Put the model on GPU model = DistributedDataParallel(model, delay_allreduce=True) # xxx Load old model from old weights if step > 0! if opts.step > 0: # get model path if opts.step_ckpt is not None: path = opts.step_ckpt else: path = f"checkpoints/step/{task_name}_{opts.name}_{opts.step - 1}.pth" # generate model from path if os.path.exists(path): step_checkpoint = torch.load(path, map_location="cpu") model.load_state_dict(step_checkpoint['model_state'], strict=False) # False because of incr. classifiers if opts.init_balanced: # implement the balanced initialization (new cls has weight of background and bias = bias_bkg - log(N+1) model.module.init_new_classifier(device) # Load state dict from the model state dict, that contains the old model parameters model_old.load_state_dict(step_checkpoint['model_state'], strict=True) # Load also here old parameters logger.info(f"[!] Previous model loaded from {path}") # clean memory del step_checkpoint['model_state'] elif opts.debug: logger.info(f"[!] WARNING: Unable to find of step {opts.step - 1}! Do you really want to do from scratch?") else: raise FileNotFoundError(path) # put the old model into distributed memory and freeze it for par in model_old.parameters(): par.requires_grad = False model_old.eval() # xxx Set up Trainer trainer_state = None # if not first step, then instance trainer from step_checkpoint if opts.step > 0 and step_checkpoint is not None: if 'trainer_state' in step_checkpoint: trainer_state = step_checkpoint['trainer_state'] # instance trainer (model must have already the previous step weights) trainer = Trainer(model, model_old, device=device, opts=opts, trainer_state=trainer_state, classes=tasks.get_per_task_classes(opts.dataset, opts.task, opts.step)) # xxx Handle checkpoint for current model (model old will always be as previous step or None) best_score = 0.0 cur_epoch = 0 if opts.ckpt is not None and os.path.isfile(opts.ckpt): checkpoint = torch.load(opts.ckpt, map_location="cpu") model.load_state_dict(checkpoint["model_state"], strict=True) optimizer.load_state_dict(checkpoint["optimizer_state"]) scheduler.load_state_dict(checkpoint["scheduler_state"]) cur_epoch = checkpoint["epoch"] + 1 best_score = checkpoint['best_score'] logger.info("[!] Model restored from %s" % opts.ckpt) # if we want to resume training, resume trainer from checkpoint if 'trainer_state' in checkpoint: trainer.load_state_dict(checkpoint['trainer_state']) del checkpoint else: if opts.step == 0: logger.info("[!] Train from scratch") # xxx Train procedure # print opts before starting training to log all parameters logger.add_table("Opts", vars(opts)) if rank == 0 and opts.sample_num > 0: sample_ids = np.random.choice(len(val_loader), opts.sample_num, replace=False) # sample idxs for visualization logger.info(f"The samples id are {sample_ids}") else: sample_ids = None label2color = utils.Label2Color(cmap=utils.color_map(opts.dataset)) # convert labels to images denorm = utils.Denormalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) # de-normalization for original images TRAIN = not opts.test val_metrics = StreamSegMetrics(n_classes) results = {} # check if random is equal here. logger.print(torch.randint(0,100, (1,1))) # train/val here while cur_epoch < opts.epochs and TRAIN: # ===== Train ===== model.train() epoch_loss = trainer.train(cur_epoch=cur_epoch, optim=optimizer, train_loader=train_loader, scheduler=scheduler, logger=logger) logger.info(f"End of Epoch {cur_epoch}/{opts.epochs}, Average Loss={epoch_loss[0]+epoch_loss[1]}," f" Class Loss={epoch_loss[0]}, Reg Loss={epoch_loss[1]}") # ===== Log metrics on Tensorboard ===== logger.add_scalar("E-Loss", epoch_loss[0]+epoch_loss[1], cur_epoch) logger.add_scalar("E-Loss-reg", epoch_loss[1], cur_epoch) logger.add_scalar("E-Loss-cls", epoch_loss[0], cur_epoch) # ===== Validation ===== if (cur_epoch + 