def resume_logger(opt): # parser = argparse.ArgumentParser() # parser.add_argument( # '-opt', type=str, required=True, help='Path to option JSON file.') # opt = option.parse(json_path, is_train=True) # opt = option.dict_to_nonedict( # opt) # Convert to NoneDict, which return None for missing key. # train from scratch OR resume training if opt['path']['resume_state']: # resuming training resume_state = torch.load(opt['path']['resume_state']) else: # training from scratch resume_state = None util.mkdir_and_rename( opt['path']['experiments_root']) # rename old folder if exists util.mkdirs((path for key, path in opt['path'].items() if not key == 'experiments_root' and 'pretrain_model' not in key and 'resume' not in key)) # config loggers. Before it, the log will not work util.setup_logger(None, opt['path']['log'], 'train', level=logging.INFO, screen=True) util.setup_logger('val', opt['path']['log'], 'val', level=logging.INFO) logger = logging.getLogger('base') return resume_state, logger
def dir_check(opt): if opt['is_train']: # starting from scratch, needs to create training directory if not opt['path']['resume_state']: util.mkdir_and_rename( opt['path']['experiments_root']) # rename old folder if exists util.mkdirs( (path for key, path in opt['path'].items() if not key == 'experiments_root' and 'pretrain_model' not in key and 'resume' not in key)) else: # create testing directory util.mkdirs((path for key, path in opt['path'].items() if not key == 'pretrain_model_G'))
def setup_logging(opt, resume_state, rank): tb_logger = None if rank <= 0: # normal training (rank -1) OR distributed training (rank 0) if resume_state is None: util.mkdir_and_rename( opt['path'] ['experiments_root']) # rename experiment folder if exists util.mkdirs( (path for key, path in opt['path'].items() if not key == 'experiments_root' and 'pretrain_model' not in key and 'resume' not in key)) # config loggers. Before it, the log will not work util.setup_logger('base', opt['path']['log'], 'train_' + opt['name'], level=logging.INFO, screen=True, tofile=True) util.setup_logger('val', opt['path']['log'], 'val_' + opt['name'], level=logging.INFO, screen=True, tofile=True) logger = logging.getLogger('base') logger.info(option.dict2str(opt)) # tensorboard logger if opt['use_tb_logger'] and 'debug' not in opt['name']: version = float(torch.__version__[0:3]) if version >= 1.1: # PyTorch 1.1 from torch.utils.tensorboard import SummaryWriter else: logger.info( 'You are using PyTorch {}. Tensorboard will use [tensorboardX]' .format(version)) from tensorboardX import SummaryWriter tb_logger = SummaryWriter(log_dir='../tb_logger/' + opt['name']) else: util.setup_logger('base', opt['path']['log'], 'train', level=logging.INFO, screen=True) logger = logging.getLogger('base') return logger, tb_logger
def main(): # options parser = argparse.ArgumentParser() parser.add_argument('-opt', type=str, required=True, help='Path to option JSON file.') opt = option.parse(parser.parse_args().opt, is_train=True) util.mkdir_and_rename( opt['path']['experiments_root']) # rename old experiments if exists util.mkdirs((path for key, path in opt['path'].items() if not key == 'experiments_root' and \ not key == 'pretrain_model_G' and not key == 'pretrain_model_D')) option.save(opt) opt = option.dict_to_nonedict( opt) # Convert to NoneDict, which return None for missing key. # print to file and std_out simultaneously sys.stdout = PrintLogger(opt['path']['log']) # random seed seed = opt['train']['manual_seed'] if seed is None: seed = random.randint(1, 10000) print("Random Seed: ", seed) random.seed(seed) torch.manual_seed(seed) # create train and val dataloader for phase, dataset_opt in opt['datasets'].items(): if phase == 'train': train_set = create_dataset(dataset_opt) train_size = int( math.ceil(len(train_set) / dataset_opt['batch_size'])) print('Number of train images: {:,d}, iters: {:,d}'.format( len(train_set), train_size)) total_iters = int(opt['train']['niter']) total_epoches = int(math.ceil(total_iters / train_size)) print('Total epoches needed: {:d} for iters {:,d}'.format( total_epoches, total_iters)) train_loader = create_dataloader(train_set, dataset_opt) elif phase == 'val': val_dataset_opt = dataset_opt val_set = create_dataset(dataset_opt) val_loader = create_dataloader(val_set, dataset_opt) print('Number of val images in [{:s}]: {:d}'.format( dataset_opt['name'], len(val_set))) else: raise NotImplementedError( 'Phase [{:s}] is not recognized.'.format(phase)) assert train_loader is not None # Create model model = create_model(opt) # create logger logger = Logger(opt) current_step = 0 start_time = time.time() print('---------- Start training -------------') for epoch in range(total_epoches): for i, train_data in enumerate(train_loader): current_step += 1 if current_step > total_iters: break # training model.feed_data(train_data) model.optimize_parameters(current_step) time_elapsed = time.time() - start_time start_time = time.time() # log if current_step % opt['logger']['print_freq'] == 0: logs = model.get_current_log() print_rlt = OrderedDict() print_rlt['model'] = opt['model'] print_rlt['epoch'] = epoch print_rlt['iters'] = current_step print_rlt['time'] = time_elapsed for k, v in logs.items(): print_rlt[k] = v print_rlt['lr'] = model.get_current_learning_rate() logger.print_format_results('train', print_rlt) # save models if current_step % opt['logger']['save_checkpoint_freq'] == 0: print('Saving the model at the end of iter {:d}.'.format( current_step)) model.save(current_step) # validation if current_step % opt['train']['val_freq'] == 0: print('---------- validation -------------') start_time = time.time() avg_psnr = 0.0 idx = 0 for val_data in val_loader: idx += 1 img_name = os.path.splitext( os.path.basename(val_data['LR_path'][0]))[0] img_dir = os.path.join(opt['path']['val_images'], img_name) util.mkdir(img_dir) model.feed_data(val_data) model.test() visuals = model.get_current_visuals() sr_img = util.tensor2img(visuals['SR']) # uint8 gt_img = util.tensor2img(visuals['HR']) # uint8 # Save SR images for reference save_img_path = os.path.join(img_dir, '{:s}_{:d}.png'.format(\ img_name, current_step)) util.save_img(sr_img, save_img_path) # calculate PSNR if opt['crop_scale'] is not None: crop_size = opt['crop_scale'] else: crop_size = opt['scale'] if crop_size <= 0: cropped_sr_img = sr_img.copy() cropped_gt_img = gt_img.copy() else: if len(gt_img.shape) < 3: cropped_sr_img = sr_img[crop_size:-crop_size, crop_size:-crop_size] cropped_gt_img = gt_img[crop_size:-crop_size, crop_size:-crop_size] else: cropped_sr_img = sr_img[crop_size:-crop_size, crop_size:-crop_size, :] cropped_gt_img = gt_img[crop_size:-crop_size, crop_size:-crop_size, :] #avg_psnr += util.psnr(cropped_sr_img, cropped_gt_img) cropped_sr_img_y = bgr2ycbcr(cropped_sr_img, only_y=True) cropped_gt_img_y = bgr2ycbcr(cropped_gt_img, only_y=True) avg_psnr += util.psnr( cropped_sr_img_y, cropped_gt_img_y) ##########only y channel avg_psnr = avg_psnr / idx time_elapsed = time.time() - start_time # Save to log print_rlt = OrderedDict() print_rlt['model'] = opt['model'] print_rlt['epoch'] = epoch print_rlt['iters'] = current_step print_rlt['time'] = time_elapsed print_rlt['psnr'] = avg_psnr logger.print_format_results('val', print_rlt) print('-----------------------------------') # update learning rate model.update_learning_rate() print('Saving the final model.') model.save('latest') print('End of training.')
def main(): # options parser = argparse.ArgumentParser() parser.add_argument('-opt', type=str, required=True, help='Path to option JSON file.') opt = option.parse(parser.parse_args().opt, is_train=True) opt = option.dict_to_nonedict( opt) # Convert to NoneDict, which return None for missing key. # train from scratch OR resume training if opt['path']['resume_state']: # resuming training resume_state = torch.load(opt['path']['resume_state']) else: # training from scratch resume_state = None util.mkdir_and_rename( opt['path']['experiments_root']) # rename old folder if exists util.mkdirs((path for key, path in opt['path'].items() if not key == 'experiments_root' and 'pretrain_model' not in key and 'resume' not in key)) # config loggers. Before it, the log will not work util.setup_logger(None, opt['path']['log'], 'train', level=logging.INFO, screen=True) util.setup_logger('val', opt['path']['log'], 'val', level=logging.INFO) logger = logging.getLogger('base') if resume_state: logger.info('Resuming training from epoch: {}, iter: {}.'.format( resume_state['epoch'], resume_state['iter'])) option.check_resume(opt) # check resume options logger.info(option.dict2str(opt)) # tensorboard logger if opt['use_tb_logger'] and 'debug' not in opt['name']: from tensorboardX import SummaryWriter tb_logger = SummaryWriter(log_dir='../tb_logger/' + opt['name']) # random seed seed = opt['train']['manual_seed'] if seed is None: seed = random.randint(1, 10000) logger.info('Random seed: {}'.format(seed)) util.set_random_seed(seed) torch.backends.cudnn.benckmark = True # torch.backends.cudnn.deterministic = True # create train and val dataloader for phase, dataset_opt in opt['datasets'].items(): if phase == 'train': train_set = create_dataset(dataset_opt) train_size = int( math.ceil(len(train_set) / dataset_opt['batch_size'])) logger.info('Number of train images: {:,d}, iters: {:,d}'.format( len(train_set), train_size)) total_iters = int(opt['train']['niter']) total_epochs = int(math.ceil(total_iters / train_size)) logger.info('Total epochs needed: {:d} for iters {:,d}'.format( total_epochs, total_iters)) train_loader = create_dataloader(train_set, dataset_opt) elif phase == 'val': val_set = create_dataset(dataset_opt) val_loader = create_dataloader(val_set, dataset_opt) logger.info('Number of val images in [{:s}]: {:d}'.format( dataset_opt['name'], len(val_set))) else: raise NotImplementedError( 'Phase [{:s}] is not recognized.'.format(phase)) assert train_loader is not None # create model model = create_model(opt) # resume training if resume_state: start_epoch = resume_state['epoch'] current_step = resume_state['iter'] model.resume_training(resume_state) # handle optimizers and schedulers else: current_step = 0 start_epoch = 0 # training logger.info('Start training from epoch: {:d}, iter: {:d}'.format( start_epoch, current_step)) for epoch in range(start_epoch, total_epochs): for _, train_data in enumerate(train_loader): current_step += 1 if current_step > total_iters: break # update learning rate model.update_learning_rate() # training model.feed_data(train_data) model.optimize_parameters(current_step) # log if current_step % opt['logger']['print_freq'] == 0: logs = model.get_current_log() message = '<epoch:{:3d}, iter:{:8,d}, lr:{:.3e}> '.format( epoch, current_step, model.get_current_learning_rate()) for k, v in logs.items(): message += '{:s}: {:.4e} '.format(k, v) # tensorboard logger if opt['use_tb_logger'] and 'debug' not in opt['name']: tb_logger.add_scalar(k, v, current_step) logger.info(message) # validation if current_step % opt['train']['val_freq'] == 0: avg_psnr = 0.0 avg_IS = 0.0 idx = 0 for val_data in val_loader: idx += 1 img_name = os.path.splitext( os.path.basename(val_data['LR_path'][0]))[0] img_dir = os.path.join(opt['path']['val_images'], img_name) util.mkdir(img_dir) model.feed_data(val_data) model.test() visuals = model.get_current_visuals() sr_img = util.tensor2img(visuals['SR']) # uint8 gt_img = util.tensor2img(visuals['HR']) # uint8 # Save SR images for reference save_img_path = os.path.join(img_dir, '{:s}_{:d}.png'.format(\ img_name, current_step)) util.save_img(sr_img, save_img_path) #calculate IS IS = model.get_IS() avg_IS += IS # calculate PSNR crop_size = opt['scale'] gt_img = gt_img / 255. sr_img = sr_img / 255. cropped_sr_img = sr_img[crop_size:-crop_size, crop_size:-crop_size, :] cropped_gt_img = gt_img[crop_size:-crop_size, crop_size:-crop_size, :] avg_psnr += util.calculate_psnr(cropped_sr_img * 255, cropped_gt_img * 255) avg_psnr = avg_psnr / idx avg_IS = avg_IS / idx # log logger.info('# Validation # PSNR: {:.4e} IS : {:.4e}'.format( avg_psnr, avg_IS)) logger_val = logging.getLogger('val') # validation logger logger_val.info( '<epoch:{:3d}, iter:{:8,d}> psnr: {:.4e} is: {:.4e} '. format(epoch, current_step, avg_psnr, avg_IS)) # tensorboard logger if opt['use_tb_logger'] and 'debug' not in opt['name']: tb_logger.add_scalar('psnr', avg_psnr, current_step) tb_logger.add_scalar('IS', avg_IS, current_step) # save models and training states if current_step % opt['logger']['save_checkpoint_freq'] == 0: logger.info('Saving models and training states.') model.save(current_step) model.save_training_state(epoch, current_step) logger.info('Saving the final model.') model.save('latest') logger.info('End of training.')
def main(): #### setup options of three networks parser = argparse.ArgumentParser() parser.add_argument('-opt_P', type=str, help='Path to option YMAL file of Predictor.') parser.add_argument('-opt_C', type=str, help='Path to option YMAL file of Corrector.') parser.add_argument('-opt_F', type=str, help='Path to option YMAL file of SFTMD_Net.') parser.add_argument('--launcher', choices=['none', 'pytorch'], default='none', help='job launcher') parser.add_argument('--local_rank', type=int, default=0) args = parser.parse_args() opt_P = option.parse(args.opt_P, is_train=True) opt_C = option.parse(args.opt_C, is_train=True) opt_F = option.parse(args.opt_F, is_train=True) # convert to NoneDict, which returns None for missing keys opt_P = option.dict_to_nonedict(opt_P) opt_C = option.dict_to_nonedict(opt_C) opt_F = option.dict_to_nonedict(opt_F) # choose small opt for SFTMD test, fill path of pre-trained model_F opt_F = opt_F['sftmd'] # create PCA matrix of enough kernel batch_ker = util.random_batch_kernel(batch=30000, l=opt_P['kernel_size'], sig_min=0.2, sig_max=4.0, rate_iso=1.0, scaling=3, tensor=False) print('batch kernel shape: {}'.format(batch_ker.shape)) b = np.size(batch_ker, 0) batch_ker = batch_ker.reshape((b, -1)) pca_matrix = util.PCA(batch_ker, k=opt_P['code_length']).float() print('PCA matrix shape: {}'.format(pca_matrix.shape)) #### distributed training settings if args.launcher == 'none': # disabled distributed training opt_P['dist'] = False opt_F['dist'] = False opt_C['dist'] = False rank = -1 print('Disabled distributed training.') else: opt_P['dist'] = True opt_F['dist'] = True opt_C['dist'] = True init_dist() world_size = torch.distributed.get_world_size( ) #Returns the number of processes in the current process group rank = torch.distributed.get_rank( ) #Returns the rank of current process group torch.backends.cudnn.benchmark = True # torch.backends.cudnn.deterministic = True ###### Predictor&Corrector train ###### #### loading resume state if exists if opt_P['path'].get('resume_state', None): # distributed resuming: all load into default GPU device_id = torch.cuda.current_device() resume_state = torch.load( opt_P['path']['resume_state'], map_location=lambda storage, loc: storage.cuda(device_id)) option.check_resume(opt_P, resume_state['iter']) # check resume options else: resume_state = None #### mkdir and loggers if rank <= 0: # normal training (rank -1) OR distributed training (rank 0-7) if resume_state is None: # Predictor path util.mkdir_and_rename( opt_P['path'] ['experiments_root']) # rename experiment folder if exists util.mkdirs( (path for key, path in opt_P['path'].items() if not key == 'experiments_root' and 'pretrain_model' not in key and 'resume' not in key)) # Corrector path util.mkdir_and_rename( opt_C['path'] ['experiments_root']) # rename experiment folder if exists util.mkdirs( (path for key, path in opt_C['path'].items() if not key == 'experiments_root' and 'pretrain_model' not in key and 'resume' not in key)) # config loggers. Before it, the log will not work util.setup_logger('base', opt_P['path']['log'], 'train_' + opt_P['name'], level=logging.INFO, screen=True, tofile=True) util.setup_logger('val', opt_P['path']['log'], 'val_' + opt_P['name'], level=logging.INFO, screen=True, tofile=True) logger = logging.getLogger('base') logger.info(option.dict2str(opt_P)) logger.info(option.dict2str(opt_C)) # tensorboard logger if opt_P['use_tb_logger'] and 'debug' not in opt_P['name']: version = float(torch.__version__[0:3]) if version >= 1.1: # PyTorch 1.1 from torch.utils.tensorboard import SummaryWriter else: logger.info( 'You are using PyTorch {}. Tensorboard will use [tensorboardX]' .format(version)) from tensorboardX import SummaryWriter tb_logger = SummaryWriter(log_dir='../tb_logger/' + opt_P['name']) else: util.setup_logger('base', opt_P['path']['log'], 'train', level=logging.INFO, screen=True) logger = logging.getLogger('base') #### random seed seed = opt_P['train']['manual_seed'] if seed is None: seed = random.randint(1, 10000) if rank <= 0: logger.info('Random seed: {}'.format(seed)) util.set_random_seed(seed) torch.backends.cudnn.benchmark = True # torch.backends.cudnn.deterministic = True #### create train and val dataloader dataset_ratio = 200 # enlarge the size of each epoch for phase, dataset_opt in opt_P['datasets'].items(): if phase == 'train': train_set = create_dataset(dataset_opt) train_size = int( math.ceil(len(train_set) / dataset_opt['batch_size'])) total_iters = int(opt_P['train']['niter']) total_epochs = int(math.ceil(total_iters / train_size)) if opt_P['dist']: train_sampler = DistIterSampler(train_set, world_size, rank, dataset_ratio) total_epochs = int( math.ceil(total_iters / (train_size * dataset_ratio))) else: train_sampler = None train_loader = create_dataloader(train_set, dataset_opt, opt_P, train_sampler) if rank <= 0: logger.info( 'Number of train images: {:,d}, iters: {:,d}'.format( len(train_set), train_size)) logger.info('Total epochs needed: {:d} for iters {:,d}'.format( total_epochs, total_iters)) elif phase == 'val': val_set = create_dataset(dataset_opt) val_loader = create_dataloader(val_set, dataset_opt, opt_P, None) if rank <= 0: logger.info('Number of val images in [{:s}]: {:d}'.format( dataset_opt['name'], len(val_set))) else: raise NotImplementedError( 'Phase [{:s}] is not recognized.'.format(phase)) assert train_loader is not None assert val_loader is not None #### create model model_F = create_model(opt_F) #load pretrained model of SFTMD model_P = create_model(opt_P) model_C = create_model(opt_C) #### resume training if resume_state: logger.info('Resuming training from epoch: {}, iter: {}.'.format( resume_state['epoch'], resume_state['iter'])) start_epoch = resume_state['epoch'] current_step = resume_state['iter'] model_P.resume_training( resume_state) # handle optimizers and schedulers else: current_step = 0 start_epoch = 0 #### training logger.info('Start training from epoch: {:d}, iter: {:d}'.format( start_epoch, current_step)) for epoch in range(start_epoch, total_epochs + 1): if opt_P['dist']: train_sampler.set_epoch(epoch) for _, train_data in enumerate(train_loader): current_step += 1 if current_step > total_iters: break #### update learning rate, schedulers # model.update_learning_rate(current_step, warmup_iter=opt_P['train']['warmup_iter']) #### preprocessing for LR_img and kernel map prepro = util.SRMDPreprocessing(opt_P['scale'], pca_matrix, para_input=opt_P['code_length'], kernel=opt_P['kernel_size'], noise=False, cuda=True, sig_min=0.2, sig_max=4.0, rate_iso=1.0, scaling=3, rate_cln=0.2, noise_high=0.0) LR_img, ker_map = prepro(train_data['GT']) #### training Predictor model_P.feed_data(LR_img, ker_map) model_P.optimize_parameters(current_step) P_visuals = model_P.get_current_visuals() est_ker_map = P_visuals['Batch_est_ker_map'] #### log of model_P if current_step % opt_P['logger']['print_freq'] == 0: logs = model_P.get_current_log() message = 'Predictor <epoch:{:3d}, iter:{:8,d}, lr:{:.3e}> '.format( epoch, current_step, model_P.get_current_learning_rate()) for k, v in logs.items(): message += '{:s}: {:.4e} '.format(k, v) # tensorboard logger if opt_P['use_tb_logger'] and 'debug' not in opt_P['name']: if rank <= 0: tb_logger.add_scalar(k, v, current_step) if rank <= 0: logger.info(message) #### training Corrector for step in range(opt_C['step']): # test SFTMD for corresponding SR image model_F.feed_data(train_data, LR_img, est_ker_map) model_F.test() F_visuals = model_F.get_current_visuals() SR_img = F_visuals['Batch_SR'] # Test SFTMD to produce SR images # train corrector given SR image and estimated kernel map model_C.feed_data(SR_img, est_ker_map, ker_map) model_C.optimize_parameters(current_step) C_visuals = model_C.get_current_visuals() est_ker_map = C_visuals['Batch_est_ker_map'] #### log of model_C if current_step % opt_C['logger']['print_freq'] == 0: logs = model_C.get_current_log() message = 'Corrector <epoch:{:3d}, iter:{:8,d}, lr:{:.3e}> '.format( epoch, current_step, model_C.get_current_learning_rate()) for k, v in logs.items(): message += '{:s}: {:.4e} '.format(k, v) # tensorboard logger if opt_C['use_tb_logger'] and 'debug' not in opt_C[ 'name']: if rank <= 0: tb_logger.add_scalar(k, v, current_step) if rank <= 0: logger.info(message) # validation, to produce ker_map_list(fake) if current_step % opt_P['train']['val_freq'] == 0 and rank <= 0: avg_psnr = 0.0 idx = 0 for _, val_data in enumerate(val_loader): prepro = util.SRMDPreprocessing( opt_P['scale'], pca_matrix, para_input=opt_P['code_length'], kernel=opt_P['kernel_size'], noise=False, cuda=True, sig_min=0.2, sig_max=4.0, rate_iso=1.0, scaling=3, rate_cln=0.2, noise_high=0.0) LR_img, ker_map = prepro(val_data['GT']) single_img_psnr = 0.0 # valid Predictor model_P.feed_data(LR_img, ker_map) model_P.test() P_visuals = model_P.get_current_visuals() est_ker_map = P_visuals['Batch_est_ker_map'] for step in range(opt_C['step']): step += 1 idx += 1 model_F.feed_data(val_data, LR_img, est_ker_map) model_F.test() F_visuals = model_F.get_current_visuals() SR_img = F_visuals['Batch_SR'] # Test SFTMD to produce SR images model_C.feed_data(SR_img, est_ker_map, ker_map) model_C.test() C_visuals = model_C.get_current_visuals() est_ker_map = C_visuals['Batch_est_ker_map'] sr_img = util.tensor2img(F_visuals['SR']) # uint8 gt_img = util.tensor2img(F_visuals['GT']) # uint8 # Save SR images for reference img_name = os.path.splitext( os.path.basename(val_data['LQ_path'][0]))[0] img_dir = os.path.join(opt_P['path']['val_images'], img_name) # img_dir = os.path.join(opt_F['path']['val_images'], str(current_step), '_', str(step)) util.mkdir(img_dir) save_img_path = os.path.join( img_dir, '{:s}_{:d}_{:d}.png'.format( img_name, current_step, step)) util.save_img(sr_img, save_img_path) # calculate PSNR crop_size = opt_P['scale'] gt_img = gt_img / 255. sr_img = sr_img / 255. cropped_sr_img = sr_img[crop_size:-crop_size, crop_size:-crop_size, :] cropped_gt_img = gt_img[crop_size:-crop_size, crop_size:-crop_size, :] step_psnr = util.calculate_psnr( cropped_sr_img * 255, cropped_gt_img * 255) logger.info( '<epoch:{:3d}, iter:{:8,d}, step:{:3d}> img:{:s}, psnr: {:.4f}' .format(epoch, current_step, step, img_name, step_psnr)) single_img_psnr += step_psnr avg_psnr += util.calculate_psnr( cropped_sr_img * 255, cropped_gt_img * 255) avg_signle_img_psnr = single_img_psnr / step logger.info( '<epoch:{:3d}, iter:{:8,d}, step:{:3d}> img:{:s}, average psnr: {:.4f}' .format(epoch, current_step, step, img_name, avg_signle_img_psnr)) avg_psnr = avg_psnr / idx # log logger.info('# Validation # PSNR: {:.4f}'.format(avg_psnr)) logger_val = logging.getLogger('val') # validation logger logger_val.info( '<epoch:{:3d}, iter:{:8,d}, step:{:3d}> psnr: {:.4f}'. format(epoch, current_step, step, avg_psnr)) # tensorboard logger if opt_P['use_tb_logger'] and 'debug' not in opt_P['name']: tb_logger.add_scalar('psnr', avg_psnr, current_step) #### save models and training states if current_step % opt_P['logger']['save_checkpoint_freq'] == 0: if rank <= 0: logger.info('Saving models and training states.') model_P.save(current_step) model_P.save_training_state(epoch, current_step) model_C.save(current_step) model_C.save_training_state(epoch, current_step) if rank <= 0: logger.info('Saving the final model.') model_P.save('latest') model_C.save('latest') logger.info('End of Predictor and Corrector training.') tb_logger.close()
def SFTMD_train(opt_F, rank, world_size, pca_matrix): #### loading resume state if exists if opt_F['path'].get('resume_state', None): # distributed resuming: all load into default GPU device_id = torch.cuda.current_device() resume_state = torch.load(opt_F['path']['resume_state'], map_location=lambda storage, loc: storage.cuda(device_id)) option.check_resume(opt_F, resume_state['iter']) # check resume options else: resume_state = None #### mkdir and loggers if rank <= 0: if resume_state is None: util.mkdir_and_rename( opt_F['path']['experiments_root']) # rename experiment folder if exists util.mkdirs((path for key, path in opt_F['path'].items() if not key == 'experiments_root' and 'pretrain_model' not in key and 'resume' not in key)) # config loggers. Before it, the log will not work util.setup_logger('base', opt_F['path']['log'], 'train_' + opt_F['name'], level=logging.INFO, screen=True, tofile=True) util.setup_logger('val', opt_F['path']['log'], 'val_' + opt_F['name'], level=logging.INFO, screen=True, tofile=True) logger = logging.getLogger('base') logger.info(option.dict2str(opt_F)) # tensorboard logger if opt_F['use_tb_logger'] and 'debug' not in opt_F['name']: version = float(torch.__version__[0:3]) if version >= 1.1: # PyTorch 1.1 from torch.utils.tensorboard import SummaryWriter else: logger.info( 'You are using PyTorch {}. Tensorboard will use [tensorboardX]'.format(version)) from tensorboardX import SummaryWriter tb_logger = SummaryWriter(log_dir='../tb_logger/' + opt_F['name']) else: util.setup_logger('base', opt_F['path']['log'], 'train', level=logging.INFO, screen=True) logger = logging.getLogger('base') #### create train and val dataloader dataset_ratio = 200 # enlarge the size of each epoch for phase, dataset_opt in opt_F['datasets'].items(): if phase == 'train': train_set = create_dataset(dataset_opt) train_size = int(math.ceil(len(train_set) / dataset_opt['batch_size'])) total_iters = int(opt_F['train']['niter']) total_epochs = int(math.ceil(total_iters / train_size)) if opt_F['dist']: train_sampler = DistIterSampler(train_set, world_size, rank, dataset_ratio) total_epochs = int(math.ceil(total_iters / (train_size * dataset_ratio))) else: train_sampler = None train_loader = create_dataloader(train_set, dataset_opt, opt_F, train_sampler) if rank <= 0: logger.info('Number of train images: {:,d}, iters: {:,d}'.format( len(train_set), train_size)) logger.info('Total epochs needed: {:d} for iters {:,d}'.format( total_epochs, total_iters)) elif phase == 'val': val_set = create_dataset(dataset_opt) val_loader = create_dataloader(val_set, dataset_opt, opt_F, None) if rank <= 0: logger.info('Number of val images in [{:s}]: {:d}'.format( dataset_opt['name'], len(val_set))) else: raise NotImplementedError('Phase [{:s}] is not recognized.'.format(phase)) assert train_loader is not None assert val_loader is not None #### create model model_F = create_model(opt_F) #### resume training if resume_state: logger.info('Resuming training from epoch: {}, iter: {}.'.format( resume_state['epoch'], resume_state['iter'])) start_epoch = resume_state['epoch'] current_step = resume_state['iter'] model_F.resume_training(resume_state) # handle optimizers and schedulers else: current_step = 0 start_epoch = 0 #### training logger.info('Start training from epoch: {:d}, iter: {:d}'.format(start_epoch, current_step)) for epoch in range(start_epoch, total_epochs + 1): if opt_F['dist']: train_sampler.set_epoch(epoch) for _, train_data in enumerate(train_loader): current_step += 1 if current_step > total_iters: break #### preprocessing for LR_img and kernel map prepro = util.SRMDPreprocessing(opt_F['scale'], pca_matrix, para_input=10, kernel=21, noise=False, cuda=True, sig_min=0.2, sig_max=4.0, rate_iso=1.0, scaling=3, rate_cln=0.2, noise_high=0.0) LR_img, ker_map = prepro(train_data['GT']) #### update learning rate, schedulers model_F.update_learning_rate(current_step, warmup_iter=opt_F['train']['warmup_iter']) #### training model_F.feed_data(train_data, LR_img, ker_map) model_F.optimize_parameters(current_step) #### log if current_step % opt_F['logger']['print_freq'] == 0: logs = model_F.get_current_log() message = '<epoch:{:3d}, iter:{:8,d}, lr:{:.3e}> '.format( epoch, current_step, model_F.get_current_learning_rate()) for k, v in logs.items(): message += '{:s}: {:.4e} '.format(k, v) # tensorboard logger if opt_F['use_tb_logger'] and 'debug' not in opt_F['name']: if rank <= 0: tb_logger.add_scalar(k, v, current_step) if rank <= 0: logger.info(message) # validation if current_step % opt_F['train']['val_freq'] == 0 and rank <= 0: avg_psnr = 0.0 idx = 0 for _, val_data in enumerate(val_loader): idx += 1 #### preprocessing for LR_img and kernel map prepro = util.SRMDPreprocessing(opt_F['scale'], pca_matrix, para_input=15, noise=False, cuda=True, sig_min=0.2, sig_max=4.0, rate_iso=1.0, scaling=3, rate_cln=0.2, noise_high=0.0) LR_img, ker_map = prepro(val_data['GT']) model_F.feed_data(val_data, LR_img, ker_map) model_F.test() visuals = model_F.get_current_visuals() sr_img = util.tensor2img(visuals['SR']) # uint8 gt_img = util.tensor2img(visuals['GT']) # uint8 # Save SR images for reference img_name = os.path.splitext(os.path.basename(val_data['LQ_path'][0]))[0] #img_dir = os.path.join(opt_F['path']['val_images'], img_name) img_dir = os.path.join(opt_F['path']['val_images'], str(current_step)) util.mkdir(img_dir) save_img_path = os.path.join(img_dir,'{:s}_{:d}.png'.format(img_name, current_step)) util.save_img(sr_img, save_img_path) # calculate PSNR crop_size = opt_F['scale'] gt_img = gt_img / 255. sr_img = sr_img / 255. cropped_sr_img = sr_img[crop_size:-crop_size, crop_size:-crop_size, :] cropped_gt_img = gt_img[crop_size:-crop_size, crop_size:-crop_size, :] avg_psnr += util.calculate_psnr(cropped_sr_img * 255, cropped_gt_img * 255) avg_psnr = avg_psnr / idx # log logger.info('# Validation # PSNR: {:.4e}'.format(avg_psnr)) logger_val = logging.getLogger('val') # validation logger logger_val.info('<epoch:{:3d}, iter:{:8,d}> psnr: {:.4e}'.format(epoch, current_step, avg_psnr)) # tensorboard logger if opt_F['use_tb_logger'] and 'debug' not in opt_F['name']: tb_logger.add_scalar('psnr', avg_psnr, current_step) #### save models and training states if current_step % opt_F['logger']['save_checkpoint_freq'] == 0: if rank <= 0: logger.info('Saving models and training states.') model_F.save(current_step) model_F.save_training_state(epoch, current_step) if rank <= 0: logger.info('Saving the final model.') model_F.save('latest') logger.info('End of SFTMD training.')
