def main(args): # need to clear the pipeline # top1 & top10 need to be chosen in the same params ??? normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) test_transform = transforms.Compose([transforms.ToTensor(), normalize]) test_loader = data_config(args.image_dir, args.anno_dir, 64, 'test', 100, test_transform) unique_image = get_image_unique(args.image_dir, args.anno_dir, 64, 'test', 100, test_transform) ac_i2t_top1_best = 0.0 ac_i2t_top10_best = 0.0 ac_t2i_top1_best = 0.0 ac_t2i_top10_best = 0.0 ac_t2i_top5_best = 0.0 ac_i2t_top5_best = 0.0 i2t_models = os.listdir(args.model_path) i2t_models.sort() model_list = [] for i2t_model in i2t_models: if i2t_model.split('.')[0] != "model_best": model_list.append(int(i2t_model.split('.')[0])) model_list.sort() logging.info('Testing on dataset: {}'.format(args.anno_dir)) for i2t_model in model_list: model_file = os.path.join(args.model_path, str(i2t_model) + '.pth.tar') if os.path.isdir(model_file): continue epoch = i2t_model if int(epoch) < args.epoch_start: continue network, _ = network_config(args, 'test', None, True, model_file) ac_top1_i2t, ac_top5_i2t, ac_top10_i2t, ac_top1_t2i, ac_top5_t2i, ac_top10_t2i, test_time = test( test_loader, network, args, unique_image) if ac_top1_t2i > ac_t2i_top1_best: ac_i2t_top1_best = ac_top1_i2t ac_i2t_top5_best = ac_top5_i2t ac_i2t_top10_best = ac_top10_i2t ac_t2i_top1_best = ac_top1_t2i ac_t2i_top5_best = ac_top5_t2i ac_t2i_top10_best = ac_top10_t2i dst_best = os.path.join(args.checkpoint_dir, 'model_best', str(epoch)) + '.pth.tar' logging.info('epoch:{}'.format(epoch)) logging.info( 'top1_t2i: {:.3f}, top5_t2i: {:.3f}, top10_t2i: {:.3f}, top1_i2t: {:.3f}, top5_i2t: {:.3f}, top10_i2t: {:.3f}' .format(ac_top1_t2i, ac_top5_t2i, ac_top10_t2i, ac_top1_i2t, ac_top5_i2t, ac_top10_i2t)) logging.info( 't2i_top1_best: {:.3f}, t2i_top5_best: {:.3f}, t2i_top10_best: {:.3f}, i2t_top1_best: {:.3f}, i2t_top5_best: {:.3f}, i2t_top10_best: {:.3f}' .format(ac_t2i_top1_best, ac_t2i_top5_best, ac_t2i_top10_best, ac_i2t_top1_best, ac_i2t_top5_best, ac_i2t_top10_best)) logging.info(args.model_path) logging.info(args.log_dir)
def main(args): # transform train_transform = transforms.Compose([ transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)) ]) val_transform = transforms.Compose([ transforms.ToTensor(), transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)) ]) # data train_loader = data_config(args.dataset_dir, args.batch_size, 'train', args.max_length, train_transform) #val_loader = data_config(args.dataset_dir, 64, 'val', args.max_length, val_transform) # loss compute_loss = Loss(args) nn.DataParallel(compute_loss).cuda() # network network, optimizer = network_config(args, 'train', compute_loss.parameters(), args.resume, args.model_path) # lr_scheduler scheduler = lr_scheduler(optimizer, args) for epoch in range(args.num_epoches - args.start_epoch): # train for one epoch train_loss, train_time, image_precision, text_precision = train( args.start_epoch + epoch, train_loader, network, optimizer, compute_loss, args) # evaluate on validation set is_best = False print('Train done for epoch-{}'.format(args.start_epoch + epoch)) state = { 'network': network.state_dict(), 'optimizer': optimizer.state_dict(), 'W': compute_loss.W, 'epoch': args.start_epoch + epoch } # 'ac': [ac_top1_i2t, ac_top10_i2t, ac_top1_t2i, ac_top10_t2i], # 'best_ac': [ac_i2t_best, ac_t2i_best]} save_checkpoint(state, epoch, args.checkpoint_dir, is_best) logging.info( 'Epoch: [{}|{}], train_time: {:.3f}, train_loss: {:.3f}'.format( args.start_epoch + epoch, args.num_epoches, train_time, train_loss)) logging.info('image_precision: {:.3f}, text_precision: {:.3f}'.format( image_precision, text_precision)) adjust_lr(optimizer, args.start_epoch + epoch, args) scheduler.step() for param in optimizer.param_groups: print('lr:{}'.format(param['lr'])) break logging.info('Train done') logging.info(args.checkpoint_dir) logging.info(args.log_dir)
def load_model(args): network, _ = network_config(args, 'test', None, True, model_path, False) network.eval() with open(test_sort_path, 'rb') as f_pkl: test_sort = pickle.load(f_pkl) with open(word_to_index_path, 'rb') as f_2_pkl: word_to_index = pickle.load(f_2_pkl) word_to_index = {k.lower(): v for k, v in word_to_index.items()} return network, test_sort, word_to_index
