with Timer("Elapsed time in update: %f"): # Main training code trainer.dis_update(images_a, images_b, config) trainer.gen_update(images_a, images_b, config) torch.cuda.synchronize() # Dump training stats in log file if (iterations + 1) % config['log_iter'] == 0: print("Iteration: %08d/%08d" % (iterations + 1, max_iter)) write_loss(iterations, trainer, train_writer) # Write images if (iterations + 1) % config['image_save_iter'] == 0: with torch.no_grad(): test_image_outputs = trainer.sample( test_display_images_a, test_display_images_b) train_image_outputs = trainer.sample( train_display_images_a, train_display_images_b) write_2images(test_image_outputs, display_size, image_directory, 'test_%08d' % (iterations + 1)) write_2images(train_image_outputs, display_size, image_directory, 'train_%08d' % (iterations + 1)) # HTML write_html(output_directory + "/index.html", iterations + 1, config['image_save_iter'], 'images') if (iterations + 1) % config['image_display_iter'] == 0: with torch.no_grad(): image_outputs = trainer.sample(train_display_images_a, train_display_images_b) write_2images(image_outputs, display_size, image_directory,
def main(): parser = argparse.ArgumentParser() parser.add_argument('--config', type=str, default='configs/edges2handbags_folder.yaml', help='Path to the config file.') parser.add_argument('--output_path', type=str, default='.', help="outputs path") #resume option => [, default='730000'] parser.add_argument("--resume", default='150000', action="store_true") parser.add_argument('--trainer', type=str, default='MUNIT', help="MUNIT|UNIT") opts = parser.parse_args() cudnn.benchmark = True # Load experiment setting config = get_config(opts.config) max_iter = config['max_iter'] display_size = config['display_size'] config['vgg_model_path'] = opts.output_path # Setup model and data loader if opts.trainer == 'MUNIT': trainer = MUNIT_Trainer(config) elif opts.trainer == 'UNIT': trainer = UNIT_Trainer(config) else: sys.exit("Only support MUNIT|UNIT") trainer.cuda() train_loader_a, train_loader_b, test_loader_a, test_loader_b = get_all_data_loaders( config) train_display_images_a = torch.stack( [train_loader_a.dataset[i] for i in range(display_size)]).cuda() train_display_images_b = torch.stack( [train_loader_b.dataset[i] for i in range(display_size)]).cuda() test_display_images_a = torch.stack( [test_loader_a.dataset[i] for i in range(display_size)]).cuda() test_display_images_b = torch.stack( [test_loader_b.dataset[i] for i in range(display_size)]).cuda() # Setup logger and output folders model_name = os.path.splitext(os.path.basename(opts.config))[0] train_writer = tensorboardX.SummaryWriter( os.path.join(opts.output_path + "/logs", model_name)) output_directory = os.path.join(opts.output_path + "/outputs", model_name) checkpoint_directory, image_directory = prepare_sub_folder( output_directory) shutil.copy(opts.config, os.path.join( output_directory, 'config.yaml')) # copy config file to output folder # Start training iterations = trainer.resume(checkpoint_directory, hyperparameters=config) if opts.resume else 0 while True: for it, (images_a, images_b) in enumerate(zip(train_loader_a, train_loader_b)): trainer.update_learning_rate() images_a, images_b = images_a.cuda().detach(), images_b.cuda( ).detach() with Timer("Elapsed time in update: %f"): # Main training code trainer.dis_update(images_a, images_b, config) trainer.gen_update(images_a, images_b, config) torch.cuda.synchronize() # Dump training stats in log file if (iterations + 1) % config['log_iter'] == 0: print("Iteration: %08d/%08d" % (iterations + 1, max_iter)) write_loss(iterations, trainer, train_writer) # Write images if (iterations + 1) % config['image_save_iter'] == 0: with torch.no_grad(): test_image_outputs = trainer.sample( test_display_images_a, test_display_images_b) train_image_outputs = trainer.sample( train_display_images_a, train_display_images_b) write_2images(test_image_outputs, display_size, image_directory, 'test_%08d' % (iterations + 1)) write_2images(train_image_outputs, display_size, image_directory, 'train_%08d' % (iterations + 1)) # HTML write_html(output_directory + "/index.html", iterations + 1, config['image_save_iter'], 'images') if (iterations + 1) % config['image_display_iter'] == 0: with torch.no_grad(): image_outputs = trainer.sample(train_display_images_a, train_display_images_b) write_2images(image_outputs, display_size, image_directory, 'train_current') # Save network weights if (iterations + 1) % config['snapshot_save_iter'] == 0: trainer.save(checkpoint_directory, iterations) iterations += 1 if iterations >= max_iter: sys.exit('Finish training')
def main(argv): (opts, args) = parser.parse_args(argv) cudnn.benchmark = True model_name = os.path.splitext(os.path.basename(opts.config))[0] # Load experiment setting config = get_config(opts.config) max_iter = config['max_iter'] display_size = config['display_size'] # Setup model and data loader trainer = MUNIT_Trainer(config) trainer.cuda() train_loader_a, train_loader_b, test_loader_a, test_loader_b = get_all_data_loaders( config) test_display_images_a = Variable(torch.stack( [test_loader_a.dataset[i] for i in range(display_size)]).cuda(), volatile=True) test_display_images_b = Variable(torch.stack( [test_loader_b.dataset[i] for i in range(display_size)]).cuda(), volatile=True) train_display_images_a = Variable(torch.stack( [train_loader_a.dataset[i] for i in range(display_size)]).cuda(), volatile=True) train_display_images_b = Variable(torch.stack( [train_loader_b.dataset[i] for i in range(display_size)]).cuda(), volatile=True) # Setup logger and output folders train_writer = tensorboard.SummaryWriter(os.path.join( opts.log, model_name)) output_directory = os.path.join(opts.outputs, model_name) checkpoint_directory, image_directory = prepare_sub_folder( output_directory) shutil.copy(opts.config, os.path.join( output_directory, 'config.yaml')) # copy config file to output folder # Start training iterations = trainer.resume(checkpoint_directory) if opts.resume else 0 while True: for it, (images_a, images_b) in enumerate(izip(train_loader_a, train_loader_b)): trainer.update_learning_rate() images_a, images_b = Variable(images_a.cuda()), Variable( images_b.cuda()) # Main training code trainer.dis_update(images_a, images_b, config) trainer.gen_update(images_a, images_b, config) # Dump training stats in log file if (iterations + 1) % config['log_iter'] == 0: print("Iteration: %08d/%08d" % (iterations + 1, max_iter)) write_loss(iterations, trainer, train_writer) # Write images if (iterations + 1) % config['image_save_iter'] == 0: # Test set images image_outputs = trainer.sample(test_display_images_a, test_display_images_b) write_images( image_outputs, display_size, '%s/gen_test%08d.jpg' % (image_directory, iterations + 1)) # Train set images image_outputs = trainer.sample(train_display_images_a, train_display_images_b) write_images( image_outputs, display_size, '%s/gen_train%08d.jpg' % (image_directory, iterations + 1)) # HTML write_html(output_directory + "/index.html", iterations + 1, config['image_save_iter'], 'images') if (iterations + 1) % config['image_save_iter'] == 0: image_outputs = trainer.sample(test_display_images_a, test_display_images_b) write_images(image_outputs, display_size, '%s/gen.jpg' % image_directory) # Save network weights if (iterations + 1) % config['snapshot_save_iter'] == 0: trainer.save(checkpoint_directory, iterations) iterations += 1 if iterations >= max_iter: return