import torch import options.options as option from utils import util # options parser = argparse.ArgumentParser() parser.add_argument('-opt', type=str, required=True, help='Path to options JSON file.') args = parser.parse_args() options_path = args.opt opt = option.parse(options_path, is_train=True) util.mkdir_and_rename( opt['path']['experiments_root']) # rename old experiments if exists util.mkdirs((path for key, path in opt['path'].items() if not key == 'experiments_root' and \ not key == 'pretrain_model_G' and not key == 'pretrain_model_D')) option.save(opt) # save option file to the opt['path']['options'] opt = option.dict_to_nonedict( opt) # Convert to NoneDict, which return None for missing key. # print to file and std_out simultaneously class PrintLogger(object): def __init__(self): self.terminal = sys.stdout self.log = open(os.path.join(opt['path']['log'], 'print_log.txt'), "a") def write(self, message): self.terminal.write(message)
def main(): # options parser = argparse.ArgumentParser() parser.add_argument('-opt', type=str, required=True, help='Path to option JSON file.') opt = option.parse(parser.parse_args().opt, is_train=True) # Convert to NoneDict, which return None for missing key. opt = option.dict_to_nonedict(opt) pytorch_ver = get_pytorch_ver() # train from scratch OR resume training if opt['path']['resume_state']: if os.path.isdir(opt['path']['resume_state']): import glob resume_state_path = util.sorted_nicely( glob.glob( os.path.normpath(opt['path']['resume_state']) + '/*.state'))[-1] else: resume_state_path = opt['path']['resume_state'] resume_state = torch.load(resume_state_path) else: # training from scratch resume_state = None # rename old folder if exists util.mkdir_and_rename(opt['path']['experiments_root']) util.mkdirs((path for key, path in opt['path'].items() if not key == 'experiments_root' and 'pretrain_model' not in key and 'resume' not in key)) # config loggers. Before it, the log will not work util.setup_logger(None, opt['path']['log'], 'train', level=logging.INFO, screen=True) util.setup_logger('val', opt['path']['log'], 'val', level=logging.INFO) logger = logging.getLogger('base') if resume_state: logger.info('Set [resume_state] to ' + resume_state_path) logger.info('Resuming training from epoch: {}, iter: {}.'.format( resume_state['epoch'], resume_state['iter'])) option.check_resume(opt) # check resume options logger.info(option.dict2str(opt)) # tensorboard logger if opt['use_tb_logger'] and 'debug' not in opt['name']: from tensorboardX import SummaryWriter try: # for version tensorboardX >= 1.7 tb_logger = SummaryWriter(logdir='../tb_logger/' + opt['name']) except: # for version tensorboardX < 1.6 tb_logger = SummaryWriter(log_dir='../tb_logger/' + opt['name']) # random seed seed = opt['train']['manual_seed'] if seed is None: seed = random.randint(1, 10000) logger.info('Random seed: {}'.format(seed)) util.set_random_seed(seed) torch.backends.cudnn.benckmark = True # torch.backends.cudnn.deterministic = True # create train and val dataloader for phase, dataset_opt in opt['datasets'].items(): if phase == 'train': train_set = create_dataset(dataset_opt) train_size = int( math.ceil(len(train_set) / dataset_opt['batch_size'])) logger.info('Number of train images: {:,d}, iters: {:,d}'.format( len(train_set), train_size)) total_iters = int(opt['train']['niter']) total_epochs = int(math.ceil(total_iters / train_size)) logger.info('Total epochs needed: {:d} for iters {:,d}'.format( total_epochs, total_iters)) train_loader = create_dataloader(train_set, dataset_opt) elif phase == 'val': val_set = create_dataset(dataset_opt) val_loader = create_dataloader(val_set, dataset_opt) logger.info('Number of val images in [{:s}]: {:d}'.format( dataset_opt['name'], len(val_set))) else: raise NotImplementedError( 'Phase [{:s}] is not recognized.'.format(phase)) assert train_loader is not None # create model model = create_model(opt) # resume training if resume_state: start_epoch = resume_state['epoch'] current_step = resume_state['iter'] model.resume_training(resume_state) # handle optimizers and schedulers # updated schedulers in case JSON configuration has changed model.update_schedulers(opt['train']) else: current_step = 0 start_epoch = 0 # training logger.info('Start training from epoch: {:d}, iter: {:d}'.format( start_epoch, current_step)) for epoch in range(start_epoch, total_epochs): for n, train_data in enumerate(train_loader, start=1): current_step += 1 if current_step > total_iters: break if pytorch_ver == "pre": # Order for PyTorch ver < 1.1.0 # update learning rate model.update_learning_rate(current_step - 1) # training model.feed_data(train_data) model.optimize_parameters(current_step) elif pytorch_ver == "post": # Order for PyTorch ver > 1.1.0 # training model.feed_data(train_data) model.optimize_parameters(current_step) # update learning rate model.update_learning_rate(current_step - 1) else: print('Error identifying PyTorch version. ', torch.__version__) break # log if current_step % opt['logger']['print_freq'] == 0: logs = model.get_current_log() message = '<epoch:{:3d}, iter:{:8,d}, lr:{:.3e}> '.format( epoch, current_step, model.get_current_learning_rate()) for k, v in logs.items(): message += '{:s}: {:.4e} '.format(k, v) # tensorboard logger if opt['use_tb_logger'] and 'debug' not in opt['name']: tb_logger.add_scalar(k, v, current_step) logger.info(message) # save models and training states (changed to save models before validation) if current_step % opt['logger']['save_checkpoint_freq'] == 0: model.save(current_step) model.save_training_state(epoch + (n >= len(train_loader)), current_step) logger.info('Models and training states saved.') # validation if current_step % opt['train']['val_freq'] == 0: avg_psnr = 0.0 avg_ssim = 0.0 avg_lpips = 0.0 idx = 0 val_sr_imgs_list = [] val_gt_imgs_list = [] for val_data in val_loader: idx += 1 img_name = os.path.splitext( os.path.basename(val_data['LR_path'][0]))[0] img_dir = os.path.join(opt['path']['val_images'], img_name) util.mkdir(img_dir) model.feed_data(val_data) model.test() visuals = model.get_current_visuals() if opt['datasets']['train'][ 'znorm']: # If the image range is [-1,1] sr_img = util.tensor2img(visuals['SR'], min_max=(-1, 1)) # uint8 gt_img = util.tensor2img(visuals['HR'], min_max=(-1, 1)) # uint8 else: # Default: Image range is [0,1] sr_img = util.tensor2img(visuals['SR']) # uint8 gt_img = util.tensor2img(visuals['HR']) # uint8 # sr_img = util.tensor2img(visuals['SR']) # uint8 # gt_img = util.tensor2img(visuals['HR']) # uint8 # print("Min. SR value:",sr_img.min()) # Debug # print("Max. SR value:",sr_img.max()) # Debug # print("Min. GT value:",gt_img.min()) # Debug # print("Max. GT value:",gt_img.max()) # Debug # Save SR images for reference save_img_path = os.path.join( img_dir, '{:s}_{:d}.png'.format(img_name, current_step)) util.save_img(sr_img, save_img_path) # calculate PSNR, SSIM and LPIPS distance crop_size = opt['scale'] gt_img = gt_img / 255. sr_img = sr_img / 255. # For training models with only one channel ndim==2, if RGB ndim==3, etc. if gt_img.ndim == 2: cropped_gt_img = gt_img[crop_size:-crop_size, crop_size:-crop_size] else: cropped_gt_img = gt_img[crop_size:-crop_size, crop_size:-crop_size, :] if sr_img.ndim == 2: cropped_sr_img = sr_img[crop_size:-crop_size, crop_size:-crop_size] else: # Default: RGB images cropped_sr_img = sr_img[crop_size:-crop_size, crop_size:-crop_size, :] # If calculating only once for all images val_gt_imgs_list.append(cropped_gt_img) # If calculating only once for all images val_sr_imgs_list.append(cropped_sr_img) # LPIPS only works for RGB images avg_psnr += util.calculate_psnr(cropped_sr_img * 255, cropped_gt_img * 255) avg_ssim += util.calculate_ssim(cropped_sr_img * 255, cropped_gt_img * 255) # If calculating for each image # avg_lpips += lpips.calculate_lpips([cropped_sr_img], [cropped_gt_img]) avg_psnr = avg_psnr / idx avg_ssim = avg_ssim / idx # avg_lpips=avg_lpips / idx # If calculating for each image # If calculating only once for all images avg_lpips = lpips.calculate_lpips(val_sr_imgs_list, val_gt_imgs_list) # log # logger.info('# Validation # PSNR: {:.5g}, SSIM: {:.5g}'.format(avg_psnr, avg_ssim)) logger.info( '# Validation # PSNR: {:.5g}, SSIM: {:.5g}, LPIPS: {:.5g}'. format(avg_psnr, avg_ssim, avg_lpips)) logger_val = logging.getLogger('val') # validation logger # logger_val.info('<epoch:{:3d}, iter:{:8,d}> psnr: {:.5g}, ssim: {:.5g}'.format( # epoch, current_step, avg_psnr, avg_ssim)) logger_val.info( '<epoch:{:3d}, iter:{:8,d}> psnr: {:.5g}, ssim: {:.5g}, lpips: {:.5g}' .format(epoch, current_step, avg_psnr, avg_ssim, avg_lpips)) # tensorboard logger if opt['use_tb_logger'] and 'debug' not in opt['name']: tb_logger.add_scalar('psnr', avg_psnr, current_step) tb_logger.add_scalar('ssim', avg_ssim, current_step) tb_logger.add_scalar('lpips', avg_lpips, current_step) logger.info('Saving the final model.') model.save('latest') logger.info('End of training.')
def main(): # options parser = argparse.ArgumentParser() parser.add_argument('-opt', type=str, required=True, help='Path to option JSON file.') opt = option.parse(parser.parse_args().opt, is_train=True) opt = option.dict_to_nonedict( opt) # Convert to NoneDict, which return None for missing key. # train from scratch OR resume training if opt['path']['resume_state']: # resuming training resume_state = torch.load(opt['path']['resume_state']) else: # training from scratch resume_state = None util.mkdir_and_rename( opt['path']['experiments_root']) # rename old folder if exists util.mkdirs((path for key, path in opt['path'].items() if not key == 'experiments_root' and 'pretrain_model' not in key and 'resume' not in key)) # config loggers. Before it, the log will not work util.setup_logger(None, opt['path']['log'], 'train', level=logging.INFO, screen=True) util.setup_logger('val', opt['path']['log'], 'val', level=logging.INFO) logger = logging.getLogger('base') if resume_state: # resume_state[] logger.info('Resuming training from epoch: {}, iter: {}.'.format( resume_state['epoch'], resume_state['iter'])) option.check_resume(opt) # check resume options logger.info(option.dict2str(opt)) # tensorboard logger if opt['use_tb_logger'] and 'debug' not in opt['name']: from tensorboardX import SummaryWriter tb_logger = SummaryWriter(logdir='../../SRN_tb_logger/' + opt['name']) # random seed seed = opt['train']['manual_seed'] if seed is None: seed = random.randint(1, 10000) logger.info('Random seed: {}'.format(seed)) util.set_random_seed(seed) torch.backends.cudnn.benckmark = True # torch.backends.cudnn.deterministic = True # create train and val dataloader for phase, dataset_opt in opt['datasets'].items(): if phase == 'train': train_set = create_dataset(dataset_opt) train_size = int( math.ceil(len(train_set) / dataset_opt['batch_size'])) logger.info('Number of train images: {:,d}, iters: {:,d}'.format( len(train_set), train_size)) total_iters = int(opt['train']['niter']) total_epochs = int(math.ceil(total_iters / train_size)) logger.info('Total epochs needed: {:d} for iters {:,d}'.format( total_epochs, total_iters)) train_loader = create_dataloader(train_set, dataset_opt) elif phase == 'val': val_set = create_dataset(dataset_opt) val_loader = create_dataloader(val_set, dataset_opt) logger.info('Number of val images in [{:s}]: {:d}'.format( dataset_opt['name'], len(val_set))) else: raise NotImplementedError( 'Phase [{:s}] is not recognized.'.format(phase)) assert train_loader is not None # create model model = create_model(opt) # resume training if resume_state: start_epoch = resume_state['epoch'] current_step = resume_state['iter'] model.resume_training(resume_state, opt['train']) # handle optimizers and schedulers else: current_step = 0 start_epoch = 0 # training logger.info('Start training from epoch: {:d}, iter: {:d}'.format( start_epoch, current_step)) for epoch in range(start_epoch, total_epochs): for _, train_data in enumerate(train_loader): current_step += 1 if current_step > total_iters: break # update learning rate model.update_learning_rate() # training model.feed_data(train_data, True) model.optimize_parameters(current_step) # log if current_step % opt['logger']['print_freq'] == 0: logs = model.get_current_log() message = '<epoch:{:3d}, iter:{:8,d}, lr:{:.3e}> '.format( epoch, current_step, model.get_current_learning_rate()) for k, v in logs.items(): message += '{:s}: {:.4e} '.format(k, v) # tensorboard logger if opt['use_tb_logger'] and 'debug' not in opt['name']: tb_logger.add_scalar(k, v, current_step) logger.info(message) # training samples if opt['train']['save_tsamples'] and current_step % opt['train'][ 'save_tsamples'] == 0: fake_LRs = os.listdir( opt['datasets']['train']['dataroot_fake_LR']) real_LRs = os.listdir( opt['datasets']['train']['dataroot_real_LR']) HRs = os.listdir(opt['datasets']['train']['dataroot_HR']) for i in range(5): random_index = np.random.choice(range(len(fake_LRs))) fake_LR_path = os.path.join( opt['datasets']['train']['dataroot_fake_LR'], fake_LRs[random_index]) real_LR_path = os.path.join( opt['datasets']['train']['dataroot_real_LR'], real_LRs[random_index]) HR_path = os.path.join( opt['datasets']['train']['dataroot_HR'], HRs[random_index]) fake_LR = np.array(Image.open(fake_LR_path)) real_LR = np.array(Image.open(real_LR_path)) HR = np.array(Image.open(HR_path)) h, w, _ = fake_LR.shape fake_LR = fake_LR[h // 2 - 64:h // 2 + 64, w // 2 - 64:w // 2 + 64, :] h, w, _ = HR.shape HR = HR[h // 2 - 64 * 4:h // 2 + 64 * 4, w // 2 - 64 * 4:w // 2 + 64 * 4, :] h, w, _ = real_LR.shape real_LR = real_LR[h // 2 - 64:h // 2 + 64, w // 2 - 64:w // 2 + 64, :] fake_LR = torch.from_numpy( np.ascontiguousarray(np.transpose( fake_LR, (2, 0, 1)))).float().unsqueeze(0) / 255 real_LR = torch.from_numpy( np.ascontiguousarray(np.transpose( real_LR, (2, 0, 1)))).float().unsqueeze(0) / 255 HR = torch.from_numpy( np.ascontiguousarray(np.transpose( HR, (2, 0, 1)))).float().unsqueeze(0) / 255 LR = torch.cat([fake_LR, real_LR], dim=0) data = {'LR': LR, 'HR': HR} model.feed_data(data, False) model.test(tsamples=True) visuals = model.get_current_visuals(tsamples=True) fake_SR = visuals['SR'][0] real_SR = visuals['SR'][1] fake_hf = visuals['hf'][0] real_hf = visuals['hf'][1] HR = visuals['HR'] HR_hf = visuals['HR_hf'][0] # image_1 = torch.cat([fake_LR[0], fake_SR[0]], dim=2) # image_2 = torch.cat([real_LR[0], real_SR[0]], dim=2) image_1 = np.clip(torch.cat([fake_SR, HR, real_SR], dim=2), 0, 1) image_2 = np.clip( torch.cat([fake_hf, HR_hf, real_hf], dim=2), 0, 1) image = torch.cat([image_1, image_2], dim=1) tb_logger.add_image( 'train/train_samples_{}'.format(str(i)), image, current_step) logger.info('Saved training Samples') # validation if current_step % opt['train']['val_freq'] == 0: avg_psnr = 0.0 idx = 0 avg_lpips = 0.0 for val_data in val_loader: idx += 1 img_name = os.path.splitext( os.path.basename(val_data['LR_path'][0]))[0] img_dir = os.path.join(opt['path']['val_images'], img_name) util.mkdir(img_dir) model.feed_data(val_data, False) model.test() visuals = model.get_current_visuals() sr_img = util.tensor2img(visuals['SR']) # uint8 if 'HR' in opt['datasets']['val']['mode']: gt_img = util.tensor2img(visuals['HR']) # uint8 log_info = '{}'.format( val_data['HR_path'][0].split('/')[-1]) if opt['val_lpips']: lpips = visuals['LPIPS'] avg_lpips += lpips log_info += ' LPIPS:{:.3f}'.format( lpips.numpy()) if opt['use_domain_distance_map']: ada_w = visuals['adaptive_weights'] log_info += ' Adaptive weights:{:.2f}'.format( ada_w.numpy()) # logger.info('{} LPIPS: {:.3f}'.format(val_data['HR_path'][0].split('/')[-1], lpips.numpy())) # print('img:', val_data['HR_path'][0].split('/')[-1], 'LPIPS: %.3f' % lpips.numpy()) # else: # print('img:', val_data['LR_path'][0].split('/')[-1]) logger.info(log_info) # Save SR images for reference save_img_path = os.path.join(img_dir, '{:s}_{:d}.png'.format(\ img_name, current_step)) util.save_img(sr_img, save_img_path) # calculate PSNR if 'HR' in opt['datasets']['val']['mode']: crop_size = opt['scale'] gt_img = gt_img / 255. sr_img = sr_img / 255. cropped_sr_img = sr_img[crop_size:-crop_size, crop_size:-crop_size, :] cropped_gt_img = gt_img[crop_size:-crop_size, crop_size:-crop_size, :] avg_psnr += util.calculate_psnr( cropped_sr_img * 255, cropped_gt_img * 255) avg_psnr = avg_psnr / idx if opt['val_lpips']: avg_lpips = avg_lpips / idx print('Mean LPIPS:', avg_lpips.numpy()) # log logger.info('# Validation # PSNR: {:.4e}'.format(avg_psnr)) logger_val = logging.getLogger('val') # validation logger if opt['val_lpips']: logger_val.info( '<epoch:{:3d}, iter:{:8,d}> psnr: {:.4e}, LPIPS: {:.4f}' .format(epoch, current_step, avg_psnr, avg_lpips)) else: logger_val.info( '<epoch:{:3d}, iter:{:8,d}> psnr: {:.4e}'.format( epoch, current_step, avg_psnr)) # tensorboard logger if opt['use_tb_logger'] and 'debug' not in opt['name']: tb_logger.add_scalar('psnr', avg_psnr, current_step) tb_logger.add_scalar('LPIPS', avg_lpips, current_step) # save models and training states if current_step % opt['logger']['save_checkpoint_freq'] == 0: logger.info('Saving models and training states.') model.save(current_step) model.save_training_state(epoch, current_step) logger.info('Saving the final model.') model.save('latest') logger.info('End of training.')