def main(args): # need to clear the pipeline # top1 & top10 need to be chosen in the same params ??? test_transform = transforms.Compose([ transforms.ToTensor(), transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)) ]) test_loader = data_config(args.image_dir, args.anno_dir, 64, 'test', args.max_length, test_transform) ac_i2t_top1_best = 0.0 ac_i2t_top10_best = 0.0 ac_t2i_top1_best = 0.0 ac_t2i_top10_best = 0.0 i2t_models = os.listdir(args.model_path) i2t_models.sort() for i2t_model in i2t_models: model_file = os.path.join(args.model_path, i2t_model) if os.path.isdir(model_file): continue epoch = i2t_model.split('.')[0] if int(epoch) >= args.epoch_ema: ema = True else: ema = False network, _ = network_config(args, 'test', None, True, model_file, ema) ac_top1_i2t, ac_top10_i2t, ac_top1_t2i, ac_top10_t2i, test_time = test( test_loader, network, args) if ac_top1_t2i > ac_t2i_top1_best: ac_i2t_top1_best = ac_top1_i2t ac_i2t_top10_best = ac_top10_i2t ac_t2i_top1_best = ac_top1_t2i ac_t2i_top10_best = ac_top10_t2i dst_best = os.path.join(args.model_path, 'model_best', str(epoch)) + '.pth.tar' shutil.copyfile(model_file, dst_best) logging.info('epoch:{}'.format(epoch)) logging.info( 'top1_t2i: {:.3f}, top10_t2i: {:.3f}, top1_i2t: {:.3f}, top10_i2t: {:.3f}' .format(ac_top1_t2i, ac_top10_t2i, ac_top1_i2t, ac_top10_i2t)) logging.info( 't2i_top1_best: {:.3f}, t2i_top10_best: {:.3f}, i2t_top1_best: {:.3f}, i2t_top10_best: {:.3f}' .format(ac_t2i_top1_best, ac_t2i_top10_best, ac_i2t_top1_best, ac_i2t_top10_best)) logging.info(args.model_path) logging.info(args.log_dir)
def main(args): # transform normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) train_transform = transforms.Compose([ transforms.RandomHorizontalFlip(), transforms.ToTensor(), normalize, ]) val_transform = transforms.Compose([ transforms.ToTensor(), normalize ]) test_transform = transforms.Compose([ transforms.ToTensor(), normalize ]) cap_transform = None # data train_loader = data_config(args.image_dir, args.anno_dir, args.batch_size, 'train', 100, train_transform, cap_transform=cap_transform) test_loader = data_config(args.image_dir, args.anno_dir, 64, 'test', 100, test_transform) unique_image = get_image_unique(args.image_dir, args.anno_dir, 64, 'test', 100, test_transform) # loss compute_loss = Loss(args) nn.DataParallel(compute_loss).cuda() # network network, optimizer = network_config(args, 'train', compute_loss.parameters(), args.resume, args.model_path) # lr_scheduler scheduler = WarmupMultiStepLR(optimizer, (20, 25, 35), 0.1, 0.01, 10, 'linear') ac_t2i_top1_best = 0.0 best_epoch = 0 for epoch in range(args.num_epoches - args.start_epoch): network.train() # train for one epoch train_loss, train_time, image_precision, text_precision = train(args.start_epoch + epoch, train_loader, network, optimizer, compute_loss, args) # evaluate on validation set is_best = False print('Train done for epoch-{}'.format(args.start_epoch + epoch)) logging.info('Epoch: [{}|{}], train_time: {:.3f}, train_loss: {:.3f}'.format(args.start_epoch + epoch, args.num_epoches, train_time, train_loss)) logging.info('image_precision: {:.3f}, text_precision: {:.3f}'.format(image_precision, text_precision)) scheduler.step() for param in optimizer.param_groups: print('lr:{}'.format(param['lr'])) if epoch >= 0: ac_top1_i2t, ac_top5_i2t, ac_top10_i2t, ac_top1_t2i, ac_top5_t2i , ac_top10_t2i, test_time = test(test_loader, network, args, unique_image) state = {'network': network.state_dict(), 'optimizer': optimizer.state_dict(), 'W': compute_loss.W, 'epoch': args.start_epoch + epoch} if ac_top1_t2i > ac_t2i_top1_best: best_epoch = epoch ac_t2i_top1_best = ac_top1_t2i save_checkpoint(state, epoch, args.checkpoint_dir, is_best) logging.info('epoch:{}'.format(epoch)) logging.info('top1_t2i: {:.3f}, top5_t2i: {:.3f}, top10_t2i: {:.3f}, top1_i2t: {:.3f}, top5_i2t: {:.3f}, top10_i2t: {:.3f}'.format( ac_top1_t2i, ac_top5_t2i, ac_top10_t2i, ac_top1_i2t, ac_top5_i2t, ac_top10_i2t)) logging.info('Best epoch:{}'.format(best_epoch)) logging.info('Train done') logging.info(args.checkpoint_dir) logging.info(args.log_dir)
def load_model(): # global graph # graph = tf.get_default_graph() global model model_file = args.model_path model, _ = network_config(args, 'test', None, True, model_file, True)