img_collect_b2a = [] dis_array_a2b = np.zeros(__TRY_NUM) dis_array_b2a = np.zeros(__TRY_NUM) for j in trange(__TRY_NUM, desc='try_num'): dis_accu_a2b = 0 dis_accu_b2a = 0 for _ in trange(__EVAL_NUM, desc='eval_num'): input_a = (dataset.sample_a().cuda()[0]).unsqueeze(dim=0) input_b = (dataset.sample_b().cuda()[0]).unsqueeze(dim=0) for _ in trange(__SAMPLE_NUM, desc='sample_num'): while len(img_collect_a2b) != 2: output_images = trainer.sample(input_a, input_b, training=False) img_collect_a2b.append(output_images[3].data) img_collect_b2a.append(output_images[-1].data) d_ab = dis_model.forward(*img_collect_a2b) d_ba = dis_model.forward(*img_collect_b2a) dis_accu_a2b += d_ab[0] dis_accu_b2a += d_ba[0] img_collect_a2b = [] img_collect_b2a = [] dis_array_a2b[j] = dis_accu_a2b / (__SAMPLE_NUM * __EVAL_NUM) dis_array_b2a[j] = dis_accu_b2a / (__SAMPLE_NUM * __EVAL_NUM)
for i, (image_a, image_b, _) in enumerate(dataloader): image_a = image_a.cuda() image_b = image_b.cuda() pbar.update(1) if i == __EVAL_NUM: break a2b_sample_array[i*im_h:(i+1)*im_h, 0:im_w, :] = \ tensor_to_image(image_a) b2a_sample_array[i*im_h:(i+1)*im_h, 0:im_w, :] = \ tensor_to_image(image_b) for j in range(__SAMPLE_NUM): output_images = trainer.sample(image_a, image_b, training=False) im_array_a2b = tensor_to_image(output_images[3]) im_array_b2a = tensor_to_image(output_images[-1]) a2b_sample_array[i*im_h:(i+1)*im_h, (j+1)*im_w:(j+2)*im_w, :] = \ im_array_a2b b2a_sample_array[i*im_h:(i+1)*im_h, (j+1)*im_w:(j+2)*im_w, :] = \ im_array_b2a imsave(os.path.join(opts.output_folder, "a2b_samples.jpg"), a2b_sample_array) imsave(os.path.join(opts.output_folder, "b2a_samples.jpg"), b2a_sample_array) pbar.close()
if trainer.train_seg: trainer.segmentation_head_update( images_as, images_bs, mask_s, sem_a, sem_b, config["adaptation"]["sem_seg_lambda"], comet_exp, ) if (iterations + 1) % config["image_save_iter"] == 0: with torch.no_grad(): test_image_outputs = trainer.sample( test_display_images_a, test_display_images_b, test_display_masks_a, test_display_masks_b, ) train_image_outputs = trainer.sample( train_display_images_a, train_display_images_b, train_display_masks_a, train_display_masks_b, ) if config["synthetic_frequency"] > 0: sim_image_outputs = trainer.sample( sim_display_images_a, sim_display_images_b, sim_display_masks, sim_display_masks,
def main(): from utils import get_all_data_loaders, prepare_sub_folder, write_html, write_loss, get_config, write_2images, Timer import argparse from torch.autograd import Variable from trainer import MUNIT_Trainer, UNIT_Trainer import torch.backends.cudnn as cudnn import torch # try: # from itertools import izip as zip # except ImportError: # will be 3.x series # pass import os import sys import tensorboardX import shutil os.environ["CUDA_VISIBLE_DEVICES"] = str(0) parser = argparse.ArgumentParser() parser.add_argument('--config', type=str, default='configs/edges2handbags_folder.yaml', help='Path to the config file.') parser.add_argument('--output_path', type=str, default='.', help="outputs path") parser.add_argument("--resume", action="store_true") parser.add_argument('--trainer', type=str, default='MUNIT', help="MUNIT|UNIT") opts = parser.parse_args() cudnn.benchmark = True ''' Note: https://www.pytorchtutorial.com/when-should-we-set-cudnn-benchmark-to-true/ 大部分情况下,设置这个 flag 可以让内置的 cuDNN 的 auto-tuner 自动寻找最适合当前配置的高效算法,来达到优化运行效率的问题 1. 如果网络的输入数据维度或类型上变化不大,设置 torch.backends.cudnn.benchmark = true 可以增加运行效率; 2. 如果网络的输入数据在每次 iteration 都变化的话,会导致 cnDNN 每次都会去寻找一遍最优配置,这样反而会降低运行效率。 ''' # Load experiment setting config = get_config(opts.config) max_iter = config['max_iter'] display_size = config['display_size'] config['vgg_model_path'] = opts.output_path # Setup model and data loader if opts.trainer == 'MUNIT': trainer = MUNIT_Trainer(config) elif opts.trainer == 'UNIT': trainer = UNIT_Trainer(config) else: sys.exit("Only support MUNIT|UNIT") trainer.cuda() train_loader_a, train_loader_b, test_loader_a, test_loader_b = get_all_data_loaders( config) train_display_images_a = torch.stack( [train_loader_a.dataset[i] for i in range(display_size)]).cuda() train_display_images_b = torch.stack( [train_loader_b.dataset[i] for i in range(display_size)]).cuda() test_display_images_a = torch.stack( [test_loader_a.dataset[i] for i in range(display_size)]).cuda() test_display_images_b = torch.stack( [test_loader_b.dataset[i] for i in range(display_size)]).cuda() # Setup logger and output folders model_name = os.path.splitext(os.path.basename(opts.config))[0] train_writer = tensorboardX.SummaryWriter( os.path.join(opts.output_path + "/logs", model_name)) output_directory = os.path.join(opts.output_path + "/outputs", model_name) checkpoint_directory, image_directory = prepare_sub_folder( output_directory) shutil.copy(opts.config, os.path.join( output_directory, 'config.yaml')) # copy config file to output folder # Start training iterations = trainer.resume(checkpoint_directory, hyperparameters=config) if opts.resume else 0 while True: for it, (images_a, images_b) in enumerate(zip(train_loader_a, train_loader_b)): trainer.update_learning_rate() images_a, images_b = images_a.cuda().detach(), images_b.cuda( ).detach() with Timer("Elapsed time in update: %f"): # Main training code trainer.dis_update(images_a, images_b, config) trainer.gen_update(images_a, images_b, config) torch.cuda.synchronize() # Dump training stats in log file if (iterations + 1) % config['log_iter'] == 0: print("Iteration: %08d/%08d" % (iterations + 1, max_iter)) write_loss(iterations, trainer, train_writer) # Write images if (iterations + 1) % config['image_save_iter'] == 0: with torch.no_grad(): test_image_outputs = trainer.sample( test_display_images_a, test_display_images_b) train_image_outputs = trainer.sample( train_display_images_a, train_display_images_b) write_2images(test_image_outputs, display_size, image_directory, 'test_%08d' % (iterations + 1)) write_2images(train_image_outputs, display_size, image_directory, 'train_%08d' % (iterations + 1)) # HTML write_html(output_directory + "/index.html", iterations + 1, config['image_save_iter'], 'images') if (iterations + 1) % config['image_display_iter'] == 0: with torch.no_grad(): image_outputs = trainer.sample(train_display_images_a, train_display_images_b) write_2images(image_outputs, display_size, image_directory, 'train_current') # Save network weights if (iterations + 1) % config['snapshot_save_iter'] == 0: trainer.save(checkpoint_directory, iterations) iterations += 1 if iterations >= max_iter: sys.exit('Finish training')
trainer.dis_update(images_a, segs_a, images_b, segs_b, config) if 0 == it % 2: trainer.gen_update(images_a, segs_a, images_b, segs_b, config) torch.cuda.synchronize() # Dump training stats in log file if (iterations + 1) % config['log_iter'] == 0: print("Iteration: %08d/%08d" % (iterations + 1, max_iter)) write_loss(iterations, trainer, train_writer) # Write images if ((iterations + 1) % config['image_save_iter'] == 0) or (iterations == 24): with torch.no_grad(): test_image_outputs = trainer.sample(test_loader_a, test_loader_b, display_size) train_image_outputs = trainer.sample(train_loader_a, train_loader_b, display_size) write_2images(test_image_outputs, display_size, image_directory, 'test_%08d' % (iterations + 1)) write_2images(train_image_outputs, display_size, image_directory, 'train_%08d' % (iterations + 1)) # HTML write_html(output_directory + "/index.html", iterations + 1, config['image_save_iter'], 'images') if ((iterations + 1) % config['image_display_iter'] == 0) or (iterations == 24): with torch.no_grad():