1) % opts.val_interval == 0: logger.info("validate on val set...") model.eval() val_loss, val_score, ret_samples = trainer.validate(loader=val_loader, metrics=val_metrics, ret_samples_ids=sample_ids, logger=logger) logger.print("Done validation") logger.info(f"End of Validation {cur_epoch}/{opts.epochs}, Validation Loss={val_loss[0]+val_loss[1]}," f" Class Loss={val_loss[0]}, Reg Loss={val_loss[1]}") logger.info(val_metrics.to_str(val_score)) # ===== Save Best Model ===== if rank == 0: # save best model at the last iteration score = val_score['Mean IoU'] # best model to build incremental steps save_ckpt(f"checkpoints/step/{task_name}_{opts.name}_{opts.step}.pth", model, trainer, optimizer, scheduler, cur_epoch, score) logger.info("[!] Checkpoint saved.") # ===== Log metrics on Tensorboard ===== # visualize validation score and samples logger.add_scalar("V-Loss", val_loss[0]+val_loss[1], cur_epoch) logger.add_scalar("V-Loss-reg", val_loss[1], cur_epoch) logger.add_scalar("V-Loss-cls", val_loss[0], cur_epoch) logger.add_scalar("Val_Overall_Acc", val_score['Overall Acc'], cur_epoch) logger.add_scalar("Val_MeanIoU", val_score['Mean IoU'], cur_epoch) logger.add_table("Val_Class_IoU", val_score['Class IoU'], cur_epoch) logger.add_table("Val_Acc_IoU", val_score['Class Acc'], cur_epoch) # logger.add_figure("Val_Confusion_Matrix", val_score['Confusion Matrix'], cur_epoch) # keep the metric to print them at the end of training results["V-IoU"] = val_score['Class IoU'] results["V-Acc"] = val_score['Class Acc'] for k, (img, target, lbl) in enumerate(ret_samples): img = (denorm(img) * 255).astype(np.uint8) target = label2color(target).transpose(2, 0, 1).astype(np.uint8) lbl = label2color(lbl).transpose(2, 0, 1).astype(np.uint8) concat_img = np.concatenate((img, target, lbl), axis=2) # concat along width logger.add_image(f'Sample_{k}', concat_img, cur_epoch) cur_epoch += 1 # ===== Save Best Model at the end of training ===== if rank == 0 and TRAIN: # save best model at the last iteration # best model to build incremental steps save_ckpt(f"checkpoints/step/{task_name}_{opts.name}_{opts.step}.pth", model, trainer, optimizer, scheduler, cur_epoch, best_score) logger.info("[!] Checkpoint saved.") torch.distributed.barrier() # xxx From here starts the test code logger.info("*** Test the model on all seen classes...") # make data loader test_loader = data.DataLoader(test_dst, batch_size=opts.batch_size if opts.crop_val else 1, sampler=DistributedSampler(test_dst, num_replicas=world_size, rank=rank), num_workers=opts.num_workers) # load best model if TRAIN: model = make_model(opts, classes=tasks.get_per_task_classes(opts.dataset, opts.task, opts.step)) # Put the model on GPU model = DistributedDataParallel(model.cuda(device)) ckpt = f"checkpoints/step/{task_name}_{opts.name}_{opts.step}.pth" checkpoint = torch.load(ckpt, map_location="cpu") model.load_state_dict(checkpoint["model_state"]) logger.info(f"*** Model restored from {ckpt}") del checkpoint trainer = Trainer(model, None, device=device, opts=opts) model.eval() val_loss, val_score, _ = trainer.validate(loader=test_loader, metrics=val_metrics, logger=logger) logger.print("Done test") logger.info(f"*** End of Test, Total Loss={val_loss[0]+val_loss[1]}," f" Class Loss={val_loss[0]}, Reg Loss={val_loss[1]}") logger.info(val_metrics.to_str(val_score)) logger.add_table("Test_Class_IoU", val_score['Class IoU']) logger.add_table("Test_Class_Acc", val_score['Class Acc']) logger.add_figure("Test_Confusion_Matrix", val_score['Confusion Matrix']) results["T-IoU"] = val_score['Class IoU'] results["T-Acc"] = val_score['Class Acc'] logger.add_results(results) logger.add_scalar("T_Overall_Acc", val_score['Overall Acc'], opts.step) logger.add_scalar("T_MeanIoU", val_score['Mean IoU'], opts.step) logger.add_scalar("T_MeanAcc", val_score['Mean Acc'], opts.step) logger.close()