def main(): ############################################ # # set options # ############################################ parser = argparse.ArgumentParser() parser.add_argument('--opt', type=str, help='Path to option YAML file.') parser.add_argument('--launcher', choices=['none', 'pytorch'], default='none', help='job launcher') parser.add_argument('--local_rank', type=int, default=0) args = parser.parse_args() opt = option.parse(args.opt, is_train=True) ############################################ # # distributed training settings # ############################################ if args.launcher == 'none': # disabled distributed training opt['dist'] = False rank = -1 print('Disabled distributed training.') else: opt['dist'] = True init_dist() world_size = torch.distributed.get_world_size() rank = torch.distributed.get_rank() print("Rank:", rank) print("World Size", world_size) print("------------------DIST-------------------------") ############################################ # # loading resume state if exists # ############################################ if opt['path'].get('resume_state', None): # distributed resuming: all load into default GPU device_id = torch.cuda.current_device() resume_state = torch.load( opt['path']['resume_state'], map_location=lambda storage, loc: storage.cuda(device_id)) option.check_resume(opt, resume_state['iter']) # check resume options else: resume_state = None ############################################ # # mkdir and loggers # ############################################ if 'debug' in opt['name']: debug_mode = True else: debug_mode = False if rank <= 0: # normal training (rank -1) OR distributed training (rank 0) if resume_state is None: util.mkdir_and_rename( opt['path'] ['experiments_root']) # rename experiment folder if exists util.mkdirs( (path for key, path in opt['path'].items() if not key == 'experiments_root' and 'pretrain_model' not in key and 'resume' not in key)) # config loggers. Before it, the log will not work util.setup_logger('base', opt['path']['log'], 'train_' + opt['name'], level=logging.INFO, screen=True, tofile=True) util.setup_logger('base_val', opt['path']['log'], 'val_' + opt['name'], level=logging.INFO, screen=True, tofile=True) logger = logging.getLogger('base') logger_val = logging.getLogger('base_val') logger.info(option.dict2str(opt)) # tensorboard logger if opt['use_tb_logger'] and 'debug' not in opt['name']: version = float(torch.__version__[0:3]) if version >= 1.1: # PyTorch 1.1 from torch.utils.tensorboard import SummaryWriter else: logger.info( 'You are using PyTorch {}. Tensorboard will use [tensorboardX]' .format(version)) from tensorboardX import SummaryWriter tb_logger = SummaryWriter(log_dir='../tb_logger/' + opt['name']) else: # config loggers. Before it, the log will not work util.setup_logger('base', opt['path']['log'], 'train_', level=logging.INFO, screen=True) print("set train log") util.setup_logger('base_val', opt['path']['log'], 'val_', level=logging.INFO, screen=True) print("set val log") logger = logging.getLogger('base') logger_val = logging.getLogger('base_val') # convert to NoneDict, which returns None for missing keys opt = option.dict_to_nonedict(opt) #### random seed seed = opt['train']['manual_seed'] if seed is None: seed = random.randint(1, 10000) if rank <= 0: logger.info('Random seed: {}'.format(seed)) util.set_random_seed(seed) torch.backends.cudnn.benchmark = True # torch.backends.cudnn.deterministic = True ############################################ # # create train and val dataloader # ############################################ #### # dataset_ratio = 200 # enlarge the size of each epoch dataset_ratio = 200 # enlarge the size of each epoch for phase, dataset_opt in opt['datasets'].items(): if phase == 'train': if opt['datasets']['train'].get('split', None): train_set, val_set = create_dataset(dataset_opt) else: train_set = create_dataset(dataset_opt) train_size = int( math.ceil(len(train_set) / dataset_opt['batch_size'])) # total_iters = int(opt['train']['niter']) # total_epochs = int(math.ceil(total_iters / train_size)) total_iters = train_size total_epochs = int(opt['train']['epoch']) if opt['dist']: train_sampler = DistIterSampler(train_set, world_size, rank, dataset_ratio) # total_epochs = int(math.ceil(total_iters / (train_size * dataset_ratio))) total_epochs = int(opt['train']['epoch']) if opt['train']['enable'] == False: total_epochs = 1 else: # train_sampler = None train_sampler = RandomBalancedSampler(train_set, train_size) train_loader = create_dataloader(train_set, dataset_opt, opt, train_sampler, vscode_debug=debug_mode) if rank <= 0: logger.info( 'Number of train images: {:,d}, iters: {:,d}'.format( len(train_set), train_size)) logger.info('Total epochs needed: {:d} for iters {:,d}'.format( total_epochs, total_iters)) elif phase == 'val': if not opt['datasets']['train'].get('split', None): val_set = create_dataset(dataset_opt) val_loader = create_dataloader(val_set, dataset_opt, opt, None, vscode_debug=debug_mode) if rank <= 0: logger.info('Number of val images in [{:s}]: {:d}'.format( dataset_opt['name'], len(val_set))) else: raise NotImplementedError( 'Phase [{:s}] is not recognized.'.format(phase)) assert train_loader is not None ############################################ # # create model # ############################################ #### model = create_model(opt) print("Model Created! ") #### resume training if resume_state: logger.info('Resuming training from epoch: {}, iter: {}.'.format( resume_state['epoch'], resume_state['iter'])) start_epoch = resume_state['epoch'] current_step = resume_state['iter'] model.resume_training(resume_state) # handle optimizers and schedulers else: current_step = 0 start_epoch = 0 print("Not Resume Training") ############################################ # # training # ############################################ logger.info('Start training from epoch: {:d}, iter: {:d}'.format( start_epoch, current_step)) model.train_AverageMeter() saved_total_loss = 10e10 saved_total_PSNR = -1 saved_total_SSIM = -1 for epoch in range(start_epoch, total_epochs): ############################################ # # Start a new epoch # ############################################ current_step = 0 if opt['dist']: train_sampler.set_epoch(epoch) for train_idx, train_data in enumerate(train_loader): # print('current_step', current_step) if 'debug' in opt['name']: img_dir = os.path.join(opt['path']['train_images']) util.mkdir(img_dir) LQs = train_data['LQs'] # B N C H W if not 'sr' in opt['name']: GTenh = train_data['GTenh'] GTinp = train_data['GTinp'] for imgs, name in zip([LQs, GTenh, GTinp], ['LQs', 'GTenh', 'GTinp']): num = imgs.size(1) for i in range(num): img = util.tensor2img(imgs[0, i, ...]) # uint8 save_img_path = os.path.join( img_dir, '{:4d}_{:s}_{:1d}.png'.format( train_idx, str(name), i)) util.save_img(img, save_img_path) else: if 'GT' in train_data: GT_name = 'GT' elif 'GTs' in train_data: GT_name = 'GTs' GT = train_data[GT_name] for imgs, name in zip([LQs, GT], ['LQs', GT_name]): if name == 'GT': num = imgs.size(0) img = util.tensor2img(imgs[0, ...]) # uint8 save_img_path = os.path.join( img_dir, '{:4d}_{:s}_{:1d}.png'.format( train_idx, str(name), 0)) util.save_img(img, save_img_path) elif name == 'GTs': num = imgs.size(1) for i in range(num): img = util.tensor2img(imgs[:, i, ...]) # uint8 save_img_path = os.path.join( img_dir, '{:4d}_{:s}_{:1d}.png'.format( train_idx, str(name), i)) util.save_img(img, save_img_path) else: num = imgs.size(1) for i in range(num): img = util.tensor2img(imgs[:, i, ...]) # uint8 save_img_path = os.path.join( img_dir, '{:4d}_{:s}_{:1d}.png'.format( train_idx, str(name), i)) util.save_img(img, save_img_path) if (train_idx >= 3): # set to 0, just do validation break # if pre-load weight first do validation and skip the first epoch # if opt['path'].get('pretrain_model_G', None) and epoch == 0: # epoch += 1 # break if opt['train']['enable'] == False: message_train_loss = 'None' break current_step += 1 if current_step > total_iters: print("Total Iteration Reached !") break #### update learning rate if opt['train']['lr_scheme'] == 'ReduceLROnPlateau': pass else: model.update_learning_rate( current_step, warmup_iter=opt['train']['warmup_iter']) #### training model.feed_data(train_data) model.optimize_parameters(current_step) model.train_AverageMeter_update() #### log if current_step % opt['logger']['print_freq'] == 0: logs_inst, logs_avg = model.get_current_log( ) # training loss mode='train' message = '[epoch:{:3d}, iter:{:8,d}, lr:('.format( epoch, current_step) for v in model.get_current_learning_rate(): message += '{:.3e},'.format(v) message += ')] ' # if 'debug' in opt['name']: # debug model print the instant loss # for k, v in logs_inst.items(): # message += '{:s}: {:.4e} '.format(k, v) # # tensorboard logger # if opt['use_tb_logger'] and 'debug' not in opt['name']: # if rank <= 0: # tb_logger.add_scalar(k, v, current_step) # for avg loss current_iters_epoch = epoch * total_iters + current_step for k, v in logs_avg.items(): message += '{:s}: {:.4e} '.format(k, v) # tensorboard logger if opt['use_tb_logger'] and 'debug' not in opt['name']: if rank <= 0: tb_logger.add_scalar(k, v, current_iters_epoch) if rank <= 0: logger.info(message) # saving models if epoch == 1: save_filename = '{:04d}_{}.pth'.format(0, 'G') save_path = os.path.join(opt['path']['models'], save_filename) if os.path.exists(save_path): os.remove(save_path) save_filename = '{:04d}_{}.pth'.format(epoch - 1, 'G') save_path = os.path.join(opt['path']['models'], save_filename) if os.path.exists(save_path): os.remove(save_path) if rank <= 0: logger.info('Saving models and training states.') save_filename = '{:04d}'.format(epoch) model.save(save_filename) # ======================================================================= # # Main validation loop # # ======================================================================= # if opt['datasets'].get('val', None): if opt['dist']: # multi-GPU testing psnr_rlt = {} # with border and center frames psnr_rlt_avg = {} psnr_total_avg = 0. ssim_rlt = {} # with border and center frames ssim_rlt_avg = {} ssim_total_avg = 0. val_loss_rlt = {} # the averaged loss val_loss_rlt_avg = {} val_loss_total_avg = 0. if rank == 0: pbar = util.ProgressBar(len(val_set)) for idx in range( rank, len(val_set), world_size): # distributed parallel validation # print('idx', idx) if 'debug' in opt['name']: if (idx >= 3): break if (idx >= 1000): break val_data = val_set[idx] # use idx method to fetch must extend batch dimension val_data['LQs'].unsqueeze_(0) val_data['GTenh'].unsqueeze_(0) val_data['GTinp'].unsqueeze_(0) key = val_data['key'][0] # IMG_0034_00809 max_idx = len(val_set) val_name = 'val_set' num = model.get_info( ) # each model has different number of loss if psnr_rlt.get(val_name, None) is None: psnr_rlt[val_name] = torch.zeros([num, max_idx], dtype=torch.float32, device='cuda') if ssim_rlt.get(val_name, None) is None: ssim_rlt[val_name] = torch.zeros([num, max_idx], dtype=torch.float32, device='cuda') if val_loss_rlt.get(val_name, None) is None: val_loss_rlt[val_name] = torch.zeros( [num, max_idx], dtype=torch.float32, device='cuda') model.feed_data(val_data) model.test() avg_loss, loss_list = model.get_loss(ret=1) save_enable = True if idx >= 100: save_enable = False psnr_list, ssim_list = model.compute_current_psnr_ssim( save=save_enable, name=key, save_path=opt['path']['val_images']) # print('psnr_list',psnr_list) assert len(loss_list) == num assert len(psnr_list) == num for i in range(num): psnr_rlt[val_name][i, idx] = psnr_list[i] ssim_rlt[val_name][i, idx] = ssim_list[i] val_loss_rlt[val_name][i, idx] = loss_list[i] # print('psnr_rlt[val_name][i, idx]',psnr_rlt[val_name][i, idx]) # print('ssim_rlt[val_name][i, idx]',ssim_rlt[val_name][i, idx]) # print('val_loss_rlt[val_name][i, idx] ',val_loss_rlt[val_name][i, idx] ) if rank == 0: for _ in range(world_size): pbar.update('Test {} - {}/{}'.format( key, idx, max_idx)) # # collect data for _, v in psnr_rlt.items(): for i in v: dist.reduce(i, 0) for _, v in ssim_rlt.items(): for i in v: dist.reduce(i, 0) for _, v in val_loss_rlt.items(): for i in v: dist.reduce(i, 0) dist.barrier() if rank == 0: psnr_rlt_avg = {} psnr_total_avg = 0. for k, v in psnr_rlt.items(): # key, value # print('k', k, 'v', v, 'v.shape', v.shape) psnr_rlt_avg[k] = [] for i in range(num): non_zero_idx = v[i, :].nonzero() # logger.info('non_zero_idx {}'.format(non_zero_idx.shape)) # check matrix = v[i, :][non_zero_idx] # print('matrix', matrix) value = torch.mean(matrix).cpu().item() # print('value', value) psnr_rlt_avg[k].append(value) psnr_total_avg += psnr_rlt_avg[k][i] psnr_total_avg = psnr_total_avg / (len(psnr_rlt) * num) log_p = '# Validation # Avg. PSNR: {:.2f},'.format( psnr_total_avg) for k, v in psnr_rlt_avg.items(): for i, it in enumerate(v): log_p += ' {}: {:.2f}'.format(i, it) logger.info(log_p) logger_val.info(log_p) # ssim ssim_rlt_avg = {} ssim_total_avg = 0. for k, v in ssim_rlt.items(): ssim_rlt_avg[k] = [] for i in range(num): non_zero_idx = v[i, :].nonzero() # print('non_zero_idx', non_zero_idx) matrix = v[i, :][non_zero_idx] # print('matrix', matrix) value = torch.mean(matrix).cpu().item() # print('value', value) ssim_rlt_avg[k].append( torch.mean(matrix).cpu().item()) ssim_total_avg += ssim_rlt_avg[k][i] ssim_total_avg /= (len(ssim_rlt) * num) log_s = '# Validation # Avg. SSIM: {:.2f},'.format( ssim_total_avg) for k, v in ssim_rlt_avg.items(): for i, it in enumerate(v): log_s += ' {}: {:.2f}'.format(i, it) logger.info(log_s) logger_val.info(log_s) # added val_loss_rlt_avg = {} val_loss_total_avg = 0. for k, v in val_loss_rlt.items(): # k, key, the folder name # v, value, the torch matrix val_loss_rlt_avg[k] = [] # loss0 - loss_N for i in range(num): non_zero_idx = v[i, :].nonzero() # print('non_zero_idx', non_zero_idx) matrix = v[i, :][non_zero_idx] # print('matrix', matrix) value = torch.mean(matrix).cpu().item() # print('value', value) val_loss_rlt_avg[k].append( torch.mean(matrix).cpu().item()) val_loss_total_avg += val_loss_rlt_avg[k][i] val_loss_total_avg /= (len(val_loss_rlt) * num) log_l = '# Validation # Avg. Loss: {:.4e},'.format( val_loss_total_avg) for k, v in val_loss_rlt_avg.items(): for i, it in enumerate(v): log_l += ' {}: {:.4e}'.format(i, it) logger.info(log_l) logger_val.info(log_l) message = '' for v in model.get_current_learning_rate(): message += '{:.5e}'.format(v) logger_val.info( 'Epoch {:02d}, LR {:s}, PSNR {:.4f}, SSIM {:.4f}, Val Loss {:.4e}' .format(epoch, message, psnr_total_avg, ssim_total_avg, val_loss_total_avg)) else: pbar = util.ProgressBar(len(val_loader)) model.val_loss_AverageMeter() model.val_AverageMeter_para() for val_inx, val_data in enumerate(val_loader): # if 'debug' in opt['name']: # if (val_inx >= 10): # break save_enable = True if val_inx >= 100: save_enable = False if val_inx >= 100: break key = val_data['key'][0] folder = key[:-6] model.feed_data(val_data) model.test() avg_loss, loss_list = model.get_loss(ret=1) model.val_loss_AverageMeter_update(loss_list, avg_loss) psnr_list, ssim_list = model.compute_current_psnr_ssim( save=save_enable, name=key, save_path=opt['path']['val_images']) model.val_AverageMeter_para_update(psnr_list, ssim_list) if 'debug' in opt['name']: msg_psnr = '' msg_ssim = '' for i, psnr in enumerate(psnr_list): msg_psnr += '{} :{:.02f} '.format(i, psnr) for i, ssim in enumerate(ssim_list): msg_ssim += '{} :{:.02f} '.format(i, ssim) logger.info('{}_{:02d} {}'.format( key, val_inx, msg_psnr)) logger.info('{}_{:02d} {}'.format( key, val_inx, msg_ssim)) pbar.update('Test {} - {}'.format(key, val_inx)) # toal validation log lr = '' for v in model.get_current_learning_rate(): lr += '{:.5e}'.format(v) logs_avg, logs_psnr_avg, psnr_total_avg, ssim_total_avg, val_loss_total_avg = model.get_current_log( mode='val') msg_logs_avg = '' for k, v in logs_avg.items(): msg_logs_avg += '{:s}: {:.4e} '.format(k, v) logger_val.info('Val-Epoch {:02d}, LR {:s}, {:s}'.format( epoch, lr, msg_logs_avg)) logger.info('Val-Epoch {:02d}, LR {:s}, {:s}'.format( epoch, lr, msg_logs_avg)) msg_logs_psnr_avg = '' for k, v in logs_psnr_avg.items(): msg_logs_psnr_avg += '{:s}: {:.4e} '.format(k, v) logger_val.info('Val-Epoch {:02d}, LR {:s}, {:s}'.format( epoch, lr, msg_logs_psnr_avg)) logger.info('Val-Epoch {:02d}, LR {:s}, {:s}'.format( epoch, lr, msg_logs_psnr_avg)) # tensorboard logger if opt['use_tb_logger'] and 'debug' not in opt['name']: tb_logger.add_scalar('val_psnr', psnr_total_avg, epoch) tb_logger.add_scalar('val_loss', val_loss_total_avg, epoch) ############################################ # # end of validation, save model # ############################################ # if rank <= 0: logger.info("Finished an epoch, Check and Save the model weights") # we check the validation loss instead of training loss. OK~ if saved_total_loss >= val_loss_total_avg: saved_total_loss = val_loss_total_avg #torch.save(model.state_dict(), args.save_path + "/best" + ".pth") model.save('best') logger.info( "Best Weights updated for decreased validation loss") else: logger.info( "Weights Not updated for undecreased validation loss") if saved_total_PSNR <= psnr_total_avg: saved_total_PSNR = psnr_total_avg model.save('bestPSNR') logger.info( "Best Weights updated for increased validation PSNR") else: logger.info( "Weights Not updated for unincreased validation PSNR") ############################################ # # end of one epoch, schedule LR # ############################################ model.train_AverageMeter_reset() # add scheduler todo if opt['train']['lr_scheme'] == 'ReduceLROnPlateau': for scheduler in model.schedulers: # scheduler.step(val_loss_total_avg) scheduler.step(val_loss_total_avg) if rank <= 0: logger.info('Saving the final model.') model.save('last') logger.info('End of training.') tb_logger.close()
def main(): #### options parser = argparse.ArgumentParser() parser.add_argument("-opt", type=str, help="Path to option YAML file.") parser.add_argument( "--launcher", choices=["none", "pytorch"], default="none", help="job launcher" ) parser.add_argument("--local_rank", type=int, default=0) args = parser.parse_args() opt = option.parse(args.opt, is_train=True) #### distributed training settings if args.launcher == "none": # disabled distributed training opt["dist"] = False rank = -1 print("Disabled distributed training.") else: opt["dist"] = True init_dist() world_size = torch.distributed.get_world_size() rank = torch.distributed.get_rank() #### loading resume state if exists if opt["path"].get("resume_state", None): # distributed resuming: all load into default GPU device_id = torch.cuda.current_device() resume_state = torch.load( opt["path"]["resume_state"], map_location=lambda storage, loc: storage.cuda(device_id), ) option.check_resume(opt, resume_state["iter"]) # check resume options else: resume_state = None #### mkdir and loggers if rank <= 0: # normal training (rank -1) OR distributed training (rank 0) if resume_state is None: util.mkdir_and_rename( opt["path"]["experiments_root"] ) # rename experiment folder if exists util.mkdirs( ( path for key, path in opt["path"].items() if not key == "experiments_root" and "pretrain_model" not in key and "resume" not in key ) ) # config loggers. Before it, the log will not work util.setup_logger( "base", opt["path"]["log"], "train_" + opt["name"], level=logging.INFO, screen=True, tofile=True, ) logger = logging.getLogger("base") logger.info(option.dict2str(opt)) # tensorboard logger if opt["use_tb_logger"] and "debug" not in opt["name"]: version = float(torch.__version__[0:3]) if version >= 1.1: # PyTorch 1.1 from torch.utils.tensorboard import SummaryWriter else: logger.info( "You are using PyTorch {}. Tensorboard will use [tensorboardX]".format( version ) ) from tensorboardX import SummaryWriter tb_logger = SummaryWriter(log_dir="../tb_logger/" + opt["name"]) else: util.setup_logger( "base", opt["path"]["log"], "train", level=logging.INFO, screen=True ) logger = logging.getLogger("base") # convert to NoneDict, which returns None for missing keys opt = option.dict_to_nonedict(opt) #### random seed seed = opt["train"]["manual_seed"] if seed is None: seed = random.randint(1, 10000) if rank <= 0: logger.info("Random seed: {}".format(seed)) util.set_random_seed(seed) torch.backends.cudnn.benchmark = True # torch.backends.cudnn.deterministic = True #### create train and val dataloader dataset_ratio = 200 # enlarge the size of each epoch for phase, dataset_opt in opt["datasets"].items(): if phase == "train": train_set = create_dataset(dataset_opt) train_size = int(math.ceil(len(train_set) / dataset_opt["batch_size"])) total_iters = int(opt["train"]["niter"]) total_epochs = int(math.ceil(total_iters / train_size)) if opt["dist"]: train_sampler = DistIterSampler( train_set, world_size, rank, dataset_ratio ) total_epochs = int( math.ceil(total_iters / (train_size * dataset_ratio)) ) else: train_sampler = None train_loader = create_dataloader(train_set, dataset_opt, opt, train_sampler) if rank <= 0: logger.info( "Number of train images: {:,d}, iters: {:,d}".format( len(train_set), train_size ) ) logger.info( "Total epochs needed: {:d} for iters {:,d}".format( total_epochs, total_iters ) ) elif phase == "val": pass # val_set = create_dataset(dataset_opt, isVal=True) # val_loader = create_dataloader(val_set, dataset_opt, opt, None) # if rank <= 0: # logger.info( # "Number of val images in [{:s}]: {:d}".format( # dataset_opt["name"], len(val_set) # ) # ) else: raise NotImplementedError("Phase [{:s}] is not recognized.".format(phase)) assert train_loader is not None #### create model # model_path = opt["path"]["pretrain_model_G"] model = create_model(opt) #### resume training if resume_state: logger.info( "Resuming training from epoch: {}, iter: {}.".format( resume_state["epoch"], resume_state["iter"] ) ) start_epoch = resume_state["epoch"] current_step = resume_state["iter"] model.resume_training(resume_state) # handle optimizers and schedulers else: current_step = 0 start_epoch = 0 #### training logger.info( "Start training from epoch: {:d}, iter: {:d}".format(start_epoch, current_step) ) for epoch in range(start_epoch, total_epochs + 1): if opt["dist"]: train_sampler.set_epoch(epoch) for _, train_data in enumerate(train_loader): current_step += 1 if current_step > total_iters: break #### update learning rate model.update_learning_rate( current_step, warmup_iter=opt["train"]["warmup_iter"] ) #### training model.feed_data(train_data) model.optimize_parameters(current_step) #### log if current_step % opt["logger"]["print_freq"] == 0: logs = model.get_current_log() message = "<epoch:{:3d}, iter:{:8,d}, lr:(".format(epoch, current_step) for v in model.get_current_learning_rate(): message += "{:.3e},".format(v) message += ")>" for k, v in logs.items(): message += "{:s}: {:.4e} ".format(k, v) # tensorboard logger if opt["use_tb_logger"] and "debug" not in opt["name"]: if rank <= 0: tb_logger.add_scalar(k, v, current_step) if rank <= 0: logger.info(message) #### validation # currently, it does not support validation during training # if current_step % opt["train"]["val_freq"] == 0: # avg_psnr = 0 # idx = 0 # for val_data in val_loader: # idx += 1 # key = ( # val_data["key"][0] # if type(val_data["key"]) is list # else val_data["key"] # ) # imgName = key + ".png" # savePath = os.path.join( # opt["path"]["val_images"], str(current_step), imgName # ) # model.feed_data(val_data) # model.test() # output = model.get_current_visuals() # hr = util.tensor2img(output["GT"]) # sr = util.tensor2img(output["restore"]) # # Cropping to calculate PSNR # hr /= 255.0 # sr /= 255.0 # scale = 4 # H, W, C = hr.shape # H_r, W_r = H % scale, W % scale # cropped_hr = hr[: H - H_r, : W - W_r, :] # cropped_sr = sr[: H - H_r, : W - W_r, :] # avg_psnr += util.calculate_psnr(cropped_sr * 255, cropped_hr * 255) # logger.info("Saving output in {}".format(savePath)) # util.mkdir(savePath) # util.save_img( # output, joinPath(savePath, str(current_step) + ".png") # ) # avg_psnr /= idx # # log # logger.info("# Validation # PSNR: {:.4e}".format(avg_psnr)) # logger_val = logging.getLogger("val") # validation logger # logger_val.info( # "<epoch:{:3d}, iter:{:8,d}> psnr: {:.4e}".format( # epoch, current_step, avg_psnr # ) # ) # # tensorboard logger # if opt["use_tb_logger"] and "debug" not in opt["name"]: # tb_logger.add_scalar("psnr", avg_psnr, current_step) #### save models and training states if current_step % opt["logger"]["save_checkpoint_freq"] == 0: if rank <= 0: # Save the experiments in case of Colab is timeout logger.info("Saving models and training states.") model.save(current_step) model.save_training_state(epoch, current_step) copy_tree( "/content/EDVR/experiments", "/content/drive/My Drive/LVTN/SuperResolution/EDVR/experiments", ) copy_tree( "/content/EDVR/tb_logger", "/content/drive/My Drive/LVTN/SuperResolution/EDVR/tb_logger", ) if rank <= 0: logger.info("Saving the final model.") model.save("latest") logger.info("End of training.") tb_logger.close()
def main(): # options parser = argparse.ArgumentParser() parser.add_argument("-opt", type=str, required=True, help="Path to option JSON file.") opt = option.parse(parser.parse_args().opt, is_train=True) opt = option.dict_to_nonedict( opt) # Convert to NoneDict, which return None for missing key. # train from scratch OR resume training if opt["path"]["resume_state"]: # resuming training resume_state = torch.load(opt["path"]["resume_state"]) else: # training from scratch resume_state = None util.mkdir_and_rename( opt["path"]["experiments_root"]) # rename old folder if exists util.mkdirs((path for key, path in opt["path"].items() if not key == "experiments_root" and "pretrain_model" not in key and "resume" not in key)) # config loggers. Before it, the log will not work util.setup_logger(None, opt["path"]["log"], "train", level=logging.INFO, screen=True) util.setup_logger("val", opt["path"]["log"], "val", level=logging.INFO) logger = logging.getLogger("base") if resume_state: logger.info("Resuming training from epoch: {}, iter: {}.".format( resume_state["epoch"], resume_state["iter"])) option.check_resume(opt) # check resume options logger.info(option.dict2str(opt)) # tensorboard logger if opt["use_tb_logger"] and "debug" not in opt["name"]: from tensorboardX import SummaryWriter tb_logger = SummaryWriter(log_dir="../tb_logger/" + opt["name"]) # random seed seed = opt["train"]["manual_seed"] if seed is None: seed = random.randint(1, 10000) logger.info("Random seed: {}".format(seed)) util.set_random_seed(seed) torch.backends.cudnn.benckmark = True # torch.backends.cudnn.deterministic = True # create train and val dataloader for phase, dataset_opt in opt["datasets"].items(): if phase == "train": train_set = create_dataset(dataset_opt) train_size = int( math.ceil(len(train_set) / dataset_opt["batch_size"])) logger.info("Number of train images: {:,d}, iters: {:,d}".format( len(train_set), train_size)) total_iters = int(opt["train"]["niter"]) total_epochs = int(math.ceil(total_iters / train_size)) logger.info("Total epochs needed: {:d} for iters {:,d}".format( total_epochs, total_iters)) train_loader = create_dataloader(train_set, dataset_opt) elif phase == "val": val_set = create_dataset(dataset_opt) val_loader = create_dataloader(val_set, dataset_opt) logger.info("Number of val images in [{:s}]: {:d}".format( dataset_opt["name"], len(val_set))) else: raise NotImplementedError( "Phase [{:s}] is not recognized.".format(phase)) assert train_loader is not None # create model model = create_model(opt) # resume training if resume_state: start_epoch = resume_state["epoch"] current_step = resume_state["iter"] model.resume_training(resume_state) # handle optimizers and schedulers else: current_step = 0 start_epoch = 0 # training logger.info("Start training from epoch: {:d}, iter: {:d}".format( start_epoch, current_step)) for epoch in range(start_epoch, total_epochs): for _, train_data in enumerate(train_loader): current_step += 1 if current_step > total_iters: break # update learning rate model.update_learning_rate() # training model.feed_data(train_data) model.optimize_parameters(current_step) # log if current_step % opt["logger"]["print_freq"] == 0: logs = model.get_current_log() message = "<epoch:{:3d}, iter:{:8,d}, lr:{:.3e}> ".format( epoch, current_step, model.get_current_learning_rate()) for k, v in logs.items(): message += "{:s}: {:.4e} ".format(k, v) # tensorboard logger if opt["use_tb_logger"] and "debug" not in opt["name"]: tb_logger.add_scalar(k, v, current_step) logger.info(message) # validation if current_step % opt["train"]["val_freq"] == 0: avg_psnr = 0.0 idx = 0 for val_data in val_loader: idx += 1 img_name = os.path.splitext( os.path.basename(val_data["LR_path"][0]))[0] img_dir = os.path.join(opt["path"]["val_images"], img_name) util.mkdir(img_dir) model.feed_data(val_data) model.test() visuals = model.get_current_visuals() sr_img = util.tensor2img(visuals["SR"]) # uint8 gt_img = util.tensor2img(visuals["HR"]) # uint8 # Save SR images for reference save_img_path = os.path.join( img_dir, "{:s}_{:d}.png".format(img_name, current_step)) util.save_img(sr_img, save_img_path) # calculate PSNR crop_size = opt["scale"] gt_img = gt_img / 255.0 sr_img = sr_img / 255.0 cropped_sr_img = sr_img[crop_size:-crop_size, crop_size:-crop_size, :] cropped_gt_img = gt_img[crop_size:-crop_size, crop_size:-crop_size, :] avg_psnr += util.calculate_psnr(cropped_sr_img * 255, cropped_gt_img * 255) avg_psnr = avg_psnr / idx # log logger.info("# Validation # PSNR: {:.4e}".format(avg_psnr)) logger_val = logging.getLogger("val") # validation logger logger_val.info( "<epoch:{:3d}, iter:{:8,d}> psnr: {:.4e}".format( epoch, current_step, avg_psnr)) # tensorboard logger if opt["use_tb_logger"] and "debug" not in opt["name"]: tb_logger.add_scalar("psnr", avg_psnr, current_step) # save models and training states if current_step % opt["logger"]["save_checkpoint_freq"] == 0: logger.info("Saving models and training states.") model.save(current_step) model.save_training_state(epoch, current_step) copy_tree( opt["path"]["experiments_root"], "/content/gdrive/My Drive/LVTN/SuperResolution/SR_models/" + "-ESRGAN/experiments/" + opt["name"], ) logger.info("Saving the final model.") model.save("latest") logger.info("End of training.")