trainer.update_learning_rate() images_a, images_b = Variable(images_a.cuda()), Variable(images_b.cuda()) # Main training code trainer.dis_update(images_a, images_b, config) trainer.gen_update(images_a, images_b, config) # Dump training stats in log file if (iterations + 1) % config['log_iter'] == 0: print("Iteration: %08d/%08d" % (iterations + 1, max_iter)) write_loss(iterations, trainer, train_writer) # Write images if (iterations + 1) % config['image_save_iter'] == 0: # Test set images image_outputs = trainer.sample(test_display_images_a, test_display_images_b) write_2images(image_outputs, display_size, image_directory, 'test_%08d' % (iterations + 1)) # Train set images image_outputs = trainer.sample(train_display_images_a, train_display_images_b) write_2images(image_outputs, display_size, image_directory, 'train_%08d' % (iterations + 1)) # HTML write_html(output_directory + "/index.html", iterations + 1, config['image_save_iter'], 'images') if (iterations + 1) % config['image_display_iter'] == 0: train_display_images_a = Variable(torch.stack([train_loader_a.dataset[i] for i in range(display_size)]).cuda(), volatile=True) train_display_images_b = Variable(torch.stack([train_loader_b.dataset[i] for i in range(display_size)]).cuda(), volatile=True) image_outputs = trainer.sample(train_display_images_a, train_display_images_b) write_2images(image_outputs, display_size, image_directory, 'train_current') # Save network weights if (iterations + 1) % config['snapshot_save_iter'] == 0:
def main(argv): (opts, args) = parser.parse_args(argv) cudnn.benchmark = True # Load experiment setting config = get_config(opts.config) max_iter = config['max_iter'] # Setup logger and output folders output_subfolders = prepare_logging_folders(config['output_root'], config['experiment_name']) logger = create_logger( os.path.join(output_subfolders['logs'], 'train_log.log')) shutil.copy(opts.config, os.path.join( output_subfolders['logs'], 'config.yaml')) # copy config file to output folder tb_logger = tensorboard_logger.Logger(output_subfolders['logs']) logger.info('============ Initialized logger ============') logger.info('Config File: {}'.format(opts.config)) # Setup model and data loader trainer = MUNIT_Trainer(config, opts) trainer.cuda() loaders = get_all_data_loaders(config) val_display_images = next(iter(loaders['val'])) logger.info('Test images: {}'.format(val_display_images['A_paths'])) # Start training iterations = trainer.resume(opts.model_path, hyperparameters=config) if opts.resume else 0 while True: for it, images in enumerate(loaders['train']): trainer.update_learning_rate() images_a = images['A'] images_b = images['B'] images_a, images_b = Variable(images_a.cuda()), Variable( images_b.cuda()) # Main training code trainer.dis_update(images_a, images_b, config) trainer.gen_update(images_a, images_b, config) # Dump training stats in log file if (iterations + 1) % config['log_iter'] == 0: for tag, value in trainer.loss.items(): tb_logger.scalar_summary(tag, value, iterations) val_output_imgs = trainer.sample( Variable(val_display_images['A'].cuda()), Variable(val_display_images['B'].cuda())) tb_imgs = [] for imgs in val_output_imgs.values(): tb_imgs.append(torch.cat(torch.unbind(imgs, 0), dim=2)) tb_logger.image_summary(list(val_output_imgs.keys()), tb_imgs, iterations) if (iterations + 1) % config['print_iter'] == 0: logger.info( "Iteration: {:08}/{:08} Discriminator Loss: {:.4f} Generator Loss: {:.4f}" .format(iterations + 1, max_iter, trainer.loss['D/total'], trainer.loss['G/total'])) # Write images # if (iterations + 1) % config['image_save_iter'] == 0: # val_output_imgs = trainer.sample( # Variable(val_display_images['A'].cuda()), # Variable(val_display_images['B'].cuda())) # # for key, imgs in val_output_imgs.items(): # key = key.replace('/', '_') # write_images(imgs, config['display_size'], '{}/{}_{:08}.jpg'.format(output_subfolders['images'], key, iterations+1)) # # logger.info('Saved images to: {}'.format(output_subfolders['images'])) # Save network weights if (iterations + 1) % config['snapshot_save_iter'] == 0: trainer.save(output_subfolders['models'], iterations) iterations += 1 if iterations >= max_iter: return