def main(): # options parser = argparse.ArgumentParser() parser.add_argument('-opt', type=str, default='train_seperate_denoiser_diffuse.json', help='Path to option JSON file.') opt = option.parse(parser.parse_args().opt, is_train=True) #打开--opt指定的JSON file opt = option.dict_to_nonedict( opt) # Convert to NoneDict, which return None for missing key. # train from scratch OR resume training if opt['path']['resume_state']: # resuming training resume_state = torch.load(opt['path']['resume_state']) util.mkdir_and_rename( opt['path']['experiments_root']) # rename old folder if exists util.mkdirs((path for key, path in opt['path'].items() if not key == 'experiments_root' and 'pretrain_model' not in key and 'resume' not in key)) else: # training from scratch resume_state = None util.mkdir_and_rename( opt['path']['experiments_root']) # rename old folder if exists util.mkdirs((path for key, path in opt['path'].items() if not key == 'experiments_root' and 'pretrain_model' not in key and 'resume' not in key)) # config loggers. Before it, the log will not work util.setup_logger(None, opt['path']['log'], 'train', level=logging.INFO, screen=True) util.setup_logger('val', opt['path']['log'], 'val', level=logging.INFO) logger = logging.getLogger('base') if resume_state: logger.info('Resuming training from epoch: {}, iter: {}.'.format( resume_state['epoch'], resume_state['iter'])) option.check_resume(opt) # check resume options logger.info(option.dict2str(opt)) # tensorboard logger if opt['use_tb_logger'] and 'debug' not in opt['name']: from tensorboardX import SummaryWriter tb_logger = SummaryWriter(log_dir='../tb_logger/' + opt['name']) # random seed seed = opt['train']['manual_seed'] if seed is None: seed = random.randint(1, 10000) logger.info('Random seed: {}'.format(seed)) util.set_random_seed(seed) torch.backends.cudnn.benckmark = True for phase, dataset_opt in opt['datasets'].items(): if phase == 'train': train_set = create_dataset(dataset_opt) #load 图片那边还需要改!!!! train_size = int( math.ceil(len(train_set) / dataset_opt['batch_size'])) logger.info('Number of train images: {:,d}, iters: {:,d}'.format( len(train_set), train_size)) total_iters = int(opt['train']['niter']) total_epochs = int(math.ceil(total_iters / train_size)) logger.info('Total epochs needed: {:d} for iters {:,d}'.format( total_epochs, total_iters)) train_loader = create_dataloader(train_set, dataset_opt) elif phase == 'val': val_set = create_dataset(dataset_opt) val_loader = create_dataloader(val_set, dataset_opt) logger.info('Number of val images in [{:s}]: {:d}'.format( dataset_opt['name'], len(val_set))) else: raise NotImplementedError( 'Phase [{:s}] is not recognized.'.format(phase)) assert train_loader is not None model = create_model(opt) # resume training if resume_state: start_epoch = resume_state['epoch'] current_step = resume_state['iter'] model.resume_training(resume_state) # handle optimizers and schedulers else: current_step = 0 start_epoch = 0 logger.info('Start training from epoch: {:d}, iter: {:d}'.format( start_epoch, current_step)) for epoch in range(start_epoch, total_epochs): for _, train_data in enumerate(train_loader): current_step += 1 if current_step > total_iters: break # update learning rate model.update_learning_rate() # training model.feed_data_diffuse(train_data) model.optimize_parameters(current_step) # log if current_step % opt['logger']['print_freq'] == 0: logs = model.get_current_log() message = '<epoch:{:3d}, iter:{:8,d}, lr:{:.3e}> '.format( epoch, current_step, model.get_current_learning_rate()) for k, v in logs.items(): message += '{:s}: {:.4e} '.format(k, v) # tensorboard logger if opt['use_tb_logger'] and 'debug' not in opt['name']: tb_logger.add_scalar(k, v, current_step) logger.info(message) if current_step % opt['train']['val_freq'] == 0: print("output") idx = 0 for val_data in val_loader: idx += 1 model.feed_data_diffuse(val_data) model.test() visuals = model.get_current_visuals() # print(visuals['feature'].shape) lr_img = util.tensor2img(visuals['NOISY'], normalize=False) sr_img = util.tensor2img(visuals['DENOISED'], normalize=False) # uint8 gt_img = util.tensor2img(visuals['GT'], normalize=False) # uint8 fea_img = util.tensor2img(visuals['feature'], normalize=False) # uint8 # print(fea_img.shape) albedo = fea_img[:, :, :3] normal = fea_img[:, :, 3:6] depth = fea_img[:, :, 6:9] # print(albedo.shape) #img_name = os.path.splitext(os.path.basename(val_data['NOISY_path'][0]))[0] img_dir = os.path.join(opt['path']['root'], 'val_images') save_DENOISED_img_path = os.path.join( img_dir, '{:d}_{:d}_1denoised.png'.format(idx, current_step)) save_NOISY_img_path = os.path.join( img_dir, '{:d}_{:d}_0noisy.png'.format(idx, current_step)) save_GT_img_path = os.path.join( img_dir, '{:d}_{:d}_2gt.png'.format(idx, current_step)) save_Fea_img_path = os.path.join( img_dir, '{:d}_{:d}_3fea.png'.format(idx, current_step)) output = np.concatenate((albedo, normal, depth), axis=1) #把他们拼起来 util.save_img(sr_img, save_DENOISED_img_path) util.save_img(gt_img, save_GT_img_path) util.save_img(lr_img, save_NOISY_img_path) util.save_img(output, save_Fea_img_path) # validation # ============================================================================= # if current_step % opt['train']['val_freq'] == 0: # avg_psnr = 0.0 # avg_ssim = 0.0 # avg_mrse = 0.0 # idx = 0 # for val_data in val_loader: # idx += 1 # img_name = os.path.splitext(os.path.basename(val_data['NOISY_path'][0]))[0] # img_dir = opt['path']['val_images'] #os.path.join(opt['path']['val_images'], img_name) # util.mkdir(img_dir) # # model.feed_data_diffuse(val_data) # model.test() # if opt["image_type"] == "exr": # y = val_data["x_offset"] # x = val_data["y_offset"] # visuals = model.get_current_visuals() # avg_mrse += util.calculate_mrse(visuals["DENOISED"].numpy(), visuals["GT"].numpy()) # lr_img = util.tensor2img(visuals['NOISY']) # sr_img = util.tensor2img(visuals['DENOISED']) # uint8 # gt_img = util.tensor2img(visuals['GT']) # uint8 # # ############################################################################################## # # sr_img = util.tensor2img(visuals['DENOISED']) # uint8 # # lr_img = util.tensor2img(visuals['NOISY']) # # gt_img = util.tensor2img(visuals['GT']) # uint8 # # # if opt["image_type"] == "exr": # # sr_img = sr_img[y:1280-y, x:1280-x, :] # # lr_img = lr_img[y:1280-y, x:1280-x, :] # # gt_img = gt_img[y:1280-y, x:1280-x, :] # # # ############################################################################################## # # # # Save DENOISED images for reference # save_DENOISED_img_path = os.path.join(img_dir, '{:s}_{:d}_1denoised.png'.format(img_name, current_step)) # save_NOISY_img_path = os.path.join(img_dir, '{:s}_{:d}_0noisy.png'.format(img_name, current_step)) # save_GT_img_path = os.path.join(img_dir, '{:s}_{:d}_2gt.png'.format(img_name, current_step)) # # if current_step % 10000 == 0 :#and idx%100 ==0: # # util.save_img(sr_img, save_DENOISED_img_path) # # util.save_img(lr_img, save_NOISY_img_path) # # util.save_img(gt_img, save_GT_img_path) # # # calculate PSNR # # crop_size = opt['scale'] # gt_img = gt_img #/ 255. # sr_img = sr_img #/ 255. # # cropped_sr_img = sr_img[crop_size:-crop_size, crop_size:-crop_size, :] # # cropped_gt_img = gt_img[crop_size:-crop_size, crop_size:-crop_size, :] # # avg_psnr += util.calculate_psnr(sr_img * 255, gt_img * 255) # avg_psnr += util.calculate_psnr(sr_img , gt_img ) # avg_ssim += util.calculate_ssim(sr_img , gt_img) # # ############################################################################################## # # if opt["image_type"] == "exr" and current_step %10000 == 0: # sr_exr = util.tensor2exr(visuals['DENOISED']) # uint8 # lr_exr = util.tensor2exr(visuals['NOISY']) # gt_exr = util.tensor2exr(visuals['GT']) # uint8 # # # sr_exr = sr_exr[y:1280-y, x:1280-x, :] # # lr_exr = lr_exr[y:1280-y, x:1280-x, :] # # gt_exr = gt_exr[y:1280-y, x:1280-x, :] # save_DENOISED_img_path = os.path.join(img_dir, '{:s}_{:d}_1denoised.exr'.format(img_name, current_step)) # save_NOISY_img_path = os.path.join(img_dir, '{:s}_{:d}_0noisy.exr'.format(img_name, current_step)) # save_GT_img_path = os.path.join(img_dir, '{:s}_{:d}_2gt.exr'.format(img_name, current_step)) # # util.saveEXRfromMatrix(save_DENOISED_img_path, sr_exr, (x, y)) # util.saveEXRfromMatrix(save_NOISY_img_path, lr_exr, (x, y)) # util.saveEXRfromMatrix(save_GT_img_path, gt_exr, (x, y)) # # ############################################################################################## # # avg_psnr = avg_psnr / idx # avg_ssim = avg_ssim / idx # avg_mrse = avg_mrse / idx # # # log # logger.info('# Validation # PSNR: {:.4e}'.format(avg_psnr)) # logger.info('# Validation # SSIM: {:.4e}'.format(avg_ssim)) # logger.info('# Validation # MRSE: {:.4e}'.format(avg_mrse)) # logger_val = logging.getLogger('val') # validation logger # logger_val.info('<epoch:{:3d}, iter:{:8,d}> psnr: {:.4e} ssim: {:.4e} mrse: {:.4e}'.format( # epoch, current_step, avg_psnr, avg_ssim, avg_mrse)) # # tensorboard logger # if opt['use_tb_logger'] and 'debug' not in opt['name']: # tb_logger.add_scalar('psnr', avg_psnr, current_step) # tb_logger.add_scalar('ssim', avg_ssim, current_step) # tb_logger.add_scalar('mrse', avg_mrse, current_step) # ============================================================================= # save models and training states if current_step % opt['logger']['save_checkpoint_freq'] == 0: logger.info('Saving models and training states.') model.save(current_step) model.save_training_state(epoch, current_step) logger.info('Saving the final model.') model.save('latest') logger.info('End of training.')
def main(): # options parser = argparse.ArgumentParser() parser.add_argument('-opt', type=str, required=True, help='Path to option JSON file.') opt = option.parse(parser.parse_args().opt, is_train=True) util.mkdir_and_rename( opt['path']['experiments_root']) # rename old experiments if exists util.mkdirs((path for key, path in opt['path'].items() if not key == 'experiments_root' and \ not key == 'pretrain_model_G' and not key == 'pretrain_model_D')) option.save(opt) opt = option.dict_to_nonedict( opt) # Convert to NoneDict, which return None for missing key. # print to file and std_out simultaneously sys.stdout = PrintLogger(opt['path']['log']) # random seed seed = opt['train']['manual_seed'] if seed is None: seed = random.randint(1, 10000) print("Random Seed: ", seed) random.seed(seed) torch.manual_seed(seed) # create train and val dataloader for phase, dataset_opt in opt['datasets'].items(): if phase == 'train': train_set = create_dataset(dataset_opt) train_size = int( math.ceil(len(train_set) / dataset_opt['batch_size'])) print('Number of train images: {:,d}, iters: {:,d}'.format( len(train_set), train_size)) total_iters = int(opt['train']['niter']) total_epoches = int(math.ceil(total_iters / train_size)) print('Total epoches needed: {:d} for iters {:,d}'.format( total_epoches, total_iters)) train_loader = create_dataloader(train_set, dataset_opt) batch_size_per_month = dataset_opt['batch_size'] batch_size_per_day = int( opt['datasets']['train']['batch_size_per_day']) num_month = int(opt['train']['num_month']) num_day = int(opt['train']['num_day']) elif phase == 'val': val_dataset_opt = dataset_opt val_set = create_dataset(dataset_opt) val_loader = create_dataloader(val_set, dataset_opt) print('Number of val images in [{:s}]: {:d}'.format( dataset_opt['name'], len(val_set))) else: raise NotImplementedError( 'Phase [{:s}] is not recognized.'.format(phase)) assert train_loader is not None # Create model model = create_model(opt) # create logger logger = Logger(opt) current_step = 0 start_time = time.time() print('---------- Start training -------------') validate(val_loader, opt, model, current_step, 0, logger) for epoch in range(num_month): for i, train_data in enumerate(train_loader): cur_month_code = get_code_for_data_two(model, train_data, opt) for j in range(num_day): current_step += 1 # get the sliced data cur_day_batch_start_idx = ( j * batch_size_per_day) % batch_size_per_month cur_day_batch_end_idx = cur_day_batch_start_idx + batch_size_per_day if cur_day_batch_end_idx > batch_size_per_month: cur_day_batch_idx = np.hstack( (np.arange(cur_day_batch_start_idx, batch_size_per_month), np.arange(cur_day_batch_end_idx - batch_size_per_month))) else: cur_day_batch_idx = slice(cur_day_batch_start_idx, cur_day_batch_end_idx) cur_day_train_data = { 'LR': train_data['LR'][cur_day_batch_idx], 'HR': train_data['HR'][cur_day_batch_idx] } code = [] for gen_code in cur_month_code: code.append(gen_code[cur_day_batch_idx]) # training model.feed_data(cur_day_train_data, code=code) model.optimize_parameters(current_step) time_elapsed = time.time() - start_time start_time = time.time() # log if current_step % opt['logger']['print_freq'] == 0: logs = model.get_current_log() print_rlt = OrderedDict() print_rlt['model'] = opt['model'] print_rlt['epoch'] = epoch print_rlt['iters'] = current_step print_rlt['time'] = time_elapsed for k, v in logs.items(): print_rlt[k] = v print_rlt['lr'] = model.get_current_learning_rate() logger.print_format_results('train', print_rlt) # save models if current_step % opt['logger']['save_checkpoint_freq'] == 0: print('Saving the model at the end of iter {:d}.'.format( current_step)) model.save(current_step) # validation if current_step % opt['train']['val_freq'] == 0: validate(val_loader, opt, model, current_step, epoch, logger) # update learning rate model.update_learning_rate() print('Saving the final model.') model.save('latest') print('End of training.')
def main(): #### options parser = argparse.ArgumentParser() parser.add_argument('-opt', type=str, default='options/train/train_EDVR_woTSA_M.yml', help='Path to option YAML file.') parser.add_argument('--set', dest='set_opt', default=None, nargs=argparse.REMAINDER, help='set options') args = parser.parse_args() opt = option.parse(args.opt, args.set_opt, is_train=True) #### loading resume state if exists if opt['path'].get('resume_state', None): # distributed resuming: all load into default GPU print('Training from state: {}'.format(opt['path']['resume_state'])) device_id = torch.cuda.current_device() resume_state = torch.load( opt['path']['resume_state'], map_location=lambda storage, loc: storage.cuda(device_id)) option.check_resume(opt, resume_state['iter']) # check resume options elif opt['auto_resume']: exp_dir = opt['path']['experiments_root'] # first time run: create dirs if not os.path.exists(exp_dir): os.makedirs(exp_dir) os.makedirs(opt['path']['models']) os.makedirs(opt['path']['training_state']) os.makedirs(opt['path']['val_images']) os.makedirs(opt['path']['tb_logger']) resume_state = None else: # detect experiment directory and get the latest state state_dir = opt['path']['training_state'] state_files = [ x for x in os.listdir(state_dir) if x.endswith('state') ] # no valid state detected if len(state_files) < 1: print( 'No previous training state found, train from start state') resume_state = None else: state_files = sorted(state_files, key=lambda x: int(x.split('.')[0])) latest_state = state_files[-1] print('Training from lastest state: {}'.format(latest_state)) latest_state_file = os.path.join(state_dir, latest_state) opt['path']['resume_state'] = latest_state_file device_id = torch.cuda.current_device() resume_state = torch.load( latest_state_file, map_location=lambda storage, loc: storage.cuda(device_id)) option.check_resume(opt, resume_state['iter']) else: resume_state = None if resume_state is None and not opt['auto_resume'] and not opt['no_log']: util.mkdir_and_rename( opt['path'] ['experiments_root']) # rename experiment folder if exists util.mkdirs((path for key, path in opt['path'].items() if not key == 'experiments_root' and 'pretrain_model' not in key and 'resume' not in key)) # config loggers. Before it, the log will not work util.setup_logger('base', opt['path']['log'], 'train_' + opt['name'], level=logging.INFO, screen=True, tofile=True) logger = logging.getLogger('base') logger.info(option.dict2str(opt)) # tensorboard logger if opt['use_tb_logger'] and 'debug' not in opt['name']: version = float(torch.__version__[0:3]) if version >= 1.2: # PyTorch 1.1 from torch.utils.tensorboard import SummaryWriter else: logger.info( 'You are using PyTorch {}. Tensorboard will use [tensorboardX]' .format(version)) from tensorboardX import SummaryWriter tb_logger = SummaryWriter(log_dir=opt['path']['tb_logger']) # convert to NoneDict, which returns None for missing keys opt = option.dict_to_nonedict(opt) #### random seed seed = opt['train']['manual_seed'] if seed is None: seed = random.randint(1, 10000) logger.info('Random seed: {}'.format(seed)) util.set_random_seed(seed) torch.backends.cudnn.benchmark = True # torch.backends.cudnn.deterministic = True #### create train and val dataloader if opt['datasets']['train']['ratio']: dataset_ratio = opt['datasets']['train']['ratio'] else: dataset_ratio = 200 # enlarge the size of each epoch for phase, dataset_opt in opt['datasets'].items(): if phase == 'train': train_set = create_dataset(dataset_opt) train_size = int( math.ceil(len(train_set) / dataset_opt['batch_size'])) total_iters = int(opt['train']['niter']) total_epochs = int( math.ceil(total_iters / (train_size * dataset_ratio))) if dataset_opt['mode'] in ['MetaREDS', 'MetaREDSOnline']: train_sampler = MetaIterSampler(train_set, dataset_opt['batch_size'], len(opt['scale']), dataset_ratio) elif dataset_opt['mode'] in ['REDS', 'MultiREDS']: train_sampler = IterSampler(train_set, dataset_opt['batch_size'], dataset_ratio) else: train_sampler = None train_loader = create_dataloader(train_set, dataset_opt, opt, train_sampler) logger.info('Number of train images: {:,d}, iters: {:,d}'.format( len(train_set), train_size)) logger.info('Total epochs needed: {:d} for iters {:,d}'.format( total_epochs, total_iters)) elif phase == 'val': val_set = create_dataset(dataset_opt) val_loader = create_dataloader(val_set, dataset_opt, opt, None) logger.info('Number of val images in [{:s}]: {:d}'.format( dataset_opt['name'], len(val_set))) else: raise NotImplementedError( 'Phase [{:s}] is not recognized.'.format(phase)) #### create model model = create_model(opt) #### resume training if resume_state: logger.info('Resuming training from epoch: {}, iter: {}.'.format( resume_state['epoch'], resume_state['iter'])) start_epoch = resume_state['epoch'] current_step = resume_state['iter'] model.resume_training(resume_state) # handle optimizers and schedulers else: current_step = 0 start_epoch = 0 #### training logger.info('Start training from epoch: {:d}, iter: {:d}'.format( start_epoch, current_step)) for epoch in range(start_epoch, total_epochs + 1): train_sampler.set_epoch(epoch) for _, train_data in enumerate(train_loader): current_step += 1 if current_step > total_iters: break #### update learning rate model.update_learning_rate(current_step, warmup_iter=opt['train']['warmup_iter']) #### training model.feed_data(train_data) model.optimize_parameters(current_step) #### log if current_step % opt['logger']['print_freq'] == 0: logs = model.get_current_log() message = '[epoch:{:3d}, iter:{:8,d}, lr:('.format( epoch, current_step) for v in model.get_current_learning_rate(): message += '{:.3e},'.format(v) message += ')] ' for k, v in logs.items(): message += '{:s}: {:.4e} '.format(k, v) # tensorboard logger if opt['use_tb_logger'] and 'debug' not in opt['name']: tb_logger.add_scalar(k, v, current_step) logger.info(message) print("PROGRESS: {:02d}%".format( int(current_step / total_iters * 100))) #### validation if opt['datasets'].get( 'val', None) and current_step % opt['train']['val_freq'] == 0: pbar = util.ProgressBar(len(val_loader)) psnr_rlt = {} # with border and center frames psnr_rlt_avg = {} psnr_total_avg = 0. for val_data in val_loader: folder = val_data['folder'][0] idx_d = val_data['idx'].item() # border = val_data['border'].item() if psnr_rlt.get(folder, None) is None: psnr_rlt[folder] = [] model.feed_data(val_data) model.test() visuals = model.get_current_visuals() rlt_img = util.tensor2img(visuals['rlt']) # uint8 gt_img = util.tensor2img(visuals['GT']) # uint8 # calculate PSNR psnr = util.calculate_psnr(rlt_img, gt_img) psnr_rlt[folder].append(psnr) pbar.update('Test {} - {}'.format(folder, idx_d)) for k, v in psnr_rlt.items(): psnr_rlt_avg[k] = sum(v) / len(v) psnr_total_avg += psnr_rlt_avg[k] psnr_total_avg /= len(psnr_rlt) log_s = '# Validation # PSNR: {:.4e}:'.format(psnr_total_avg) for k, v in psnr_rlt_avg.items(): log_s += ' {}: {:.4e}'.format(k, v) logger.info(log_s) if opt['use_tb_logger'] and 'debug' not in opt['name']: tb_logger.add_scalar('psnr_avg', psnr_total_avg, current_step) for k, v in psnr_rlt_avg.items(): tb_logger.add_scalar(k, v, current_step) #### save models and training states if current_step % opt['logger']['save_checkpoint_freq'] == 0: logger.info('Saving models and training states.') model.save(current_step) model.save_training_state(epoch, current_step) logger.info('Saving the final model.') model.save('latest') logger.info('End of training.') tb_logger.close()
def main(): PreUp = False # options parser = argparse.ArgumentParser() parser.add_argument('-opt', type=str, required=True, help='Path to option JSON file.') opt = option.parse(parser.parse_args().opt, is_train=True) opt = option.dict_to_nonedict(opt) # Convert to NoneDict, which return None for missing key. ratio = opt["scale"] if PreUp == True: ratio=5 # train from scratch OR resume training if opt['path']['resume_state']: # resuming training resume_state = torch.load(opt['path']['resume_state']) else: # training from scratch resume_state = None util.mkdir_and_rename(opt['path']['experiments_root']) # rename old folder if exists util.mkdirs((path for key, path in opt['path'].items() if not key == 'experiments_root' and 'pretrain_model' not in key and 'resume' not in key)) # config loggers. Before it, the log will not work util.setup_logger(None, opt['path']['log'], 'train', level=logging.INFO, screen=True) util.setup_logger('val', opt['path']['log'], 'val', level=logging.INFO) logger = logging.getLogger('base') if resume_state: logger.info('Resuming training from epoch: {}, iter: {}.'.format( resume_state['epoch'], resume_state['iter'])) option.check_resume(opt) # check resume options logger.info(option.dict2str(opt)) # tensorboard logger if opt['use_tb_logger'] and 'debug' not in opt['name']: from tensorboardX import SummaryWriter tb_logger_train = SummaryWriter(log_dir='/mnt/gpid07/users/luis.salgueiro/git/mnt/BasicSR/tb_logger/' + opt['name'] + "/train") tb_logger_val = SummaryWriter(log_dir='//mnt/gpid07/users/luis.salgueiro/git/mnt/BasicSR/tb_logger/' + opt['name'] + "/val" ) # random seed seed = opt['train']['manual_seed'] if seed is None: seed = 100 #random.randint(1, 10000) logger.info('Random seed: {}'.format(seed)) util.set_random_seed(seed) torch.backends.cudnn.benckmark = True # torch.backends.cudnn.deterministic = True # print("OLAAAA_-...", os.environ['CUDA_VISIBLE_DEVICES']) # ######################################### # ######## DATA LOADER #################### # ######################################### # create train and val dataloader for phase, dataset_opt in opt['datasets'].items(): if phase == 'train': print("Entro DATASET train......") train_set = create_dataset(dataset_opt) print("CREO DATASET train_set ", train_set) train_size = int(math.ceil(len(train_set) / dataset_opt['batch_size'])) logger.info('Number of train images: {:,d}, iters: {:,d}'.format( len(train_set), train_size)) total_iters = int(opt['train']['niter']) total_epochs = int(math.ceil(total_iters / train_size)) logger.info('Total epochs needed: {:d} for iters {:,d}'.format( total_epochs, total_iters)) train_loader = create_dataloader(train_set, dataset_opt) print("CREO train loader: ", train_loader) elif phase == 'val': print("Entro en phase VAL....") val_set = create_dataset(dataset_opt) val_loader = create_dataloader(val_set, dataset_opt) logger.info('Number of val images in [{:s}]: {:d}'.format(dataset_opt['name'], len(val_set))) # for _,ii in enumerate(val_loader): # print("VAL LOADER:........", ii) # print(val_loader[0]) else: raise NotImplementedError('Phase [{:s}] is not recognized.'.format(phase)) assert train_loader is not None assert val_loader is not None # create model model = create_model(opt) #print("PASO..... MODEL ") # resume training if resume_state: print("RESUMING state") start_epoch = resume_state['epoch'] current_step = resume_state['iter'] model.resume_training(resume_state) # handle optimizers and schedulers else: current_step = 0 start_epoch = 0 print("PASO..... INIT ") # ######################################### # ######### training ################ # ######################################### # ii=0 logger.info('Start training from epoch: {:d}, iter: {:d}'.format(start_epoch, current_step)) for epoch in range(start_epoch, total_epochs): # print("Entro EPOCH...", ii) for _, train_data in enumerate(train_loader): # print("Entro TRAIN_LOADER...") current_step += 1 if current_step > total_iters: break # update learning rate model.update_learning_rate() # training #print("....... TRAIN DATA..........", train_data) model.feed_data(train_data) model.optimize_parameters(current_step) # log train if current_step % opt['logger']['print_freq'] == 0: logs = model.get_current_log() message = '<epoch:{:3d}, iter:{:8,d}, lr:{:.3e}> '.format( epoch, current_step, model.get_current_learning_rate()) # print(".............MESSAGE: ", message) for k, v in logs.items(): message += '{:s}: {:.4e} '.format(k, v) #print("MSG: ", message) # tensorboard logger if opt['use_tb_logger'] and 'debug' not in opt['name']: # print("K: ", k) # print("V: ", v) if "test" in k: tb_logger_val.add_scalar(k, v, current_step) else: tb_logger_train.add_scalar(k, v, current_step) logger.info(message) if current_step % opt['train']['val_freq'] == 0: avg_psnr_sr = 0.0 avg_psnr_lr = 0.0 avg_psnr_dif = 0.0 avg_ssim_lr, avg_ssim_sr, avg_ssim_dif = 0.0, 0.0, 0.0 avg_ergas_lr, avg_ergas_sr, avg_ergas_dif = 0.0, 0.0, 0.0 idx = 0 # for val_data in val_loader: for _, val_data in enumerate(val_loader): idx += 1 img_name = os.path.splitext(os.path.basename(val_data['LR_path'][0]))[0] img_dir = os.path.join(opt['path']['val_images'], img_name) # print("Img nameVaL: ", img_name) model.feed_data(val_data) model.test() visuals = model.get_current_visuals() sr_img = util.tensor2imgNorm(visuals['SR'],out_type=np.uint8, min_max=(0, 1), MinVal=val_data["LR_min"], MaxVal=val_data["LR_max"]) # uint16 gt_img = util.tensor2imgNorm(visuals['HR'],out_type=np.uint8, min_max=(0, 1), MinVal=val_data["HR_min"], MaxVal=val_data["HR_max"]) # uint16 lr_img = util.tensor2imgNorm(visuals['LR'], out_type=np.uint8, min_max=(0, 1), MinVal=val_data["LR_min"], MaxVal=val_data["LR_max"]) # uint16 # Save SR images for reference if idx < 10: # print(idx) util.mkdir(img_dir) save_img_path = os.path.join(img_dir, '{:s}_{:d}'.format(img_name, current_step)) util.save_imgSR(sr_img, save_img_path) util.save_imgHR(gt_img, save_img_path) util.save_imgLR(lr_img, save_img_path) print("SAVING CROPS") util.save_imgCROP(lr_img,gt_img,sr_img , save_img_path, ratio, PreUp=PreUp) if PreUp==False: dim2 = (gt_img.shape[1], gt_img.shape[1]) print("DIM:", dim2) print("LR image shape ", lr_img.shape) print("HR image shape ", gt_img.shape) lr_img = cv2.resize(np.transpose(lr_img,(1,2,0)), dim2, interpolation=cv2.INTER_NEAREST) lr_img = np.transpose(lr_img,(2,0,1)) print("LR image 2 shape ", lr_img.shape) print("LR image 2 shape ", lr_img.shape) avg_psnr_sr += util.calculate_psnr2(sr_img, gt_img) avg_psnr_lr += util.calculate_psnr2(lr_img, gt_img) avg_ssim_lr += util.calculate_ssim2(lr_img, gt_img) avg_ssim_sr += util.calculate_ssim2(sr_img, gt_img) avg_ergas_lr += util.calculate_ergas(lr_img, gt_img, pixratio=ratio) avg_ergas_sr += util.calculate_ergas(sr_img, gt_img, pixratio=ratio) #avg_psnr += util.calculate_psnr2(cropped_sr_img, cropped_gt_img) avg_psnr_sr = avg_psnr_sr / idx avg_psnr_lr = avg_psnr_lr / idx avg_psnr_dif = avg_psnr_lr - avg_psnr_sr avg_ssim_lr = avg_ssim_lr / idx avg_ssim_sr = avg_ssim_sr / idx avg_ssim_dif = avg_ssim_lr - avg_ssim_sr avg_ergas_lr = avg_ergas_lr / idx avg_ergas_sr = avg_ergas_sr / idx avg_ergas_dif = avg_ergas_lr - avg_ergas_sr # print("IDX: ", idx) # log VALIDATION logger.info('# Validation # PSNR: {:.4e}'.format(avg_psnr_sr)) logger.info('# Validation # SSIM: {:.4e}'.format(avg_ssim_sr)) logger.info('# Validation # ERGAS: {:.4e}'.format(avg_ergas_sr)) logger_val = logging.getLogger('val') # validation logger logger_val.info('<epoch:{:3d}, iter:{:8,d}> psnr_SR: {:.4e}'.format( epoch, current_step, avg_psnr_sr)) logger_val.info('<epoch:{:3d}, iter:{:8,d}> psnr_LR: {:.4e}'.format( epoch, current_step, avg_psnr_lr)) logger_val.info('<epoch:{:3d}, iter:{:8,d}> psnr_DIF: {:.4e}'.format( epoch, current_step, avg_psnr_dif)) logger_val.info('<epoch:{:3d}, iter:{:8,d}> ssim_LR: {:.4e}'.format( epoch, current_step, avg_ssim_lr)) logger_val.info('<epoch:{:3d}, iter:{:8,d}> ssim_SR: {:.4e}'.format( epoch, current_step, avg_ssim_sr)) logger_val.info('<epoch:{:3d}, iter:{:8,d}> ssim_DIF: {:.4e}'.format( epoch, current_step, avg_ssim_dif)) logger_val.info('<epoch:{:3d}, iter:{:8,d}> ergas_LR: {:.4e}'.format( epoch, current_step, avg_ergas_lr)) logger_val.info('<epoch:{:3d}, iter:{:8,d}> ergas_SR: {:.4e}'.format( epoch, current_step, avg_ergas_sr)) logger_val.info('<epoch:{:3d}, iter:{:8,d}> ergas_DIF: {:.4e}'.format( epoch, current_step, avg_ergas_dif)) # tensorboard logger if opt['use_tb_logger'] and 'debug' not in opt['name']: tb_logger_val.add_scalar('dif_PSNR', avg_psnr_dif, current_step) # tb_logger.add_scalar('psnr', avg_psnr, current_step) tb_logger_val.add_scalar('dif_SSIM', avg_ssim_dif, current_step) tb_logger_val.add_scalar('dif_ERGAS', avg_ergas_dif, current_step) tb_logger_val.add_scalar('psnr_LR', avg_psnr_lr, current_step) # tb_logger.add_scalar('psnr', avg_psnr, current_step) tb_logger_val.add_scalar('ssim_LR', avg_ssim_lr, current_step) tb_logger_val.add_scalar('ERGAS_LR', avg_ergas_lr, current_step) tb_logger_val.add_scalar('psnr_SR', avg_psnr_sr, current_step) # tb_logger.add_scalar('psnr', avg_psnr, current_step) tb_logger_val.add_scalar('ssim_SR', avg_ssim_sr, current_step) tb_logger_val.add_scalar('ERGAS_SR', avg_ergas_sr, current_step) print("****** SR_IMG: ", sr_img.shape) print("****** LR_IMG: ", lr_img.shape) print("****** GT_IMG: ", gt_img.shape) fig1,ax1 = ep.plot_rgb(sr_img, rgb=[2, 1, 0], stretch=True) tb_logger_val.add_figure("SR_plt", fig1, current_step,close=True) fig2, ax2 = ep.plot_rgb(gt_img, rgb=[2, 1, 0], stretch=True) tb_logger_val.add_figure("GT_plt", fig2, current_step, close=True) fig3, ax3 = ep.plot_rgb(lr_img, rgb=[2, 1, 0], stretch=True) tb_logger_val.add_figure("LR_plt", fig3, current_step, close=True) # print("TERMINO GUARDAR IMG TB") # save models and training states if current_step % opt['logger']['save_checkpoint_freq'] == 0: logger.info('Saving models and training states.') model.save(current_step) model.save_training_state(epoch, current_step) # ii=ii+1 logger.info('Saving the final model.') model.save('latest') logger.info('End of training.')
def main(): ### parser ### diff : cannot support distribution parser: ArgumentParser = argparse.ArgumentParser() parser.add_argument("-opt", type=str, help="Path to YAML file.") args = parser.parse_args() opt: Dict[str, Any] = option.parse(args.opt, is_train=True) ### mkdir and loggers util.mkdir_and_rename( opt["path"]["experiments_root"] ) # rename experiment folder if exists util.mkdirs( ( path for key, path in opt["path"].items() if not key == "experiments_root" and "pretrain_model" not in key and "resume" not in key ) ) util.setup_logger( "base", PATH, "train_" + NAME, level=logging.INFO, screen=True, tofile=False ) util.setup_logger( "val", PATH, "val_" + NAME, level=logging.INFO, screen=True, tofile=False ) logger: Logger = logging.getLogger("base") opt = option.dict_to_nonedict(opt) # TODO : tensorboard logger ### random seed seed: int = MANUAL_SEED logger.info(f"Random seed: {seed}") util.set_random_seed(seed) # https://discuss.pytorch.org/t/what-does-torch-backends-cudnn-benchmark-do/5936/2 torch.backends.cudnn.benchmark = True ### create train and val dataloader phase: str dataset_opt: Dict[str, Any] for phase, dataset_opt in opt["datasets"].items(): if phase == "train": train_set: Union[LQGTDataset] = create_dataset(dataset_opt) train_size: int = int(math.ceil(len(train_set) / dataset_opt["batch_size"])) total_iters: int = int(opt["train"]["niter"]) total_epochs: int = int(math.ceil(total_iters / train_size)) train_loader = create_dataloader(train_set, dataset_opt, opt, None) logger.info( "Number of train images: {:,d}, iters: {:,d}".format( len(train_set), train_size ) ) logger.info( "Total epochs needed: {:d} for iters {:,d}".format( total_epochs, total_iters ) ) elif phase == "val": val_set = create_dataset(dataset_opt) val_loader = create_dataloader(val_set, dataset_opt, opt, None) else: raise NotImplementedError(f"Phase [{phase:s}] is not recognized") model = create_model(opt) current_step: int = 0 start_epoch: int = 0 # TODO : training logger.info(f"Start training from epoch: {start_epoch}, iter: {current_step}") # for epoch in range(start_epoch, total_epochs + 1): for epoch in range(start_epoch, total_epochs + 1): for _, train_data in enumerate(train_loader): current_step += 1 if current_step > total_iters: break ### training model.feed_data(train_data) model.optimize_parameters(current_step) model.update_learning_rate( current_step, warmup_iter=opt["train"]["warmup_iter"] ) if current_step % opt["logger"]["print_freq"] == 0: logs = model.get_current_log() message = f"<epoch:{epoch:3d}, iter:{current_step:8d}, lr:{model.get_current_learning_rate():3e}> " for k, v in logs.items(): message += f"{k:s}: {v:.4e} " # TODO: tensorboard logger.info(message) # validation if current_step % opt["train"]["val_freq"] == 0: avg_psnr: float = 0.0 idx: int = 0 for val_data in val_loader: idx += 1 img_name = os.path.splitext( os.path.basename(val_data["LQ_path"][0]) )[0] img_dir = os.path.join(opt["path"]["val_images"], img_name) util.mkdir(img_dir) model.feed_data(val_data) # TODO: fix if model.test() == False: continue visuals = model.get_current_visuals() sr_img = util.tensor2img(visuals["SR"]) # uint8 gt_img = util.tensor2img(visuals["GT"]) # uint8 lr_img = util.tensor2img(visuals["LR"]) gtl_img = util.tensor2img(visuals["LR_ref"]) # Save SR images for reference save_img_path = os.path.join( img_dir, "{:s}_{:d}.png".format(img_name, current_step) ) util.save_img(sr_img, save_img_path) # Save LR images save_img_path_L = os.path.join( img_dir, "{:s}_forwLR_{:d}.png".format(img_name, current_step) ) util.save_img(lr_img, save_img_path_L) # Save ground truth if current_step == opt["train"]["val_freq"]: save_img_path_gt = os.path.join( img_dir, "{:s}_GT_{:d}.png".format(img_name, current_step) ) util.save_img(gt_img, save_img_path_gt) save_img_path_gtl = os.path.join( img_dir, "{:s}_LR_ref_{:d}.png".format(img_name, current_step), ) util.save_img(gtl_img, save_img_path_gtl) # calculate PSNR crop_size = opt["scale"] gt_img = gt_img / 255.0 sr_img = sr_img / 255.0 cropped_sr_img = sr_img[ crop_size:-crop_size, crop_size:-crop_size, : ] cropped_gt_img = gt_img[ crop_size:-crop_size, crop_size:-crop_size, : ] avg_psnr += util.calculate_psnr( cropped_sr_img * 255, cropped_gt_img * 255 ) avg_psnr = avg_psnr / idx # log logger.info("# Validation # PSNR: {:.4e}.".format(avg_psnr)) logger_val = logging.getLogger("val") # validation logger logger_val.info( "<epoch:{:3d}, iter:{:8,d}> psnr: {:.4e}.".format( epoch, current_step, avg_psnr ) ) # TODO: tensorboard if current_step % opt["logger"]["save_checkpoint_freq"] == 0: logger.info("Saving models and training states.") model.save(current_step) model.save_training_state(epoch, current_step)
def main(): #### options parser = argparse.ArgumentParser() parser.add_argument('-opt', type=str, help='Path to option YMAL file.') parser.add_argument('--launcher', choices=['none', 'pytorch'], default='none', help='job launcher') parser.add_argument('--local_rank', type=int, default=0) args = parser.parse_args() opt = option.parse(args.opt, is_train=True) #### distributed training settings opt['dist'] = False rank = -1 print('Disabled distributed training.') #### loading resume state if exists if opt['path'].get('resume_state', None): resume_state_path, _ = get_resume_paths(opt) print('\n\t************************') print('resume_state_path: ', resume_state_path) print('\n\t************************') # distributed resuming: all load into default GPU if resume_state_path is None: resume_state = None else: device_id = torch.cuda.current_device() resume_state = torch.load( resume_state_path, map_location=lambda storage, loc: storage.cuda(device_id)) option.check_resume(opt, resume_state['iter']) # check resume options else: resume_state = None #### mkdir and loggers if rank <= 0: # normal training (rank -1) OR distributed training (rank 0) if resume_state is None: util.mkdir_and_rename( opt['path'] ['experiments_root']) # rename experiment folder if exists util.mkdirs( (path for key, path in opt['path'].items() if not key == 'experiments_root' and 'pretrain_model' not in key and 'resume' not in key)) # config loggers. Before it, the log will not work util.setup_logger('base', opt['path']['log'], 'train_' + opt['name'], level=logging.INFO, screen=True, tofile=True) util.setup_logger('val', opt['path']['log'], 'val_' + opt['name'], level=logging.INFO, screen=True, tofile=True) logger = logging.getLogger('base') logger.info(option.dict2str(opt)) # tensorboard logger if opt.get('use_tb_logger', False) and 'debug' not in opt['name']: version = float(torch.__version__[0:3]) if version >= 1.1: # PyTorch 1.1 from torch.utils.tensorboard import SummaryWriter else: logger.info( 'You are using PyTorch {}. Tensorboard will use [tensorboardX]' .format(version)) from tensorboardX import SummaryWriter conf_name = basename(args.opt).replace(".yml", "") exp_dir = opt['path']['experiments_root'] log_dir_train = os.path.join(exp_dir, 'tb', conf_name, 'train') log_dir_valid = os.path.join(exp_dir, 'tb', conf_name, 'valid') tb_logger_train = SummaryWriter(log_dir=log_dir_train) tb_logger_valid = SummaryWriter(log_dir=log_dir_valid) else: util.setup_logger('base', opt['path']['log'], 'train', level=logging.INFO, screen=True) logger = logging.getLogger('base') # convert to NoneDict, which returns None for missing keys opt = option.dict_to_nonedict(opt) #### random seed seed = opt['train']['manual_seed'] if seed is None: seed = random.randint(1, 10000) if rank <= 0: logger.info('Random seed: {}'.format(seed)) util.set_random_seed(seed) torch.backends.cudnn.benchmark = True # torch.backends.cudnn.deterministic = True #### create train and val dataloader dataset_ratio = 200 # enlarge the size of each epoch for phase, dataset_opt in opt['datasets'].items(): if phase == 'train': train_set = create_dataset(dataset_opt) print('Dataset created') train_size = int( math.ceil(len(train_set) / dataset_opt['batch_size'])) total_iters = int(opt['train']['niter']) total_epochs = int(math.ceil(total_iters / train_size)) train_sampler = None train_loader = create_dataloader(train_set, dataset_opt, opt, train_sampler) if rank <= 0: logger.info( 'Number of train images: {:,d}, iters: {:,d}'.format( len(train_set), train_size)) logger.info('Total epochs needed: {:d} for iters {:,d}'.format( total_epochs, total_iters)) elif phase == 'val': val_set = create_dataset(dataset_opt) val_loader = create_dataloader(val_set, dataset_opt, opt, None) if rank <= 0: logger.info('Number of val images in [{:s}]: {:d}'.format( dataset_opt['name'], len(val_set))) else: raise NotImplementedError( 'Phase [{:s}] is not recognized.'.format(phase)) assert train_loader is not None #### create model current_step = 0 if resume_state is None else resume_state['iter'] model = create_model(opt, current_step) #### resume training if resume_state: logger.info('Resuming training from epoch: {}, iter: {}.'.format( resume_state['epoch'], resume_state['iter'])) start_epoch = resume_state['epoch'] current_step = resume_state['iter'] model.resume_training(resume_state) # handle optimizers and schedulers else: current_step = 0 start_epoch = 0 #### training timer = Timer() logger.info('Start training from epoch: {:d}, iter: {:d}'.format( start_epoch, current_step)) timerData = TickTock() for epoch in range(start_epoch, total_epochs + 1): if opt['dist']: train_sampler.set_epoch(epoch) timerData.tick() for _, train_data in enumerate(train_loader): timerData.tock() current_step += 1 if current_step > total_iters: break #### training model.feed_data(train_data) #### update learning rate model.update_learning_rate(current_step, warmup_iter=opt['train']['warmup_iter']) nll = model.optimize_parameters(current_step) """ try: nll = model.optimize_parameters(current_step) except RuntimeError as e: nll = None print("Skipping ERROR caught in nll = model.optimize_parameters(current_step): ") print(e) """ if nll is None: nll = 0 #### log def eta(t_iter): return (t_iter * (opt['train']['niter'] - current_step)) / 3600 if current_step % opt['logger']['print_freq'] == 0 \ or current_step - (resume_state['iter'] if resume_state else 0) < 25: avg_time = timer.get_average_and_reset() avg_data_time = timerData.get_average_and_reset() message = '<epoch:{:3d}, iter:{:8,d}, lr:{:.3e}, t:{:.2e}, td:{:.2e}, eta:{:.2e}, nll:{:.3e}> '.format( epoch, current_step, model.get_current_learning_rate(), avg_time, avg_data_time, eta(avg_time), nll) print(message) timer.tick() # Reduce number of logs if current_step % 5 == 0: tb_logger_train.add_scalar('loss/nll', nll, current_step) tb_logger_train.add_scalar('lr/base', model.get_current_learning_rate(), current_step) tb_logger_train.add_scalar('time/iteration', timer.get_last_iteration(), current_step) tb_logger_train.add_scalar('time/data', timerData.get_last_iteration(), current_step) tb_logger_train.add_scalar('time/eta', eta(timer.get_last_iteration()), current_step) for k, v in model.get_current_log().items(): tb_logger_train.add_scalar(k, v, current_step) # validation if current_step % opt['train']['val_freq'] == 0 and rank <= 0: avg_psnr = 0.0 idx = 0 nlls = [] for val_data in val_loader: idx += 1 img_name = os.path.splitext( os.path.basename(val_data['LQ_path'][0]))[0] img_dir = os.path.join(opt['path']['val_images'], img_name) util.mkdir(img_dir) model.feed_data(val_data) nll = model.test() if nll is None: nll = 0 nlls.append(nll) visuals = model.get_current_visuals() sr_img = None # Save SR images for reference if hasattr(model, 'heats'): for heat in model.heats: for i in range(model.n_sample): sr_img = util.tensor2img( visuals['SR', heat, i]) # uint8 save_img_path = os.path.join( img_dir, '{:s}_{:09d}_h{:03d}_s{:d}.png'.format( img_name, current_step, int(heat * 100), i)) util.save_img(sr_img, save_img_path) else: sr_img = util.tensor2img(visuals['SR']) # uint8 save_img_path = os.path.join( img_dir, '{:s}_{:d}.png'.format(img_name, current_step)) util.save_img(sr_img, save_img_path) assert sr_img is not None # Save LQ images for reference save_img_path_lq = os.path.join( img_dir, '{:s}_LQ.png'.format(img_name)) if not os.path.isfile(save_img_path_lq): lq_img = util.tensor2img(visuals['LQ']) # uint8 util.save_img( cv2.resize(lq_img, dsize=None, fx=opt['scale'], fy=opt['scale'], interpolation=cv2.INTER_NEAREST), save_img_path_lq) # Save GT images for reference gt_img = util.tensor2img(visuals['GT']) # uint8 save_img_path_gt = os.path.join( img_dir, '{:s}_GT.png'.format(img_name)) if not os.path.isfile(save_img_path_gt): util.save_img(gt_img, save_img_path_gt) # calculate PSNR crop_size = opt['scale'] gt_img = gt_img / 255. sr_img = sr_img / 255. cropped_sr_img = sr_img[crop_size:-crop_size, crop_size:-crop_size, :] cropped_gt_img = gt_img[crop_size:-crop_size, crop_size:-crop_size, :] avg_psnr += util.calculate_psnr(cropped_sr_img * 255, cropped_gt_img * 255) avg_psnr = avg_psnr / idx avg_nll = sum(nlls) / len(nlls) # log logger.info('# Validation # PSNR: {:.4e}'.format(avg_psnr)) logger_val = logging.getLogger('val') # validation logger logger_val.info( '<epoch:{:3d}, iter:{:8,d}> psnr: {:.4e}'.format( epoch, current_step, avg_psnr)) # tensorboard logger tb_logger_valid.add_scalar('loss/psnr', avg_psnr, current_step) tb_logger_valid.add_scalar('loss/nll', avg_nll, current_step) tb_logger_train.flush() tb_logger_valid.flush() #### save models and training states if current_step % opt['logger']['save_checkpoint_freq'] == 0: if rank <= 0: logger.info('Saving models and training states.') model.save(current_step) model.save_training_state(epoch, current_step) timerData.tick() with open(os.path.join(opt['path']['root'], "TRAIN_DONE"), 'w') as f: f.write("TRAIN_DONE") if rank <= 0: logger.info('Saving the final model.') model.save('latest') logger.info('End of training.')
def main(): ############################################ # # set options # ############################################ parser = argparse.ArgumentParser() parser.add_argument('--opt', type=str, help='Path to option YAML file.') parser.add_argument('--launcher', choices=['none', 'pytorch'], default='none', help='job launcher') parser.add_argument('--local_rank', type=int, default=0) args = parser.parse_args() opt = option.parse(args.opt, is_train=True) ############################################ # # distributed training settings # ############################################ if args.launcher == 'none': # disabled distributed training opt['dist'] = False rank = -1 print('Disabled distributed training.') else: opt['dist'] = True init_dist() world_size = torch.distributed.get_world_size() rank = torch.distributed.get_rank() print("Rank:", rank) print("------------------DIST-------------------------") ############################################ # # loading resume state if exists # ############################################ if opt['path'].get('resume_state', None): # distributed resuming: all load into default GPU device_id = torch.cuda.current_device() resume_state = torch.load( opt['path']['resume_state'], map_location=lambda storage, loc: storage.cuda(device_id)) option.check_resume(opt, resume_state['iter']) # check resume options else: resume_state = None ############################################ # # mkdir and loggers # ############################################ if rank <= 0: # normal training (rank -1) OR distributed training (rank 0) if resume_state is None: util.mkdir_and_rename( opt['path'] ['experiments_root']) # rename experiment folder if exists util.mkdirs( (path for key, path in opt['path'].items() if not key == 'experiments_root' and 'pretrain_model' not in key and 'resume' not in key)) # config loggers. Before it, the log will not work util.setup_logger('base', opt['path']['log'], 'train_' + opt['name'], level=logging.INFO, screen=True, tofile=True) util.setup_logger('base_val', opt['path']['log'], 'val_' + opt['name'], level=logging.INFO, screen=True, tofile=True) logger = logging.getLogger('base') logger_val = logging.getLogger('base_val') logger.info(option.dict2str(opt)) # tensorboard logger if opt['use_tb_logger'] and 'debug' not in opt['name']: version = float(torch.__version__[0:3]) if version >= 1.1: # PyTorch 1.1 from torch.utils.tensorboard import SummaryWriter else: logger.info( 'You are using PyTorch {}. Tensorboard will use [tensorboardX]' .format(version)) from tensorboardX import SummaryWriter tb_logger = SummaryWriter(log_dir='../tb_logger/' + opt['name']) else: # config loggers. Before it, the log will not work util.setup_logger('base', opt['path']['log'], 'train_', level=logging.INFO, screen=True) print("set train log") util.setup_logger('base_val', opt['path']['log'], 'val_', level=logging.INFO, screen=True) print("set val log") logger = logging.getLogger('base') logger_val = logging.getLogger('base_val') # convert to NoneDict, which returns None for missing keys opt = option.dict_to_nonedict(opt) #### random seed seed = opt['train']['manual_seed'] if seed is None: seed = random.randint(1, 10000) if rank <= 0: logger.info('Random seed: {}'.format(seed)) util.set_random_seed(seed) torch.backends.cudnn.benchmark = True # torch.backends.cudnn.deterministic = True ############################################ # # create train and val dataloader # ############################################ #### # dataset_ratio = 200 # enlarge the size of each epoch, todo: what it is dataset_ratio = 1 # enlarge the size of each epoch, todo: what it is for phase, dataset_opt in opt['datasets'].items(): if phase == 'train': train_set = create_dataset(dataset_opt) train_size = int( math.ceil(len(train_set) / dataset_opt['batch_size'])) # total_iters = int(opt['train']['niter']) # total_epochs = int(math.ceil(total_iters / train_size)) total_iters = train_size total_epochs = int(opt['train']['epoch']) if opt['dist']: train_sampler = DistIterSampler(train_set, world_size, rank, dataset_ratio) # total_epochs = int(math.ceil(total_iters / (train_size * dataset_ratio))) total_epochs = int(opt['train']['epoch']) if opt['train']['enable'] == False: total_epochs = 1 else: train_sampler = None train_loader = create_dataloader(train_set, dataset_opt, opt, train_sampler) if rank <= 0: logger.info( 'Number of train images: {:,d}, iters: {:,d}'.format( len(train_set), train_size)) logger.info('Total epochs needed: {:d} for iters {:,d}'.format( total_epochs, total_iters)) elif phase == 'val': val_set = create_dataset(dataset_opt) val_loader = create_dataloader(val_set, dataset_opt, opt, None) if rank <= 0: logger.info('Number of val images in [{:s}]: {:d}'.format( dataset_opt['name'], len(val_set))) else: raise NotImplementedError( 'Phase [{:s}] is not recognized.'.format(phase)) assert train_loader is not None ############################################ # # create model # ############################################ #### model = create_model(opt) print("Model Created! ") #### resume training if resume_state: logger.info('Resuming training from epoch: {}, iter: {}.'.format( resume_state['epoch'], resume_state['iter'])) start_epoch = resume_state['epoch'] current_step = resume_state['iter'] model.resume_training(resume_state) # handle optimizers and schedulers else: current_step = 0 start_epoch = 0 print("Not Resume Training") ############################################ # # training # ############################################ #### #### logger.info('Start training from epoch: {:d}, iter: {:d}'.format( start_epoch, current_step)) Avg_train_loss = AverageMeter() # total if (opt['train']['pixel_criterion'] == 'cb+ssim'): Avg_train_loss_pix = AverageMeter() Avg_train_loss_ssim = AverageMeter() elif (opt['train']['pixel_criterion'] == 'cb+ssim+vmaf'): Avg_train_loss_pix = AverageMeter() Avg_train_loss_ssim = AverageMeter() Avg_train_loss_vmaf = AverageMeter() elif (opt['train']['pixel_criterion'] == 'ssim'): Avg_train_loss_ssim = AverageMeter() elif (opt['train']['pixel_criterion'] == 'msssim'): Avg_train_loss_msssim = AverageMeter() elif (opt['train']['pixel_criterion'] == 'cb+msssim'): Avg_train_loss_pix = AverageMeter() Avg_train_loss_msssim = AverageMeter() saved_total_loss = 10e10 saved_total_PSNR = -1 for epoch in range(start_epoch, total_epochs): ############################################ # # Start a new epoch # ############################################ # Turn into training mode #model = model.train() # reset total loss Avg_train_loss.reset() current_step = 0 if (opt['train']['pixel_criterion'] == 'cb+ssim'): Avg_train_loss_pix.reset() Avg_train_loss_ssim.reset() elif (opt['train']['pixel_criterion'] == 'cb+ssim+vmaf'): Avg_train_loss_pix.reset() Avg_train_loss_ssim.reset() Avg_train_loss_vmaf.reset() elif (opt['train']['pixel_criterion'] == 'ssim'): Avg_train_loss_ssim = AverageMeter() elif (opt['train']['pixel_criterion'] == 'msssim'): Avg_train_loss_msssim = AverageMeter() elif (opt['train']['pixel_criterion'] == 'cb+msssim'): Avg_train_loss_pix = AverageMeter() Avg_train_loss_msssim = AverageMeter() if opt['dist']: train_sampler.set_epoch(epoch) for train_idx, train_data in enumerate(train_loader): if 'debug' in opt['name']: img_dir = os.path.join(opt['path']['train_images']) util.mkdir(img_dir) LQ = train_data['LQs'] GT = train_data['GT'] GT_img = util.tensor2img(GT) # uint8 save_img_path = os.path.join( img_dir, '{:4d}_{:s}.png'.format(train_idx, 'debug_GT')) util.save_img(GT_img, save_img_path) for i in range(5): LQ_img = util.tensor2img(LQ[0, i, ...]) # uint8 save_img_path = os.path.join( img_dir, '{:4d}_{:s}_{:1d}.png'.format(train_idx, 'debug_LQ', i)) util.save_img(LQ_img, save_img_path) if (train_idx >= 3): break if opt['train']['enable'] == False: message_train_loss = 'None' break current_step += 1 if current_step > total_iters: print("Total Iteration Reached !") break #### update learning rate if opt['train']['lr_scheme'] == 'ReduceLROnPlateau': pass else: model.update_learning_rate( current_step, warmup_iter=opt['train']['warmup_iter']) #### training model.feed_data(train_data) # if opt['train']['lr_scheme'] == 'ReduceLROnPlateau': # model.optimize_parameters_without_schudlue(current_step) # else: model.optimize_parameters(current_step) if (opt['train']['pixel_criterion'] == 'cb+ssim'): Avg_train_loss.update(model.log_dict['total_loss'], 1) Avg_train_loss_pix.update(model.log_dict['l_pix'], 1) Avg_train_loss_ssim.update(model.log_dict['ssim_loss'], 1) elif (opt['train']['pixel_criterion'] == 'cb+ssim+vmaf'): Avg_train_loss.update(model.log_dict['total_loss'], 1) Avg_train_loss_pix.update(model.log_dict['l_pix'], 1) Avg_train_loss_ssim.update(model.log_dict['ssim_loss'], 1) Avg_train_loss_vmaf.update(model.log_dict['vmaf_loss'], 1) elif (opt['train']['pixel_criterion'] == 'ssim'): Avg_train_loss.update(model.log_dict['total_loss'], 1) Avg_train_loss_ssim.update(model.log_dict['ssim_loss'], 1) elif (opt['train']['pixel_criterion'] == 'msssim'): Avg_train_loss.update(model.log_dict['total_loss'], 1) Avg_train_loss_msssim.update(model.log_dict['msssim_loss'], 1) elif (opt['train']['pixel_criterion'] == 'cb+msssim'): Avg_train_loss.update(model.log_dict['total_loss'], 1) Avg_train_loss_pix.update(model.log_dict['l_pix'], 1) Avg_train_loss_msssim.update(model.log_dict['msssim_loss'], 1) else: Avg_train_loss.update(model.log_dict['l_pix'], 1) # add total train loss if (opt['train']['pixel_criterion'] == 'cb+ssim'): message_train_loss = ' pix_avg_loss: {:.4e}'.format( Avg_train_loss_pix.avg) message_train_loss += ' ssim_avg_loss: {:.4e}'.format( Avg_train_loss_ssim.avg) message_train_loss += ' total_avg_loss: {:.4e}'.format( Avg_train_loss.avg) elif (opt['train']['pixel_criterion'] == 'cb+ssim+vmaf'): message_train_loss = ' pix_avg_loss: {:.4e}'.format( Avg_train_loss_pix.avg) message_train_loss += ' ssim_avg_loss: {:.4e}'.format( Avg_train_loss_ssim.avg) message_train_loss += ' vmaf_avg_loss: {:.4e}'.format( Avg_train_loss_vmaf.avg) message_train_loss += ' total_avg_loss: {:.4e}'.format( Avg_train_loss.avg) elif (opt['train']['pixel_criterion'] == 'ssim'): message_train_loss = ' ssim_avg_loss: {:.4e}'.format( Avg_train_loss_ssim.avg) message_train_loss += ' total_avg_loss: {:.4e}'.format( Avg_train_loss.avg) elif (opt['train']['pixel_criterion'] == 'msssim'): message_train_loss = ' msssim_avg_loss: {:.4e}'.format( Avg_train_loss_msssim.avg) message_train_loss += ' total_avg_loss: {:.4e}'.format( Avg_train_loss.avg) elif (opt['train']['pixel_criterion'] == 'cb+msssim'): message_train_loss = ' pix_avg_loss: {:.4e}'.format( Avg_train_loss_pix.avg) message_train_loss += ' msssim_avg_loss: {:.4e}'.format( Avg_train_loss_msssim.avg) message_train_loss += ' total_avg_loss: {:.4e}'.format( Avg_train_loss.avg) else: message_train_loss = ' train_avg_loss: {:.4e}'.format( Avg_train_loss.avg) #### log if current_step % opt['logger']['print_freq'] == 0: logs = model.get_current_log() message = '[epoch:{:3d}, iter:{:8,d}, lr:('.format( epoch, current_step) for v in model.get_current_learning_rate(): message += '{:.3e},'.format(v) message += ')] ' for k, v in logs.items(): message += '{:s}: {:.4e} '.format(k, v) # tensorboard logger if opt['use_tb_logger'] and 'debug' not in opt['name']: if rank <= 0: tb_logger.add_scalar(k, v, current_step) message += message_train_loss if rank <= 0: logger.info(message) ############################################ # # end of one epoch, save epoch model # ############################################ #### save models and training states # if current_step % opt['logger']['save_checkpoint_freq'] == 0: # if rank <= 0: # logger.info('Saving models and training states.') # model.save(current_step) # model.save('latest') # # model.save_training_state(epoch, current_step) # # todo delete previous weights # previous_step = current_step - opt['logger']['save_checkpoint_freq'] # save_filename = '{}_{}.pth'.format(previous_step, 'G') # save_path = os.path.join(opt['path']['models'], save_filename) # if os.path.exists(save_path): # os.remove(save_path) if epoch == 1: save_filename = '{:04d}_{}.pth'.format(0, 'G') save_path = os.path.join(opt['path']['models'], save_filename) if os.path.exists(save_path): os.remove(save_path) save_filename = '{:04d}_{}.pth'.format(epoch - 1, 'G') save_path = os.path.join(opt['path']['models'], save_filename) if os.path.exists(save_path): os.remove(save_path) if rank <= 0: logger.info('Saving models and training states.') save_filename = '{:04d}'.format(epoch) model.save(save_filename) # model.save('latest') # model.save_training_state(epoch, current_step) ############################################ # # end of one epoch, do validation # ############################################ #### validation #if opt['datasets'].get('val', None) and current_step % opt['train']['val_freq'] == 0: if opt['datasets'].get('val', None): if opt['model'] in [ 'sr', 'srgan' ] and rank <= 0: # image restoration validation # does not support multi-GPU validation pbar = util.ProgressBar(len(val_loader)) avg_psnr = 0. idx = 0 for val_data in val_loader: idx += 1 img_name = os.path.splitext( os.path.basename(val_data['LQ_path'][0]))[0] img_dir = os.path.join(opt['path']['val_images'], img_name) util.mkdir(img_dir) model.feed_data(val_data) model.test() visuals = model.get_current_visuals() sr_img = util.tensor2img(visuals['rlt']) # uint8 gt_img = util.tensor2img(visuals['GT']) # uint8 # Save SR images for reference save_img_path = os.path.join( img_dir, '{:s}_{:d}.png'.format(img_name, current_step)) #util.save_img(sr_img, save_img_path) # calculate PSNR sr_img, gt_img = util.crop_border([sr_img, gt_img], opt['scale']) avg_psnr += util.calculate_psnr(sr_img, gt_img) pbar.update('Test {}'.format(img_name)) avg_psnr = avg_psnr / idx # log logger.info('# Validation # PSNR: {:.4e}'.format(avg_psnr)) # tensorboard logger if opt['use_tb_logger'] and 'debug' not in opt['name']: tb_logger.add_scalar('psnr', avg_psnr, current_step) else: # video restoration validation if opt['dist']: # todo : multi-GPU testing psnr_rlt = {} # with border and center frames psnr_rlt_avg = {} psnr_total_avg = 0. ssim_rlt = {} # with border and center frames ssim_rlt_avg = {} ssim_total_avg = 0. val_loss_rlt = {} val_loss_rlt_avg = {} val_loss_total_avg = 0. if rank == 0: pbar = util.ProgressBar(len(val_set)) for idx in range(rank, len(val_set), world_size): print('idx', idx) if 'debug' in opt['name']: if (idx >= 3): break val_data = val_set[idx] val_data['LQs'].unsqueeze_(0) val_data['GT'].unsqueeze_(0) folder = val_data['folder'] idx_d, max_idx = val_data['idx'].split('/') idx_d, max_idx = int(idx_d), int(max_idx) if psnr_rlt.get(folder, None) is None: psnr_rlt[folder] = torch.zeros(max_idx, dtype=torch.float32, device='cuda') if ssim_rlt.get(folder, None) is None: ssim_rlt[folder] = torch.zeros(max_idx, dtype=torch.float32, device='cuda') if val_loss_rlt.get(folder, None) is None: val_loss_rlt[folder] = torch.zeros( max_idx, dtype=torch.float32, device='cuda') # tmp = torch.zeros(max_idx, dtype=torch.float32, device='cuda') model.feed_data(val_data) # model.test() # model.test_stitch() if opt['stitch'] == True: model.test_stitch() else: model.test() # large GPU memory # visuals = model.get_current_visuals() visuals = model.get_current_visuals( save=True, name='{}_{}'.format(folder, idx), save_path=opt['path']['val_images']) rlt_img = util.tensor2img(visuals['rlt']) # uint8 gt_img = util.tensor2img(visuals['GT']) # uint8 # calculate PSNR psnr = util.calculate_psnr(rlt_img, gt_img) psnr_rlt[folder][idx_d] = psnr # calculate SSIM ssim = util.calculate_ssim(rlt_img, gt_img) ssim_rlt[folder][idx_d] = ssim # calculate Val loss val_loss = model.get_loss() val_loss_rlt[folder][idx_d] = val_loss logger.info( '{}_{:02d} PSNR: {:.4f}, SSIM: {:.4f}'.format( folder, idx, psnr, ssim)) if rank == 0: for _ in range(world_size): pbar.update('Test {} - {}/{}'.format( folder, idx_d, max_idx)) # # collect data for _, v in psnr_rlt.items(): dist.reduce(v, 0) for _, v in ssim_rlt.items(): dist.reduce(v, 0) for _, v in val_loss_rlt.items(): dist.reduce(v, 0) dist.barrier() if rank == 0: psnr_rlt_avg = {} psnr_total_avg = 0. for k, v in psnr_rlt.items(): psnr_rlt_avg[k] = torch.mean(v).cpu().item() psnr_total_avg += psnr_rlt_avg[k] psnr_total_avg /= len(psnr_rlt) log_s = '# Validation # PSNR: {:.4e}:'.format( psnr_total_avg) for k, v in psnr_rlt_avg.items(): log_s += ' {}: {:.4e}'.format(k, v) logger.info(log_s) # ssim ssim_rlt_avg = {} ssim_total_avg = 0. for k, v in ssim_rlt.items(): ssim_rlt_avg[k] = torch.mean(v).cpu().item() ssim_total_avg += ssim_rlt_avg[k] ssim_total_avg /= len(ssim_rlt) log_s = '# Validation # PSNR: {:.4e}:'.format( ssim_total_avg) for k, v in ssim_rlt_avg.items(): log_s += ' {}: {:.4e}'.format(k, v) logger.info(log_s) # added val_loss_rlt_avg = {} val_loss_total_avg = 0. for k, v in val_loss_rlt.items(): val_loss_rlt_avg[k] = torch.mean(v).cpu().item() val_loss_total_avg += val_loss_rlt_avg[k] val_loss_total_avg /= len(val_loss_rlt) log_l = '# Validation # Loss: {:.4e}:'.format( val_loss_total_avg) for k, v in val_loss_rlt_avg.items(): log_l += ' {}: {:.4e}'.format(k, v) logger.info(log_l) message = '' for v in model.get_current_learning_rate(): message += '{:.5e}'.format(v) logger_val.info( 'Epoch {:02d}, LR {:s}, PSNR {:.4f}, SSIM {:.4f} Train {:s}, Val Total Loss {:.4e}' .format(epoch, message, psnr_total_avg, ssim_total_avg, message_train_loss, val_loss_total_avg)) if opt['use_tb_logger'] and 'debug' not in opt['name']: tb_logger.add_scalar('psnr_avg', psnr_total_avg, current_step) for k, v in psnr_rlt_avg.items(): tb_logger.add_scalar(k, v, current_step) # add val loss tb_logger.add_scalar('val_loss_avg', val_loss_total_avg, current_step) for k, v in val_loss_rlt_avg.items(): tb_logger.add_scalar(k, v, current_step) else: # Todo: our function One GPU pbar = util.ProgressBar(len(val_loader)) psnr_rlt = {} # with border and center frames psnr_rlt_avg = {} psnr_total_avg = 0. ssim_rlt = {} # with border and center frames ssim_rlt_avg = {} ssim_total_avg = 0. val_loss_rlt = {} val_loss_rlt_avg = {} val_loss_total_avg = 0. for val_inx, val_data in enumerate(val_loader): if 'debug' in opt['name']: if (val_inx >= 5): break folder = val_data['folder'][0] # idx_d = val_data['idx'].item() idx_d = val_data['idx'] # border = val_data['border'].item() if psnr_rlt.get(folder, None) is None: psnr_rlt[folder] = [] if ssim_rlt.get(folder, None) is None: ssim_rlt[folder] = [] if val_loss_rlt.get(folder, None) is None: val_loss_rlt[folder] = [] # process the black blank [B N C H W] print(val_data['LQs'].size()) H_S = val_data['LQs'].size(3) # 540 W_S = val_data['LQs'].size(4) # 960 print(H_S) print(W_S) blank_1_S = 0 blank_2_S = 0 print(val_data['LQs'][0, 2, 0, :, :].size()) for i in range(H_S): if not sum(val_data['LQs'][0, 2, 0, i, :]) == 0: blank_1_S = i - 1 # assert not sum(data_S[:, :, 0][i+1]) == 0 break for i in range(H_S): if not sum(val_data['LQs'][0, 2, 0, :, H_S - i - 1]) == 0: blank_2_S = (H_S - 1) - i - 1 # assert not sum(data_S[:, :, 0][blank_2_S-1]) == 0 break print('LQ :', blank_1_S, blank_2_S) if blank_1_S == -1: print('LQ has no blank') blank_1_S = 0 blank_2_S = H_S # val_data['LQs'] = val_data['LQs'][:,:,:,blank_1_S:blank_2_S,:] print("LQ", val_data['LQs'].size()) # end of process the black blank model.feed_data(val_data) if opt['stitch'] == True: model.test_stitch() else: model.test() # large GPU memory # process blank blank_1_L = blank_1_S << 2 blank_2_L = blank_2_S << 2 print(blank_1_L, blank_2_L) print(model.fake_H.size()) if not blank_1_S == 0: # model.fake_H = model.fake_H[:,:,blank_1_L:blank_2_L,:] model.fake_H[:, :, 0:blank_1_L, :] = 0 model.fake_H[:, :, blank_2_L:H_S, :] = 0 # end of # process blank visuals = model.get_current_visuals( save=True, name='{}_{:02d}'.format(folder, val_inx), save_path=opt['path']['val_images']) rlt_img = util.tensor2img(visuals['rlt']) # uint8 gt_img = util.tensor2img(visuals['GT']) # uint8 # calculate PSNR psnr = util.calculate_psnr(rlt_img, gt_img) psnr_rlt[folder].append(psnr) # calculate SSIM ssim = util.calculate_ssim(rlt_img, gt_img) ssim_rlt[folder].append(ssim) # val loss val_loss = model.get_loss() val_loss_rlt[folder].append(val_loss.item()) logger.info( '{}_{:02d} PSNR: {:.4f}, SSIM: {:.4f}'.format( folder, val_inx, psnr, ssim)) pbar.update('Test {} - {}'.format(folder, idx_d)) # average PSNR for k, v in psnr_rlt.items(): psnr_rlt_avg[k] = sum(v) / len(v) psnr_total_avg += psnr_rlt_avg[k] psnr_total_avg /= len(psnr_rlt) log_s = '# Validation # PSNR: {:.4e}:'.format( psnr_total_avg) for k, v in psnr_rlt_avg.items(): log_s += ' {}: {:.4e}'.format(k, v) logger.info(log_s) # average SSIM for k, v in ssim_rlt.items(): ssim_rlt_avg[k] = sum(v) / len(v) ssim_total_avg += ssim_rlt_avg[k] ssim_total_avg /= len(ssim_rlt) log_s = '# Validation # SSIM: {:.4e}:'.format( ssim_total_avg) for k, v in ssim_rlt_avg.items(): log_s += ' {}: {:.4e}'.format(k, v) logger.info(log_s) # average VMAF # average Val LOSS for k, v in val_loss_rlt.items(): val_loss_rlt_avg[k] = sum(v) / len(v) val_loss_total_avg += val_loss_rlt_avg[k] val_loss_total_avg /= len(val_loss_rlt) log_l = '# Validation # Loss: {:.4e}:'.format( val_loss_total_avg) for k, v in val_loss_rlt_avg.items(): log_l += ' {}: {:.4e}'.format(k, v) logger.info(log_l) # toal validation log message = '' for v in model.get_current_learning_rate(): message += '{:.5e}'.format(v) logger_val.info( 'Epoch {:02d}, LR {:s}, PSNR {:.4f}, SSIM {:.4f} Train {:s}, Val Total Loss {:.4e}' .format(epoch, message, psnr_total_avg, ssim_total_avg, message_train_loss, val_loss_total_avg)) # end add if opt['use_tb_logger'] and 'debug' not in opt['name']: tb_logger.add_scalar('psnr_avg', psnr_total_avg, current_step) for k, v in psnr_rlt_avg.items(): tb_logger.add_scalar(k, v, current_step) # tb_logger.add_scalar('ssim_avg', ssim_total_avg, current_step) # for k, v in ssim_rlt_avg.items(): # tb_logger.add_scalar(k, v, current_step) # add val loss tb_logger.add_scalar('val_loss_avg', val_loss_total_avg, current_step) for k, v in val_loss_rlt_avg.items(): tb_logger.add_scalar(k, v, current_step) ############################################ # # end of validation, save model # ############################################ # logger.info("Finished an epoch, Check and Save the model weights") # we check the validation loss instead of training loss. OK~ if saved_total_loss >= val_loss_total_avg: saved_total_loss = val_loss_total_avg #torch.save(model.state_dict(), args.save_path + "/best" + ".pth") model.save('best') logger.info( "Best Weights updated for decreased validation loss") else: logger.info( "Weights Not updated for undecreased validation loss") if saved_total_PSNR <= psnr_total_avg: saved_total_PSNR = psnr_total_avg model.save('bestPSNR') logger.info( "Best Weights updated for increased validation PSNR") else: logger.info( "Weights Not updated for unincreased validation PSNR") ############################################ # # end of one epoch, schedule LR # ############################################ # add scheduler todo if opt['train']['lr_scheme'] == 'ReduceLROnPlateau': for scheduler in model.schedulers: # scheduler.step(val_loss_total_avg) scheduler.step(val_loss_total_avg) if rank <= 0: logger.info('Saving the final model.') model.save('last') logger.info('End of training.') tb_logger.close()
def main(): """The main function.""" parser = argparse.ArgumentParser() parser.add_argument('-opt', type=str, required=True, help='Path to option JSON file.') options = option.parse( parser.parse_args().opt) # load settings and initialize settings util.mkdir_and_rename( options['path']['results_root']) # rename old experiments if exists util.mkdirs((path for key, path in options['path'].items() if not key == 'experiments_root' and not key == 'saved_model' and not key == 'fc_root' and not key == 'db_root')) option.save(options) options = option.dict_to_nonedict( options) # Convert to NoneDict, which return None for missing key. # Redirect all writes to the "txt" file sys.stdout = PrintLogger(options['path']['log']) # # logger = Logger(opt) if options['manual_seed'] is not None: random.seed(options['manual_seed']) index = None if options['dataset'] == 'pipal': index = list(range(0, 200)) else: raise NotImplementedError lr_backup = options['base_lr'] lr_steps_backup = options['lr_steps'] epochs_backup = options['epochs'] srcc_all = np.zeros((1, options['epoch']), dtype=np.float) plcc_all = np.zeros((1, options['epoch']), dtype=np.float) print('Total epochs:' + str(options['epoch'])) for i in range(0, options['epoch']): # randomly split train-test set random.shuffle(index) train_index = index[0:round(0.8 * len(index))] test_index = index[round(0.8 * len(index)):len(index)] options['train_index'] = train_index options['test_index'] = test_index # train the fully connected layer only print('[No.%d/%d] Training the FC layer...' % (i, options['epoch'])) options['fc'] = True options['base_lr'] = lr_backup options['lr_steps'] = lr_steps_backup options['epochs'] = epochs_backup manager = IQAManager(options) best_srcc1, best_plcc1 = manager.train() # fine-tune all model print('[No.%d/%d] Fine-tune all model...' % (i, options['epoch'])) options['fc'] = False options['base_lr'] = options['base_lr_step2'] options['lr_steps'] = options['lr_steps_step2'] options['epochs'] = options['epochs_step2'] manager = IQAManager(options) best_srcc2, best_plcc2 = manager.train() srcc_all[0][i] = np.max([best_srcc1, best_srcc2]) plcc_all[0][i] = np.max([best_plcc1, best_plcc2]) # srcc_all[0][i] = best_srcc1 # plcc_all[0][i] = best_plcc1 srcc_mean = np.mean(srcc_all) srcc_median = np.median(srcc_all) plcc_mean = np.mean(plcc_all) plcc_median = np.median(plcc_all) print(srcc_all) print('average srcc:%4.4f' % srcc_mean) print('median srcc:%4.4f' % srcc_median) print(plcc_all) print('average plcc:%4.4f' % plcc_mean) print('median plcc:%4.4f' % plcc_median) print('--------------Finish! [' + options['name'] + ']--------------') util.mkdir_and_rename(options['path']['results_root'], ('done_{:.4f}'.format(srcc_median)))
def main(): # options parser = argparse.ArgumentParser() parser.add_argument('--opt', type=str, required=True, help='Path to option JSON file.') opt = option.parse(parser.parse_args(), is_train=True) opt = option.dict_to_nonedict( opt) # Convert to NoneDict, which return None for missing key. # train from scratch OR resume training if opt['path']['resume_state']: # resuming training resume_state = torch.load(opt['path']['resume_state']) else: # training from scratch resume_state = None util.mkdir_and_rename( opt['path']['experiments_root']) # rename old folder if exists util.mkdirs((path for key, path in opt['path'].items() if not key == 'experiments_root' and 'pretrain_model' not in key and 'resume' not in key)) # config loggers. Before it, the log will not work util.setup_logger(None, opt['path']['log'], 'train', level=logging.INFO, screen=True) util.setup_logger('val', opt['path']['log'], 'val', level=logging.INFO) logger = logging.getLogger('base') if resume_state: logger.info('Resuming training from epoch: {}, iter: {}.'.format( resume_state['epoch'], resume_state['iter'])) option.check_resume(opt) # check resume options logger.info(option.dict2str(opt)) # tensorboard logger if opt['use_tb_logger'] and 'debug' not in opt['name']: from tensorboardX import SummaryWriter tb_logger = SummaryWriter(log_dir='../tb_logger/' + opt['name']) # random seed seed = opt['train']['manual_seed'] if seed is None: seed = random.randint(1, 10000) logger.info('Random seed: {}'.format(seed)) util.set_random_seed(seed) torch.backends.cudnn.benckmark = True # torch.backends.cudnn.deterministic = True # create train and val dataloader for phase, dataset_opt in opt['datasets'].items(): if phase == 'train': train_set = create_dataset(dataset_opt) train_size = int( math.ceil(len(train_set) / dataset_opt['batch_size'])) logger.info('Number of train images: {:,d}, iters: {:,d}'.format( len(train_set), train_size)) total_iters = int(opt['train']['niter']) total_epochs = int(math.ceil(total_iters / train_size)) logger.info('Total epochs needed: {:d} for iters {:,d}'.format( total_epochs, total_iters)) train_loader = create_dataloader(train_set, dataset_opt) elif phase == 'val': val_set = create_dataset(dataset_opt) val_loader = create_dataloader(val_set, dataset_opt) logger.info('Number of val images in [{:s}]: {:d}'.format( dataset_opt['name'], len(val_set))) else: raise NotImplementedError( 'Phase [{:s}] is not recognized.'.format(phase)) assert train_loader is not None # create model model = create_model(opt) # resume training if resume_state: start_epoch = resume_state['epoch'] current_step = resume_state['iter'] model.resume_training(resume_state) # handle optimizers and schedulers else: current_step = 0 start_epoch = 0 # training logger.info('Start training from epoch: {:d}, iter: {:d}'.format( start_epoch, current_step)) for epoch in range(start_epoch, total_epochs): for _, train_data in enumerate(train_loader): current_step += 1 if current_step > total_iters: break # validation if current_step % opt['train']['val_freq'] == 0: visuals = model.get_current_visuals(isTrain=True) os.makedirs('valid/' + opt['name'], exist_ok=True) util.save_img( util.tensor2img(visuals['LR']), 'valid/' + opt['name'] + '/' + str(current_step) + '_LR.png') util.save_img( util.tensor2img(visuals['HR']), 'valid/' + opt['name'] + '/' + str(current_step) + '_HR.png') util.save_img( util.tensor2img(visuals['SR']), 'valid/' + opt['name'] + '/' + str(current_step) + '_SR.png') util.save_img( util.tensor2img(visuals['SRgray']), 'valid/' + opt['name'] + '/' + str(current_step) + '_SRgray.png') util.save_img( util.tensor2img(visuals['HRgray']), 'valid/' + opt['name'] + '/' + str(current_step) + '_HRgray.png') # update learning rate model.update_learning_rate() # training model.feed_data(train_data) model.optimize_parameters(current_step) # log if current_step % opt['logger']['print_freq'] == 0: logs = model.get_current_log() message = '<epoch:{:3d}, iter:{:8,d}, lr:{:.3e}> '.format( epoch, current_step, model.get_current_learning_rate()) for k, v in logs.items(): message += '{:s}: {:.4e} '.format(k, v) # tensorboard logger if opt['use_tb_logger'] and 'debug' not in opt['name']: tb_logger.add_scalar(k, v, current_step) logger.info(message) # save models and training states if current_step % opt['logger']['save_checkpoint_freq'] == 0: logger.info('Saving models and training states.') model.save(current_step) model.save_training_state(epoch, current_step) logger.info('Saving the final model.') model.save('latest') logger.info('End of training.')
def main(): #### options parser = argparse.ArgumentParser() parser.add_argument('-opt', type=str, help='Path to option YAML file.') parser.add_argument('--launcher', choices=['none', 'pytorch'], default='none', help='job launcher') parser.add_argument('--local_rank', type=int, default=0) args = parser.parse_args() opt = option.parse(args.opt, is_train=True) #### distributed training settings if args.launcher == 'none': # disabled distributed training opt['dist'] = False rank = -1 print('Disabled distributed training.') else: opt['dist'] = True init_dist() world_size = torch.distributed.get_world_size() rank = torch.distributed.get_rank() #### loading resume state if exists if opt['path'].get('resume_state', None): # distributed resuming: all load into default GPU device_id = torch.cuda.current_device() resume_state = torch.load( opt['path']['resume_state'], map_location=lambda storage, loc: storage.cuda(device_id)) option.check_resume(opt, resume_state['iter']) # check resume options else: resume_state = None #### mkdir and loggers if rank <= 0: # normal training (rank -1) OR distributed training (rank 0) if resume_state is None: util.mkdir_and_rename( opt['path'] ['experiments_root']) # rename experiment folder if exists util.mkdirs( (path for key, path in opt['path'].items() if not key == 'experiments_root' and 'pretrain_model' not in key and 'resume' not in key and 'wandb_load_run_path' not in key)) # config loggers. Before it, the log will not work util.setup_logger('base', opt['path']['log'], 'train_' + opt['name'], level=logging.INFO, screen=True, tofile=True) logger = logging.getLogger('base') logger.info(option.dict2str(opt)) # tensorboard logger if opt['use_tb_logger'] and 'debug' not in opt['name']: version = float(torch.__version__[0:3]) if version >= 1.1: # PyTorch 1.1 from torch.utils.tensorboard import SummaryWriter else: logger.info( 'You are using PyTorch {}. Tensorboard will use [tensorboardX]' .format(version)) from tensorboardX import SummaryWriter tb_logger = SummaryWriter(log_dir='../tb_logger/' + opt['name']) if opt['use_wandb_logger'] and 'debug' not in opt['name']: json_path = os.path.join(os.path.expanduser('~'), '.wandb_api_keys.json') if os.path.exists(json_path): with open(json_path, 'r') as j: json_file = json.loads(j.read()) os.environ['WANDB_API_KEY'] = json_file['ryul99'] wandb.init(project="mmsr", config=opt, sync_tensorboard=True) else: util.setup_logger('base', opt['path']['log'], 'train', level=logging.INFO, screen=True) logger = logging.getLogger('base') # convert to NoneDict, which returns None for missing keys opt = option.dict_to_nonedict(opt) #### random seed seed = opt['train']['manual_seed'] if seed is None: seed = random.randint(1, 10000) if rank <= 0: logger.info('Random seed: {}'.format(seed)) if opt['use_wandb_logger'] and 'debug' not in opt['name']: wandb.config.update({'random_seed': seed}) util.set_random_seed(seed) torch.backends.cudnn.benchmark = True # torch.backends.cudnn.deterministic = True #### create train and val dataloader dataset_ratio = 200 # enlarge the size of each epoch for phase, dataset_opt in opt['datasets'].items(): if phase == 'train': train_set = create_dataset(dataset_opt) train_size = int( math.ceil(len(train_set) / dataset_opt['batch_size'])) total_iters = int(opt['train']['niter']) total_epochs = int(math.ceil(total_iters / train_size)) if opt['dist']: train_sampler = DistIterSampler(train_set, world_size, rank, dataset_ratio) total_epochs = int( math.ceil(total_iters / (train_size * dataset_ratio))) else: train_sampler = None train_loader = create_dataloader(train_set, dataset_opt, opt, train_sampler) if rank <= 0: logger.info( 'Number of train images: {:,d}, iters: {:,d}'.format( len(train_set), train_size)) logger.info('Total epochs needed: {:d} for iters {:,d}'.format( total_epochs, total_iters)) elif phase == 'val': val_set = create_dataset(dataset_opt) val_loader = create_dataloader(val_set, dataset_opt, opt, None) if rank <= 0: logger.info('Number of val images in [{:s}]: {:d}'.format( dataset_opt['name'], len(val_set))) else: raise NotImplementedError( 'Phase [{:s}] is not recognized.'.format(phase)) assert train_loader is not None #### create model model = create_model(opt) #### resume training if resume_state: logger.info('Resuming training from epoch: {}, iter: {}.'.format( resume_state['epoch'], resume_state['iter'])) start_epoch = resume_state['epoch'] current_step = resume_state['iter'] model.resume_training(resume_state) # handle optimizers and schedulers else: current_step = 0 start_epoch = 0 #### training logger.info('Start training from epoch: {:d}, iter: {:d}'.format( start_epoch, current_step)) for epoch in range(start_epoch, total_epochs + 1): if opt['dist']: train_sampler.set_epoch(epoch) for _, train_data in enumerate(train_loader): current_step += 1 if current_step > total_iters: break #### update learning rate model.update_learning_rate(current_step, warmup_iter=opt['train']['warmup_iter']) #### training model.feed_data(train_data, noise_mode=opt['datasets']['train']['noise_mode'], noise_rate=opt['datasets']['train']['noise_rate']) model.optimize_parameters(current_step) #### log if current_step % opt['logger']['print_freq'] == 0: logs = model.get_current_log() message = '[epoch:{:3d}, iter:{:8,d}, lr:('.format( epoch, current_step) for v in model.get_current_learning_rate(): message += '{:.3e},'.format(v) message += ')] ' for k, v in logs.items(): message += '{:s}: {:.4e} '.format(k, v) # tensorboard logger if opt['use_tb_logger'] and 'debug' not in opt['name']: if rank <= 0: tb_logger.add_scalar(k, v, current_step) if opt['use_wandb_logger'] and 'debug' not in opt['name']: if rank <= 0: wandb.log({k: v}, step=current_step) if rank <= 0: logger.info(message) #### validation if opt['datasets'].get( 'val', None) and current_step % opt['train']['val_freq'] == 0: if opt['model'] in [ 'sr', 'srgan' ] and rank <= 0: # image restoration validation # does not support multi-GPU validation pbar = util.ProgressBar(len(val_loader)) avg_psnr = 0. idx = 0 for val_data in val_loader: idx += 1 img_name = os.path.splitext( os.path.basename(val_data['LQ_path'][0]))[0] img_dir = os.path.join(opt['path']['val_images'], img_name) util.mkdir(img_dir) model.feed_data( val_data, noise_mode=opt['datasets']['val']['noise_mode'], noise_rate=opt['datasets']['val']['noise_rate']) model.test() visuals = model.get_current_visuals() sr_img = util.tensor2img(visuals['rlt']) # uint8 gt_img = util.tensor2img(visuals['GT']) # uint8 # Save SR images for reference save_img_path = os.path.join( img_dir, '{:s}_{:d}.png'.format(img_name, current_step)) util.save_img(sr_img, save_img_path) # calculate PSNR sr_img, gt_img = util.crop_border([sr_img, gt_img], opt['scale']) avg_psnr += util.calculate_psnr(sr_img, gt_img) pbar.update('Test {}'.format(img_name)) avg_psnr = avg_psnr / idx # log logger.info('# Validation # PSNR: {:.4e}'.format(avg_psnr)) # tensorboard logger if opt['use_tb_logger'] and 'debug' not in opt['name']: tb_logger.add_scalar('psnr', avg_psnr, current_step) if opt['use_wandb_logger'] and 'debug' not in opt['name']: wandb.log({'psnr': avg_psnr}, step=current_step) else: # video restoration validation if opt['dist']: # multi-GPU testing psnr_rlt = {} # with border and center frames if rank == 0: pbar = util.ProgressBar(len(val_set)) for idx in range(rank, len(val_set), world_size): val_data = val_set[idx] val_data['LQs'].unsqueeze_(0) val_data['GT'].unsqueeze_(0) folder = val_data['folder'] idx_d, max_idx = val_data['idx'].split('/') idx_d, max_idx = int(idx_d), int(max_idx) if psnr_rlt.get(folder, None) is None: psnr_rlt[folder] = torch.zeros( max_idx, dtype=torch.float32, device='cuda') # tmp = torch.zeros(max_idx, dtype=torch.float32, device='cuda') model.feed_data(val_data, noise_mode=opt['datasets']['val'] ['noise_mode'], noise_rate=opt['datasets']['val'] ['noise_rate']) model.test() visuals = model.get_current_visuals() rlt_img = util.tensor2img(visuals['rlt']) # uint8 gt_img = util.tensor2img(visuals['GT']) # uint8 # calculate PSNR psnr_rlt[folder][idx_d] = util.calculate_psnr( rlt_img, gt_img) if rank == 0: for _ in range(world_size): pbar.update('Test {} - {}/{}'.format( folder, idx_d, max_idx)) # # collect data for _, v in psnr_rlt.items(): dist.reduce(v, 0) dist.barrier() if rank == 0: psnr_rlt_avg = {} psnr_total_avg = 0. for k, v in psnr_rlt.items(): psnr_rlt_avg[k] = torch.mean(v).cpu().item() psnr_total_avg += psnr_rlt_avg[k] psnr_total_avg /= len(psnr_rlt) log_s = '# Validation # PSNR: {:.4e}:'.format( psnr_total_avg) for k, v in psnr_rlt_avg.items(): log_s += ' {}: {:.4e}'.format(k, v) logger.info(log_s) if opt['use_tb_logger'] and 'debug' not in opt[ 'name']: tb_logger.add_scalar('psnr_avg', psnr_total_avg, current_step) for k, v in psnr_rlt_avg.items(): tb_logger.add_scalar(k, v, current_step) if opt['use_wandb_logger'] and 'debug' not in opt[ 'name']: lq_img, rlt_img, gt_img = map( util.tensor2img, [ visuals['LQ'], visuals['rlt'], visuals['GT'] ]) wandb.log({'psnr_avg': psnr_total_avg}, step=current_step) wandb.log(psnr_rlt_avg, step=current_step) wandb.log( { 'Validation Image': [ wandb.Image(lq_img[:, :, [2, 1, 0]], caption='LQ'), wandb.Image(rlt_img[:, :, [2, 1, 0]], caption='output'), wandb.Image(gt_img[:, :, [2, 1, 0]], caption='GT'), ] }, step=current_step) else: pbar = util.ProgressBar(len(val_loader)) psnr_rlt = {} # with border and center frames psnr_rlt_avg = {} psnr_total_avg = 0. for val_data in val_loader: folder = val_data['folder'][0] idx_d = val_data['idx'].item() # border = val_data['border'].item() if psnr_rlt.get(folder, None) is None: psnr_rlt[folder] = [] model.feed_data(val_data, noise_mode=opt['datasets']['val'] ['noise_mode'], noise_rate=opt['datasets']['val'] ['noise_rate']) model.test() visuals = model.get_current_visuals() rlt_img = util.tensor2img(visuals['rlt']) # uint8 gt_img = util.tensor2img(visuals['GT']) # uint8 # calculate PSNR psnr = util.calculate_psnr(rlt_img, gt_img) psnr_rlt[folder].append(psnr) pbar.update('Test {} - {}'.format(folder, idx_d)) for k, v in psnr_rlt.items(): psnr_rlt_avg[k] = sum(v) / len(v) psnr_total_avg += psnr_rlt_avg[k] psnr_total_avg /= len(psnr_rlt) log_s = '# Validation # PSNR: {:.4e}:'.format( psnr_total_avg) for k, v in psnr_rlt_avg.items(): log_s += ' {}: {:.4e}'.format(k, v) logger.info(log_s) if opt['use_tb_logger'] and 'debug' not in opt['name']: tb_logger.add_scalar('psnr_avg', psnr_total_avg, current_step) for k, v in psnr_rlt_avg.items(): tb_logger.add_scalar(k, v, current_step) if opt['use_wandb_logger'] and 'debug' not in opt[ 'name']: lq_img, rlt_img, gt_img = map( util.tensor2img, [visuals['LQ'], visuals['rlt'], visuals['GT']]) wandb.log({'psnr_avg': psnr_total_avg}, step=current_step) wandb.log(psnr_rlt_avg, step=current_step) wandb.log( { 'Validation Image': [ wandb.Image(lq_img[:, :, [2, 1, 0]], caption='LQ'), wandb.Image(rlt_img[:, :, [2, 1, 0]], caption='output'), wandb.Image(gt_img[:, :, [2, 1, 0]], caption='GT'), ] }, step=current_step) #### save models and training states if current_step % opt['logger']['save_checkpoint_freq'] == 0: if rank <= 0: logger.info('Saving models and training states.') model.save(current_step) model.save_training_state(epoch, current_step) if rank <= 0: logger.info('Saving the final model.') model.save('latest') logger.info('End of training.') if opt['use_tb_logger'] and 'debug' not in opt['name']: tb_logger.close()
# img_list = glob.glob('/mnt/lustrenew/dongrunmin/dongrunmin/SR/RefSR/github_pub_round1/dataset/val/val_1st/LR/*') # HR_path = '/mnt/lustrenew/dongrunmin/dongrunmin/SR/RefSR/github_pub_round1/dataset/val/val_1st/HR' #elif dataset == 'val_2nd': # img_list = glob.glob('/mnt/lustrenew/dongrunmin/dongrunmin/SR/RefSR/github_pub_round1/dataset/val/val_2nd/LR/*') # HR_path = '/mnt/lustrenew/dongrunmin/dongrunmin/SR/RefSR/github_pub_round1/dataset/val/val_2nd/HR' #elif dataset == 'val_3rd': # img_list = glob.glob('/mnt/lustrenew/dongrunmin/dongrunmin/SR/RefSR/github_pub_round1/dataset/val/val_3rd/LR/*') # HR_path = '/mnt/lustrenew/dongrunmin/dongrunmin/SR/RefSR/github_pub_round1/dataset/val/val_3rd/HR' #elif dataset == 'val_4th': # img_list = glob.glob('/mnt/lustrenew/dongrunmin/dongrunmin/SR/RefSR/github_pub_round1/dataset/val/val_4th/LR/*') # HR_path = '/mnt/lustrenew/dongrunmin/dongrunmin/SR/RefSR/github_pub_round1/dataset/val/val_4th/HR' img_list = glob.glob( '/data/zlx/rfsr/github_pub_round1/dataset/val/val_1st/LR/*') HR_path = '/data/zlx/rfsr/github_pub_round1/dataset/val/val_1st/HR' util.mkdir_and_rename(osp.join(save_path, '{}_exp{}'.format(dataset, exp_name))) PSNR_avg = 0 SSIM_avg = 0 stat_time = 0 for ii, img_path in enumerate(sorted(img_list)): # if '160' in img_path: base_name = osp.splitext(osp.basename(img_path))[0] use_name = base_name + '.jpg' img_GT = cv2.imread(osp.join(HR_path, use_name)) img_LR = cv2.imread(img_path) / 255. #img_LR = img_LR[:, :, [2, 1, 0]]
def main(): #### options parser = argparse.ArgumentParser() parser.add_argument('-opt', type=str, help='Path to option YMAL file.') parser.add_argument('--launcher', choices=['none', 'pytorch'], default='none', help='job launcher') parser.add_argument('--local_rank', type=int, default=0) args = parser.parse_args() opt = option.parse(args.opt, is_train=True) #### distributed training settings if args.launcher == 'none': # disabled distributed training opt['dist'] = False rank = -1 print('Disabled distributed training.') else: opt['dist'] = True init_dist() world_size = torch.distributed.get_world_size() rank = torch.distributed.get_rank() #### loading resume state if exists if opt['path'].get('resume_state', None): # distributed resuming: all load into default GPU device_id = torch.cuda.current_device() resume_state = torch.load( opt['path']['resume_state'], map_location=lambda storage, loc: storage.cuda(device_id)) option.check_resume(opt, resume_state['iter']) # check resume options else: resume_state = None #### mkdir and loggers if rank <= 0: # normal training (rank -1) OR distributed training (rank 0) if resume_state is None: util.mkdir_and_rename( opt['path'] ['experiments_root']) # rename experiment folder if exists util.mkdirs( (path for key, path in opt['path'].items() if not key == 'experiments_root' and 'pretrain_model' not in key and 'resume' not in key)) # config loggers. Before it, the log will not work util.setup_logger('base', opt['path']['log'], 'train_' + opt['name'], level=logging.INFO, screen=True, tofile=True) util.setup_logger('val', opt['path']['log'], 'val_' + opt['name'], level=logging.INFO, screen=True, tofile=True) logger = logging.getLogger('base') logger.info(option.dict2str(opt)) # tensorboard logger if opt['use_tb_logger'] and 'debug' not in opt['name']: version = float(torch.__version__[0:3]) if version >= 1.1: # PyTorch 1.1 from torch.utils.tensorboard import SummaryWriter else: logger.info( 'You are using PyTorch {}. Tensorboard will use [tensorboardX]' .format(version)) from tensorboardX import SummaryWriter tb_logger = SummaryWriter(log_dir='../tb_logger/' + opt['name']) else: util.setup_logger('base', opt['path']['log'], 'train', level=logging.INFO, screen=True) logger = logging.getLogger('base') # convert to NoneDict, which returns None for missing keys opt = option.dict_to_nonedict(opt) #### random seed seed = opt['train']['manual_seed'] if seed is None: seed = random.randint(1, 10000) if rank <= 0: logger.info('Random seed: {}'.format(seed)) util.set_random_seed(seed) torch.backends.cudnn.benchmark = True # torch.backends.cudnn.deterministic = True #### create train and val dataloader dataset_ratio = 200 # enlarge the size of each epoch for phase, dataset_opt in opt['datasets'].items(): if phase == 'train': train_set = create_dataset(dataset_opt) train_size = int( math.ceil(len(train_set) / dataset_opt['batch_size'])) total_iters = int(opt['train']['niter']) total_epochs = int(math.ceil(total_iters / train_size)) if opt['dist']: train_sampler = DistIterSampler(train_set, world_size, rank, dataset_ratio) total_epochs = int( math.ceil(total_iters / (train_size * dataset_ratio))) else: train_sampler = None train_loader = create_dataloader(train_set, dataset_opt, opt, train_sampler) if rank <= 0: logger.info( 'Number of train images: {:,d}, iters: {:,d}'.format( len(train_set), train_size)) logger.info('Total epochs needed: {:d} for iters {:,d}'.format( total_epochs, total_iters)) elif phase == 'val': val_set = create_dataset(dataset_opt) val_loader = create_dataloader(val_set, dataset_opt, opt, None) if rank <= 0: logger.info('Number of val images in [{:s}]: {:d}'.format( dataset_opt['name'], len(val_set))) else: raise NotImplementedError( 'Phase [{:s}] is not recognized.'.format(phase)) assert train_loader is not None #### create model model = create_model(opt) #### resume training if resume_state: logger.info('Resuming training from epoch: {}, iter: {}.'.format( resume_state['epoch'], resume_state['iter'])) start_epoch = resume_state['epoch'] current_step = resume_state['iter'] model.resume_training(resume_state) # handle optimizers and schedulers else: current_step = 0 start_epoch = 0 #### training logger.info('Start training from epoch: {:d}, iter: {:d}'.format( start_epoch, current_step)) for epoch in range(start_epoch, total_epochs + 1): if opt['dist']: train_sampler.set_epoch(epoch) for _, train_data in enumerate(train_loader): current_step += 1 if current_step > total_iters: break #### update learning rate model.update_learning_rate(current_step, warmup_iter=opt['train']['warmup_iter']) #### training model.feed_data(train_data) model.optimize_parameters(current_step) #### log if current_step % opt['logger']['print_freq'] == 0: logs = model.get_current_log() message = '<epoch:{:3d}, iter:{:8,d}, lr:{:.3e}> '.format( epoch, current_step, model.get_current_learning_rate()) for k, v in logs.items(): message += '{:s}: {:.4e} '.format(k, v) # tensorboard logger if opt['use_tb_logger'] and 'debug' not in opt['name']: if rank <= 0: tb_logger.add_scalar(k, v, current_step) if rank <= 0: logger.info(message) # validation if current_step % opt['train']['val_freq'] == 0 and rank <= 0: avg_psnr = 0.0 idx = 0 for val_data in val_loader: idx += 1 img_name = os.path.splitext( os.path.basename(val_data['LQ_path'][0]))[0] img_dir = os.path.join(opt['path']['val_images'], img_name) util.mkdir(img_dir) model.feed_data(val_data) model.test() visuals = model.get_current_visuals() sr_img = util.tensor2img(visuals['SR']) # uint8 gt_img = util.tensor2img(visuals['GT']) # uint8 # Save SR images for reference save_img_path = os.path.join( img_dir, '{:s}_{:d}.png'.format(img_name, current_step)) util.save_img(sr_img, save_img_path) # calculate PSNR crop_size = opt['scale'] gt_img = gt_img / 255. sr_img = sr_img / 255. cropped_sr_img = sr_img[crop_size:-crop_size, crop_size:-crop_size, :] cropped_gt_img = gt_img[crop_size:-crop_size, crop_size:-crop_size, :] avg_psnr += util.calculate_psnr(cropped_sr_img * 255, cropped_gt_img * 255) avg_psnr = avg_psnr / idx # log logger.info('# Validation # PSNR: {:.4e}'.format(avg_psnr)) logger_val = logging.getLogger('val') # validation logger logger_val.info( '<epoch:{:3d}, iter:{:8,d}> psnr: {:.4e}'.format( epoch, current_step, avg_psnr)) # tensorboard logger if opt['use_tb_logger'] and 'debug' not in opt['name']: tb_logger.add_scalar('psnr', avg_psnr, current_step) #### save models and training states if current_step % opt['logger']['save_checkpoint_freq'] == 0: if rank <= 0: logger.info('Saving models and training states.') model.save(current_step) model.save_training_state(epoch, current_step) if rank <= 0: logger.info('Saving the final model.') model.save('latest') logger.info('End of training.')
def main(): # options parser = argparse.ArgumentParser() parser.add_argument('-opt', type=str, required=True, help='Path to option JSON file.') parser.add_argument('-name', type=str, help='Name of the experiment.') parser.add_argument('-seed', type=int, default=2, help='Random seed.') parser.add_argument('-gsize', type=int, default=1, help='Num of devices per task.') parser.add_argument('-val_freq', type=int, help='Num of transfer frequency.') parser.add_argument('-train', action='store_true', help='Specify if training.') parser.add_argument( '-lmdb', action='store_true', help='create or use lmdb datasets for accelerating I/O.') parser.add_argument('-droot', type=str, help='dataroot if need specify.') parser.add_argument('-rancl', action='store_true', help='random task for transfer?') parser.add_argument('-create_val', action='store_true', help='create validation set from training set') parser.add_argument('-model', type=str, help='specify model to use') parser.add_argument('-k', action='store_true', help='Use knowledge distillation based transfer.') parser.add_argument('-a', action='store_true', help='Use attention based transfer.') parser.add_argument('-f', action='store_true', help='Use fsp matrix based transfer.') parser.add_argument('-w', action='store_true', help='Use weights transfer.') parser.add_argument('-ws', action='store_true', help='Use weights statistical transfer.') parser.add_argument('-VL', action='store_true', help='Variate on loss function.') parser.add_argument('-VO', action='store_true', help='Variate on optimization algorithm.') parser.add_argument('-VH', action='store_true', help='Variate on hyperparameters.') parser.add_argument('-VD', action='store_true', help='Variate on datasets.') parser.add_argument('-VF', action='store_true', help='Variate on cross validation.') parser.add_argument('-VS', type=float, help='fraction of bagging dataset.') opt = parser.parse_args() if not opt.train and opt.gsize > 1: warnings.warn( '1 device for validation is enough! The gsize will be reset to 1') opt.gsize = 1 if not opt.w and opt.gsize > 1: warnings.warn( 'The exploit-explorer mode is not activated! The gsize will be reset to 1' ) opt.gsize = 1 comm = MPI.COMM_WORLD rank = comm.Get_rank() gpus_per_node = torch.cuda.device_count() if gpus_per_node == 0: device_id = 'cpu' else: device_id = rank % gpus_per_node print('start running! {} gpus are in use'.format(gpus_per_node)) print('Torch version {}'.format(torch.__version__)) # run to here is fine!! # torch.cuda.set_device(device_id) device_per_task = opt.gsize task_id = rank // device_per_task task_exploit_rank = rank - rank % device_per_task world_size = comm.Get_size() opt.ntasks = world_size opt = option.parse(opt, task_id) opt = option.dict_to_nonedict( opt) # Convert to NoneDict, which return None for missing key. # Setup directory if not opt['resume'] and task_exploit_rank == rank and not opt['path'][ 'resume_state']: try: if opt['is_train'] and rank == 0: util.mkdir_and_rename( opt['path'] ['experiments_root']) # rename old folder if exists except FileNotFoundError: raise FileNotFoundError("Issue from task {} and rank {}".format( task_id, rank)) comm.Barrier() if not opt['resume'] and task_exploit_rank == rank and not opt['path'][ 'resume_state']: util.mkdirs((path for key, path in opt['path'].items() if key not in [ 'experiments_root', 'log', 'root', 'pretrain_model', 'pretrain_model_G', 'resume_state', 'data_config' ])) # save json file to task folder if opt['is_train']: with open(os.path.join(opt['path']['task'], 'task.json'), 'w') as outfile: json.dump(opt, outfile, indent=4, ensure_ascii=False) comm.Barrier() # config loggers. Before it, the log will not work util.setup_logger(str(task_id), opt['path']['log'], 'train', level=logging.INFO, screen=True, task_id=task_id, rank=rank) logger = logging.getLogger(str(task_id)) if task_exploit_rank == rank: if task_id == 0: logger.info(option.dict2str(opt)) # display options else: logger.info( 'Auxiliary task {} configuration: network: {}, optim: {}, loss: {}, data: {}' .format(task_id, opt['network'], opt['train']['optim'], opt['train']['loss'], opt['datasets']['train']['name'])) # tensorboard logger tb_logger = None if opt['use_tb_logger'] and opt[ 'is_train'] and task_exploit_rank == rank: # and 'debug' not in opt['name'] from tensorboardX import SummaryWriter if rank == 0: util.mkdir_and_rename(os.path.join('../tb_logger/', opt['name'])) comm.barrier() tb_logger_path = os.path.join('../tb_logger/', opt['name'], str(task_id)) tb_logger = SummaryWriter(log_dir=tb_logger_path) # create task # If you don't use the following setting, the results will still be non-deterministic. # However, if you set them so, the speed may be slower, depending on models. torch.backends.cudnn.benchmark = False torch.backends.cudnn.deterministic = True task = create_task(opt, comm=comm, device=device_id) # Start training or testing start = time() if opt['is_train']: task.resume_training() task.full_training(opt['logger']['print_freq'], opt['val_freq'], opt['logger']['save_checkpoint_freq'], tb_logger) if rank == 0: if opt['varyOnCV']: all_hmean = task.cv_validation() results = [r['acc'] for r in all_hmean] idx = np.argmax(results) best_task_id = all_hmean[idx]['id'] best_task_rank = device_per_task * best_task_id logger.info( 'All hmean is {} and select rank {} to run validation'. format(all_hmean, best_task_rank)) if task_exploit_rank == best_task_rank: task.validation(verbose=True, report=True, split='test', best=True) else: task.validation(verbose=True, report=True, split='test', best=True) logger.info('End of training.') else: logger.info('Task {} start validation'.format(task_id)) task.validation(verbose=True, report=True, split='test', best=True) duration = time() - start logger.info('The program tasks time {}'.format(duration))
def main(): # options parser = argparse.ArgumentParser() parser.add_argument("-opt", type=str, help="Path to option YAML file.") parser.add_argument("--launcher", choices=["none", "pytorch"], default="none", help="job launcher") parser.add_argument("--local_rank", type=int, default=0) args = parser.parse_args() opt = option.parse(args.opt, is_train=True) # distributed training settings if args.launcher == "none": # disabled distributed training opt["dist"] = False rank = -1 print("Disabled distributed training.") else: opt["dist"] = True init_dist() world_size = torch.distributed.get_world_size() rank = torch.distributed.get_rank() # loading resume state if exists if opt["path"].get("resume_state", None): # distributed resuming: all load into default GPU device_id = torch.cuda.current_device() resume_state = torch.load( opt["path"]["resume_state"], map_location=lambda storage, loc: storage.cuda(device_id)) option.check_resume(opt, resume_state["iter"]) # check resume options else: resume_state = None # mkdir and loggers if rank <= 0: # normal training (rank -1) OR distributed training (rank 0) if resume_state is None: util.mkdir_and_rename( opt["path"] ["experiments_root"]) # rename experiment folder if exists util.mkdirs( (path for key, path in opt["path"].items() if not key == "experiments_root" and "pretrain_model" not in key and "resume" not in key)) # config loggers. Before it, the log will not work util.setup_logger("base", opt["path"]["log"], "train_" + opt["name"], level=logging.INFO, screen=True, tofile=True) logger = logging.getLogger("base") logger.info(option.dict2str(opt)) # tensorboard logger if opt["use_tb_logger"] and "debug" not in opt["name"]: version = float(torch.__version__[0:3]) if version >= 1.1: # PyTorch 1.1 from torch.utils.tensorboard import SummaryWriter else: logger.info("You are using PyTorch {}. \ Tensorboard will use [tensorboardX]".format( version)) from tensorboardX import SummaryWriter tb_logger = SummaryWriter(log_dir="../tb_logger/" + opt["name"]) else: util.setup_logger("base", opt["path"]["log"], "train", level=logging.INFO, screen=True) logger = logging.getLogger("base") # convert to NoneDict, which returns None for missing keys opt = option.dict_to_nonedict(opt) # random seed seed = opt["train"]["manual_seed"] if seed is None: seed = random.randint(1, 10000) if rank <= 0: logger.info("Random seed: {}".format(seed)) util.set_random_seed(seed) torch.backends.cudnn.benchmark = True # torch.backends.cudnn.deterministic = True # create train and val dataloader dataset_ratio = 200 # enlarge the size of each epoch for phase, dataset_opt in opt["datasets"].items(): if phase == "train": train_set = create_dataset(dataset_opt) train_size = int( math.ceil(len(train_set) / dataset_opt["batch_size"])) total_iters = int(opt["train"]["niter"]) total_epochs = int(math.ceil(total_iters / train_size)) if opt["dist"]: train_sampler = DistIterSampler(train_set, world_size, rank, dataset_ratio) total_epochs = int( math.ceil(total_iters / (train_size * dataset_ratio))) else: train_sampler = None train_loader = create_dataloader(train_set, dataset_opt, opt, train_sampler) if rank <= 0: logger.info( "Number of train images: {:,d}, iters: {:,d}".format( len(train_set), train_size)) logger.info("Total epochs needed: {:d} for iters {:,d}".format( total_epochs, total_iters)) elif phase == "val": val_set = create_dataset(dataset_opt) val_loader = create_dataloader(val_set, dataset_opt, opt, None) if rank <= 0: logger.info("Number of val images in [{:s}]: {:d}".format( dataset_opt["name"], len(val_set))) else: raise NotImplementedError( "Phase [{:s}] is not recognized.".format(phase)) assert train_loader is not None # create model model = create_model(opt) print("Model created!") # resume training if resume_state: logger.info("Resuming training from epoch: {}, iter: {}.".format( resume_state["epoch"], resume_state["iter"])) start_epoch = resume_state["epoch"] current_step = resume_state["iter"] model.resume_training(resume_state) # handle optimizers and schedulers else: current_step = 0 start_epoch = 0 # training logger.info("Start training from epoch: {:d}, iter: {:d}".format( start_epoch, current_step)) for epoch in range(start_epoch, total_epochs + 1): if opt["dist"]: train_sampler.set_epoch(epoch) for _, train_data in enumerate(train_loader): current_step += 1 if current_step > total_iters: break # update learning rate model.update_learning_rate(current_step, warmup_iter=opt["train"]["warmup_iter"]) # training model.feed_data(train_data) model.optimize_parameters(current_step) # log if current_step % opt["logger"]["print_freq"] == 0: logs = model.get_current_log() message = "[epoch:{:3d}, iter:{:8,d}, lr:(".format( epoch, current_step) for v in model.get_current_learning_rate(): message += "{:.3e},".format(v) message += ")] " for k, v in logs.items(): message += "{:s}: {:.4e} ".format(k, v) # tensorboard logger if opt["use_tb_logger"] and "debug" not in opt["name"]: if rank <= 0: tb_logger.add_scalar(k, v, current_step) if rank <= 0: logger.info(message) # validation if opt["datasets"].get( "val", None) and current_step % opt["train"]["val_freq"] == 0: # image restoration validation if opt["model"] in ["sr", "srgan"] and rank <= 0: # does not support multi-GPU validation pbar = util.ProgressBar(len(val_loader)) avg_psnr = 0.0 idx = 0 for val_data in val_loader: idx += 1 img_name = os.path.splitext( os.path.basename(val_data["LQ_path"][0]))[0] img_dir = os.path.join(opt["path"]["val_images"], img_name) util.mkdir(img_dir) model.feed_data(val_data) model.test() visuals = model.get_current_visuals() sr_img = util.tensor2img(visuals["rlt"]) # uint8 gt_img = util.tensor2img(visuals["GT"]) # uint8 # Save SR images for reference save_img_path = os.path.join( img_dir, "{:s}_{:d}.png".format(img_name, current_step)) util.save_img(sr_img, save_img_path) # calculate PSNR sr_img, gt_img = util.crop_border([sr_img, gt_img], opt["scale"]) avg_psnr += util.calculate_psnr(sr_img, gt_img) pbar.update("Test {}".format(img_name)) avg_psnr = avg_psnr / idx # log logger.info("# Validation # PSNR: {:.4e}".format(avg_psnr)) # tensorboard logger if opt["use_tb_logger"] and "debug" not in opt["name"]: tb_logger.add_scalar("psnr", avg_psnr, current_step) else: # video restoration validation if opt["dist"]: # multi-GPU testing psnr_rlt = {} # with border and center frames if rank == 0: pbar = util.ProgressBar(len(val_set)) for idx in range(rank, len(val_set), world_size): val_data = val_set[idx] val_data["LQs"].unsqueeze_(0) val_data["GT"].unsqueeze_(0) folder = val_data["folder"] idx_d, max_idx = val_data["idx"].split("/") idx_d, max_idx = int(idx_d), int(max_idx) if psnr_rlt.get(folder, None) is None: psnr_rlt[folder] = torch.zeros( max_idx, dtype=torch.float32, device="cuda") model.feed_data(val_data) model.test() visuals = model.get_current_visuals() rlt_img = util.tensor2img(visuals["rlt"]) # uint8 gt_img = util.tensor2img(visuals["GT"]) # uint8 # calculate PSNR psnr_rlt[folder][idx_d] = util.calculate_psnr( rlt_img, gt_img) if rank == 0: for _ in range(world_size): pbar.update("Test {} - {}/{}".format( folder, idx_d, max_idx)) # collect data for _, v in psnr_rlt.items(): dist.reduce(v, 0) dist.barrier() if rank == 0: psnr_rlt_avg = {} psnr_total_avg = 0.0 for k, v in psnr_rlt.items(): psnr_rlt_avg[k] = torch.mean(v).cpu().item() psnr_total_avg += psnr_rlt_avg[k] psnr_total_avg /= len(psnr_rlt) log_s = "# Validation # PSNR: {:.4e}:".format( psnr_total_avg) for k, v in psnr_rlt_avg.items(): log_s += " {}: {:.4e}".format(k, v) logger.info(log_s) if opt["use_tb_logger"] and "debug" not in opt[ "name"]: tb_logger.add_scalar("psnr_avg", psnr_total_avg, current_step) for k, v in psnr_rlt_avg.items(): tb_logger.add_scalar(k, v, current_step) else: pbar = util.ProgressBar(len(val_loader)) psnr_rlt = {} # with border and center frames psnr_rlt_avg = {} psnr_total_avg = 0.0 for val_data in val_loader: folder = val_data["folder"][0] idx_d, max_id = val_data["idx"][0].split("/") # border = val_data['border'].item() if psnr_rlt.get(folder, None) is None: psnr_rlt[folder] = [] model.feed_data(val_data) model.test() visuals = model.get_current_visuals() rlt_img = util.tensor2img(visuals["rlt"]) # uint8 gt_img = util.tensor2img(visuals["GT"]) # uint8 lq_img = util.tensor2img(visuals["LQ"][2]) # uint8 img_dir = opt["path"]["val_images"] util.mkdir(img_dir) save_img_path = os.path.join( img_dir, "{}.png".format(idx_d)) util.save_img(np.hstack((lq_img, rlt_img, gt_img)), save_img_path) # calculate PSNR psnr = util.calculate_psnr(rlt_img, gt_img) psnr_rlt[folder].append(psnr) pbar.update("Test {} - {}".format(folder, idx_d)) for k, v in psnr_rlt.items(): psnr_rlt_avg[k] = sum(v) / len(v) psnr_total_avg += psnr_rlt_avg[k] psnr_total_avg /= len(psnr_rlt) log_s = "# Validation # PSNR: {:.4e}:".format( psnr_total_avg) for k, v in psnr_rlt_avg.items(): log_s += " {}: {:.4e}".format(k, v) logger.info(log_s) if opt["use_tb_logger"] and "debug" not in opt["name"]: tb_logger.add_scalar("psnr_avg", psnr_total_avg, current_step) for k, v in psnr_rlt_avg.items(): tb_logger.add_scalar(k, v, current_step) # save models and training states if current_step % opt["logger"]["save_checkpoint_freq"] == 0: if rank <= 0: logger.info("Saving models and training states.") model.save(current_step) model.save_training_state(epoch, current_step) if rank <= 0: logger.info("Saving the final model.") model.save("latest") logger.info("End of training.") tb_logger.close()
def main(): #### setup options of three networks parser = argparse.ArgumentParser() parser.add_argument("-opt", type=str, help="Path to option YMAL file of Predictor.") parser.add_argument("--launcher", choices=["none", "pytorch"], default="none", help="job launcher") parser.add_argument("--local_rank", type=int, default=0) args = parser.parse_args() opt = option.parse(args.opt, is_train=True) # convert to NoneDict, which returns None for missing keys opt = option.dict_to_nonedict(opt) # choose small opt for SFTMD test, fill path of pre-trained model_F #### set random seed seed = opt["train"]["manual_seed"] if seed is None: seed = random.randint(1, 10000) util.set_random_seed(seed) # load PCA matrix of enough kernel print("load PCA matrix") pca_matrix = torch.load(opt["pca_matrix_path"], map_location=lambda storage, loc: storage) print("PCA matrix shape: {}".format(pca_matrix.shape)) #### distributed training settings if args.launcher == "none": # disabled distributed training opt["dist"] = False opt["dist"] = False rank = -1 print("Disabled distributed training.") else: opt["dist"] = True opt["dist"] = True init_dist() world_size = ( torch.distributed.get_world_size() ) # Returns the number of processes in the current process group rank = torch.distributed.get_rank( ) # Returns the rank of current process group torch.backends.cudnn.benchmark = True # torch.backends.cudnn.deterministic = True ###### Predictor&Corrector train ###### #### loading resume state if exists if opt["path"].get("resume_state", None): # distributed resuming: all load into default GPU device_id = torch.cuda.current_device() resume_state = torch.load( opt["path"]["resume_state"], map_location=lambda storage, loc: storage.cuda(device_id), ) option.check_resume(opt, resume_state["iter"]) # check resume options else: resume_state = None #### mkdir and loggers if rank <= 0: # normal training (rank -1) OR distributed training (rank 0-7) if resume_state is None: # Predictor path util.mkdir_and_rename( opt["path"] ["experiments_root"]) # rename experiment folder if exists util.mkdirs( (path for key, path in opt["path"].items() if not key == "experiments_root" and "pretrain_model" not in key and "resume" not in key)) os.system("rm ./log") os.symlink(os.path.join(opt["path"]["experiments_root"], ".."), "./log") # config loggers. Before it, the log will not work util.setup_logger( "base", opt["path"]["log"], "train_" + opt["name"], level=logging.INFO, screen=True, tofile=True, ) util.setup_logger( "val", opt["path"]["log"], "val_" + opt["name"], level=logging.INFO, screen=True, tofile=True, ) logger = logging.getLogger("base") logger.info(option.dict2str(opt)) # tensorboard logger if opt["use_tb_logger"] and "debug" not in opt["name"]: version = float(torch.__version__[0:3]) if version >= 1.1: # PyTorch 1.1 from torch.utils.tensorboard import SummaryWriter else: logger.info( "You are using PyTorch {}. Tensorboard will use [tensorboardX]" .format(version)) from tensorboardX import SummaryWriter tb_logger = SummaryWriter(log_dir="log/tb_logger/" + opt["name"]) else: util.setup_logger("base", opt["path"]["log"], "train", level=logging.INFO, screen=True) logger = logging.getLogger("base") torch.backends.cudnn.benchmark = True # torch.backends.cudnn.deterministic = True #### create train and val dataloader dataset_ratio = 200 # enlarge the size of each epoch for phase, dataset_opt in opt["datasets"].items(): if phase == "train": train_set = create_dataset(dataset_opt) train_size = int( math.ceil(len(train_set) / dataset_opt["batch_size"])) total_iters = int(opt["train"]["niter"]) total_epochs = int(math.ceil(total_iters / train_size)) if opt["dist"]: train_sampler = DistIterSampler(train_set, world_size, rank, dataset_ratio) total_epochs = int( math.ceil(total_iters / (train_size * dataset_ratio))) else: train_sampler = None train_loader = create_dataloader(train_set, dataset_opt, opt, train_sampler) if rank <= 0: logger.info( "Number of train images: {:,d}, iters: {:,d}".format( len(train_set), train_size)) logger.info("Total epochs needed: {:d} for iters {:,d}".format( total_epochs, total_iters)) elif phase == "val": val_set = create_dataset(dataset_opt) val_loader = create_dataloader(val_set, dataset_opt, opt, None) if rank <= 0: logger.info("Number of val images in [{:s}]: {:d}".format( dataset_opt["name"], len(val_set))) else: raise NotImplementedError( "Phase [{:s}] is not recognized.".format(phase)) assert train_loader is not None assert val_loader is not None #### create model model = create_model(opt) # load pretrained model of SFTMD #### resume training if resume_state: logger.info("Resuming training from epoch: {}, iter: {}.".format( resume_state["epoch"], resume_state["iter"])) start_epoch = resume_state["epoch"] current_step = resume_state["iter"] model.resume_training(resume_state) # handle optimizers and schedulers else: current_step = 0 start_epoch = 0 prepro = util.SRMDPreprocessing( opt["scale"], pca_matrix, random=True, para_input=opt["code_length"], kernel=opt["kernel_size"], noise=False, cuda=True, sig=None, sig_min=opt["sig_min"], sig_max=opt["sig_max"], rate_iso=1.0, scaling=3, rate_cln=0.2, noise_high=0.0, ) #### training logger.info("Start training from epoch: {:d}, iter: {:d}".format( start_epoch, current_step)) for epoch in range(start_epoch, total_epochs + 1): if opt["dist"]: train_sampler.set_epoch(epoch) for _, train_data in enumerate(train_loader): current_step += 1 if current_step > total_iters: break #### preprocessing for LR_img and kernel map LR_img, ker_map = prepro(train_data["GT"]) LR_img = (LR_img * 255).round() / 255 #### training Predictor model.feed_data(LR_img, train_data["GT"], ker_map) model.optimize_parameters(current_step) model.update_learning_rate(current_step, warmup_iter=opt["train"]["warmup_iter"]) visuals = model.get_current_visuals() #### log of model_P if current_step % opt["logger"]["print_freq"] == 0: logs = model.get_current_log() message = "Predictor <epoch:{:3d}, iter:{:8,d}, lr:{:.3e}> ".format( epoch, current_step, model.get_current_learning_rate()) for k, v in logs.items(): message += "{:s}: {:.4e} ".format(k, v) # tensorboard logger if opt["use_tb_logger"] and "debug" not in opt["name"]: if rank <= 0: tb_logger.add_scalar(k, v, current_step) if rank <= 0: logger.info(message) # validation, to produce ker_map_list(fake) if current_step % opt["train"]["val_freq"] == 0 and rank <= 0: avg_psnr = 0.0 idx = 0 for _, val_data in enumerate(val_loader): # LR_img, ker_map = prepro(val_data['GT']) LR_img = val_data["LQ"] lr_img = util.tensor2img( LR_img) # save LR image for reference # valid Predictor model.feed_data(LR_img, val_data["GT"]) model.test() visuals = model.get_current_visuals() # Save images for reference img_name = os.path.splitext( os.path.basename(val_data["LQ_path"][0]))[0] img_dir = os.path.join(opt["path"]["val_images"], img_name) # img_dir = os.path.join(opt['path']['val_images'], str(current_step), '_', str(step)) util.mkdir(img_dir) save_lr_path = os.path.join(img_dir, "{:s}_LR.png".format(img_name)) util.save_img(lr_img, save_lr_path) sr_img = util.tensor2img(visuals["SR"]) # uint8 gt_img = util.tensor2img(visuals["GT"]) # uint8 save_img_path = os.path.join( img_dir, "{:s}_{:d}.png".format(img_name, current_step)) util.save_img(sr_img, save_img_path) # calculate PSNR crop_size = opt["scale"] gt_img = gt_img / 255.0 sr_img = sr_img / 255.0 cropped_sr_img = sr_img[crop_size:-crop_size, crop_size:-crop_size, :] cropped_gt_img = gt_img[crop_size:-crop_size, crop_size:-crop_size, :] avg_psnr += util.calculate_psnr(cropped_sr_img * 255, cropped_gt_img * 255) idx += 1 avg_psnr = avg_psnr / idx # log logger.info("# Validation # PSNR: {:.6f}".format(avg_psnr)) logger_val = logging.getLogger("val") # validation logger logger_val.info( "<epoch:{:3d}, iter:{:8,d}, psnr: {:.6f}".format( epoch, current_step, avg_psnr)) # tensorboard logger if opt["use_tb_logger"] and "debug" not in opt["name"]: tb_logger.add_scalar("psnr", avg_psnr, current_step) #### save models and training states if current_step % opt["logger"]["save_checkpoint_freq"] == 0: if rank <= 0: logger.info("Saving models and training states.") model.save(current_step) model.save_training_state(epoch, current_step) if rank <= 0: logger.info("Saving the final model.") model.save("latest") logger.info("End of Predictor and Corrector training.") tb_logger.close()
def parse(opt_path): # remove comments starting with '//' json_str = '' with open(opt_path, 'r') as f: for line in f: line = line.split('//')[0] + '\n' json_str += line opt = json.loads(json_str, object_pairs_hook=OrderedDict) opt['timestamp'] = get_timestamp() scale = opt['scale'] rgb_range = opt['rgb_range'] # export CUDA_VISIBLE_DEVICES # Need do it before import torch if opt['gpu_ids'] is not None: gpu_list = ','.join(str(x) for x in opt['gpu_ids']) os.environ['CUDA_VISIBLE_DEVICES'] = gpu_list print('===> Export CUDA_VISIBLE_DEVICES = [' + gpu_list + ']') else: print('===> CPU mode is set (NOTE: GPU is recommended)') # datasets for phase, dataset in opt['datasets'].items(): phase = phase.split('_')[0] dataset['phase'] = phase dataset['scale'] = scale dataset['rgb_range'] = rgb_range # for network initialize opt['networks']['scale'] = opt['scale'] network_opt = opt['networks'] config_str = '%s_in%df%d_x%d' % (network_opt['which_model'].upper(), network_opt['in_channels'], network_opt['num_features'], opt['scale']) exp_path = os.path.join(os.getcwd(), 'experiments_round2', config_str) if opt['is_train'] and opt['solver']['pretrain']: if 'pretrained_path' not in list(opt['solver'].keys()): raise ValueError( "[Error] The 'pretrained_path' does not declarate in *.json") exp_path = os.path.dirname( os.path.dirname(opt['solver']['pretrained_path'])) if opt['solver']['pretrain'] == 'finetune': exp_path += '_finetune' #exp_path = os.path.relpath(exp_path) path_opt = OrderedDict() path_opt['exp_root'] = exp_path path_opt['epochs'] = os.path.join(exp_path, 'epochs') path_opt['visual'] = os.path.join(exp_path, 'visual') path_opt['records'] = os.path.join(exp_path, 'records') opt['path'] = path_opt if opt['is_train']: # create folders if opt['solver']['pretrain'] == 'resume': opt = dict_to_nonedict(opt) else: util.mkdir_and_rename( opt['path']['exp_root']) # rename old experiments if exists util.mkdirs((path for key, path in opt['path'].items() if not key == 'exp_root')) save(opt) opt = dict_to_nonedict(opt) print("===> Experimental DIR: [%s]" % exp_path) return opt
def main(): parser = argparse.ArgumentParser() parser.add_argument('-opt', type=str, default='options/train/train_ESRCNN_S2self.json', help='Path to option JSON file.') opt = option.parse(parser.parse_args().opt, is_train=True) opt = option.dict_to_nonedict(opt) if opt['path']['resume_state']: resume_state = torch.load(opt['path']['resume_state']) else: resume_state = None util.mkdir_and_rename(opt['path']['experiments_root']) util.mkdirs((path for key, path in opt['path'].items() if not key == 'experiments_root' and 'pretrain_model' not in key and 'resume' not in key)) util.setup_logger(None, opt['path']['log'], 'train', level=logging.INFO, screen=True) util.setup_logger('val', opt['path']['log'], 'val', level=logging.INFO) logger = logging.getLogger('base') if resume_state: logger.info('Resuming training from epoch: {}, iter: {}.'.format( resume_state['epoch'], resume_state['iter'])) option.check_resume(opt) logger.info(option.dict2str(opt)) if opt['use_tb_logger'] and 'debug' not in opt['name']: from tensorboardX import SummaryWriter tb_logger = SummaryWriter(log_dir='./tb_logger/' + opt['name']) seed = opt['train']['manual_seed'] if seed is None: seed = random.randint(1, 10000) logger.info('Random seed: {}'.format(seed)) util.set_random_seed(seed) torch.backends.cudnn.benckmark = True # Setup TrainDataLoader trainloader = DataLoader(opt['datasets']['train']['dataroot'], split='train') train_size = int( math.ceil(len(trainloader) / opt['datasets']['train']['batch_size'])) logger.info('Number of train images: {:,d}, iters: {:,d}'.format( len(trainloader), train_size)) total_iters = int(opt['train']['niter']) total_epochs = int(math.ceil(total_iters / train_size)) logger.info('Total epochs needed: {:d} for iters {:,d}'.format( total_epochs, total_iters)) TrainDataLoader = data.DataLoader( trainloader, batch_size=opt['datasets']['train']['batch_size'], num_workers=12, shuffle=True) #Setup for validate valloader = DataLoader(opt['datasets']['train']['dataroot'], split='val') VALDataLoader = data.DataLoader( valloader, batch_size=opt['datasets']['train']['batch_size'] // 5, num_workers=1, shuffle=True) logger.info('Number of val images:{:d}'.format(len(valloader))) # Setup Model model = get_model('esrcnn_s2self', opt) if resume_state: start_epoch = resume_state['epoch'] current_step = resume_state['iter'] model.resume_training(resume_state) else: current_step = 0 start_epoch = 0 logger.info('Start training from epoch: {:d}, iter: {:d}'.format( start_epoch, current_step)) for epoch in range(start_epoch, total_epochs): for i, train_data in enumerate(TrainDataLoader): current_step += 1 if current_step > total_iters: break model.update_learning_rate() model.feed_data(train_data) model.optimize_parameters(current_step) if current_step % opt['logger']['print_freq'] == 0: logs = model.get_current_log() message = '<epoch:{:3d}, iter:{:8,d}, lr:{:.3e}>'.format( epoch, current_step, model.get_current_learning_rate()) for k, v in logs.items(): message += '{:s}: {:.4e} '.format(k, v[0]) if opt['use_tb_logger'] and 'debug' not in opt['name']: tb_logger.add_scalar(k, v[0], current_step) logger.info(message) if current_step % opt['train']['val_freq'] == 0: avg_psnr = 0.0 idx = 0 for i_val, val_data in enumerate(VALDataLoader): idx += 1 img_name = val_data[3][0].split('.')[0] model.feed_data(val_data) model.val() visuals = model.get_current_visuals() pred_img = util.tensor2img(visuals['Pred']) gt_img = util.tensor2img(visuals['label']) avg_psnr += util.calculate_psnr(pred_img, gt_img) avg_psnr = avg_psnr / idx logger.info('# Validation #PSNR: {:.4e}'.format(avg_psnr)) logger_val = logging.getLogger('val') logger_val.info( '<epoch:{:3d}, iter:{:8,d}> psnr:{:.4e}'.format( epoch, current_step, avg_psnr)) if opt['use_tb_logger'] and 'debug' not in opt['name']: tb_logger.add_scalar('psnr', avg_psnr, current_step) if current_step % opt['logger']['save_checkpoint_freq'] == 0: logger.info('Saving models and training states.') model.save(current_step) model.save_training_state(epoch, current_step) logger.info('Saving the final model.') model.save('latest') logger.info('End of training')