def main(): import argparse parser = argparse.ArgumentParser( description="imsitu VSRL. Training, evaluation and prediction.") parser.add_argument("--gpuid", default=-1, help="put GPU id > -1 in GPU mode", type=int) #parser.add_argument("--command", choices = ["train", "eval", "resume", 'predict'], required = True) parser.add_argument('--resume_training', action='store_true', help='Resume training from the model [resume_model]') parser.add_argument('--resume_model', type=str, default='', help='The model we resume') parser.add_argument('--verb_module', type=str, default='', help='pretrained verb module') parser.add_argument('--role_module', type=str, default='', help='pretrained role module') parser.add_argument('--train_role', action='store_true', help='cnn fix, verb fix, role train from the scratch') parser.add_argument( '--finetune_verb', action='store_true', help='cnn fix, verb finetune, role train from the scratch') parser.add_argument( '--finetune_cnn', action='store_true', help='cnn finetune, verb finetune, role train from the scratch') parser.add_argument('--output_dir', type=str, default='./trained_models', help='Location to output the model') parser.add_argument('--evaluate', action='store_true', help='Only use the testing mode') parser.add_argument('--test', action='store_true', help='Only use the testing mode') parser.add_argument('--dataset_folder', type=str, default='./imSitu', help='Location of annotations') parser.add_argument('--imgset_dir', type=str, default='./resized_256', help='Location of original images') parser.add_argument('--frcnn_feat_dir', type=str, help='Location of output from detectron') #todo: train role module separately with gt verbs args = parser.parse_args() batch_size = 640 #lr = 5e-6 lr = 0.0001 lr_max = 5e-4 lr_gamma = 0.1 lr_step = 15 clip_norm = 0.5 weight_decay = 1e-4 n_epoch = 500 n_worker = 3 #dataset_folder = 'imSitu' #imgset_folder = 'resized_256' dataset_folder = args.dataset_folder imgset_folder = args.imgset_dir print('model spec :, top down att with role q ') train_set = json.load(open(dataset_folder + "/updated_train_new.json")) imsitu_roleq = json.load(open("imsitu_data/imsitu_questions_prev.json")) verb_templates = json.load( open("imsitu_data/verb_questions_template.json")) encoder = imsitu_encoder(train_set, imsitu_roleq, verb_templates) model = model_verbq_final_deprole_verbqaft.BaseModel(encoder, args.gpuid) # To group up the features cnn_features, role_features = utils.group_features_noun(model) #cnn_features, role_features = utils.group_features_joint_reverb(model) train_set = imsitu_loader_roleq_updated(imgset_folder, train_set, encoder, model.train_preprocess()) train_loader = torch.utils.data.DataLoader(train_set, batch_size=64, shuffle=True, num_workers=n_worker) dev_set = json.load(open(dataset_folder + "/dev.json")) dev_set = imsitu_loader_roleq_updated(imgset_folder, dev_set, encoder, model.dev_preprocess()) dev_loader = torch.utils.data.DataLoader(dev_set, batch_size=64, shuffle=True, num_workers=n_worker) test_set = json.load(open(dataset_folder + "/test.json")) test_set = imsitu_loader_roleq_updated(imgset_folder, test_set, encoder, model.dev_preprocess()) test_loader = torch.utils.data.DataLoader(test_set, batch_size=64, shuffle=True, num_workers=n_worker) traindev_set = json.load(open(dataset_folder + "/dev.json")) traindev_set = imsitu_loader_roleq_updated(imgset_folder, traindev_set, encoder, model.dev_preprocess()) traindev_loader = torch.utils.data.DataLoader(traindev_set, batch_size=8, shuffle=True, num_workers=n_worker) if args.resume_training: print('Resume training ') args.train_all = True '''if len(args.resume_model) == 0: raise Exception('[pretrained verb module] not specified')''' utils.load_net_deproleqa0td(args.resume_model, [model]) optimizer_select = 0 model_name = 'resume_all' else: utils.load_net(args.verb_module, [model.verb_module]) utils.load_net(args.role_module, [model.role_module]) model_name = 'train_full' if not os.path.exists(args.output_dir): os.mkdir(args.output_dir) torch.manual_seed(1234) if args.gpuid >= 0: #print('GPU enabled') model.cuda() torch.cuda.manual_seed(1234) torch.backends.cudnn.deterministic = True '''optimizer = torch.optim.Adam([ {'params': cnn_features, 'lr': 5e-5}, {'params': role_features} ], lr=1e-3)''' utils.set_trainable_param(cnn_features, False) optimizer = torch.optim.Adam(role_features, lr=5e-5) #optimizer = torch.optim.Adam(model.parameters(), lr=lr, weight_decay=weight_decay) #scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=lr_step, gamma=lr_gamma) #gradient clipping, grad check scheduler = torch.optim.lr_scheduler.ExponentialLR(optimizer, gamma=0.9) if args.evaluate: top1, top5, val_loss = eval(model, dev_loader, encoder, args.gpuid, write_to_file=True) top1_avg = top1.get_average_results() top5_avg = top5.get_average_results() avg_score = top1_avg["verb"] + top1_avg["value"] + top1_avg["value-all"] + top5_avg["verb"] + \ top5_avg["value"] + top5_avg["value-all"] + top5_avg["value*"] + top5_avg["value-all*"] avg_score /= 8 print('Dev average :{:.2f} {} {}'.format( avg_score * 100, utils.format_dict(top1_avg, '{:.2f}', '1-'), utils.format_dict(top5_avg, '{:.2f}', '5-'))) #write results to csv file role_dict = top1.role_dict fail_val_all = top1.value_all_dict pass_val_dict = top1.vall_all_correct with open('role_pred_data.json', 'w') as fp: json.dump(role_dict, fp, indent=4) with open('fail_val_all.json', 'w') as fp: json.dump(fail_val_all, fp, indent=4) with open('pass_val_all.json', 'w') as fp: json.dump(pass_val_dict, fp, indent=4) print('Writing predictions to file completed !') elif args.test: top1, top5, val_loss = eval(model, test_loader, encoder, args.gpuid, write_to_file=True) top1_avg = top1.get_average_results() top5_avg = top5.get_average_results() avg_score = top1_avg["verb"] + top1_avg["value"] + top1_avg["value-all"] + top5_avg["verb"] + \ top5_avg["value"] + top5_avg["value-all"] + top5_avg["value*"] + top5_avg["value-all*"] avg_score /= 8 print('Test average :{:.2f} {} {}'.format( avg_score * 100, utils.format_dict(top1_avg, '{:.2f}', '1-'), utils.format_dict(top5_avg, '{:.2f}', '5-'))) else: print('Model training started!') train(model, train_loader, dev_loader, traindev_loader, optimizer, scheduler, n_epoch, args.output_dir, encoder, args.gpuid, clip_norm, lr_max, model_name, args)
def main(): import argparse parser = argparse.ArgumentParser( description="imsitu VSRL. Training, evaluation and prediction.") parser.add_argument("--gpuid", default=-1, help="put GPU id > -1 in GPU mode", type=int) parser.add_argument( "--command", choices=["train", "eval", "resume", 'predict', 'finetune'], required=True) parser.add_argument("--batch_size", '-b', type=int, default=64) parser.add_argument("--weights_file", help="the model to start from") parser.add_argument( '--finetune_verb', action='store_true', help='verb classifier train from the scratch, all others fixed') parser.add_argument('--verb_module', type=str, default='', help='pretrained verb module') args = parser.parse_args() batch_size = args.batch_size #lr = 1e-5 lr = 1e-4 lr_max = 5e-4 lr_gamma = 0.1 lr_step = 25 clip_norm = 50 weight_decay = 1e-5 n_epoch = 500 n_worker = 4 # print('LR scheme : lr decay, vgg, fc as per gnn paper batch 64', 1e-5, 0.1,25) dataset_folder = 'imSitu' imgset_folder = 'resized_256' model_dir = 'trained_models' train_set = json.load(open(dataset_folder + "/train.json")) encoder = imsitu_encoder(train_set) model = model_verb_embd.RelationNetworks(encoder, args.gpuid) train_set = imsitu_loader(imgset_folder, train_set, encoder, model.train_preprocess()) train_loader = torch.utils.data.DataLoader(train_set, batch_size=batch_size, shuffle=True, num_workers=n_worker) dev_set = json.load(open(dataset_folder + "/dev.json")) dev_set = imsitu_loader(imgset_folder, dev_set, encoder, model.dev_preprocess()) dev_loader = torch.utils.data.DataLoader(dev_set, batch_size=batch_size, shuffle=True, num_workers=n_worker) traindev_set = json.load(open(dataset_folder + "/dev.json")) traindev_set = imsitu_loader(imgset_folder, traindev_set, encoder, model.train_preprocess()) traindev_loader = torch.utils.data.DataLoader(traindev_set, batch_size=batch_size, shuffle=True, num_workers=n_worker) if args.command == "resume": print("loading model weights...") model.load_state_dict(torch.load(args.weights_file)) elif args.finetune_verb: print( 'CNN fix, Verb fc fixed, train verb classifier layer from the scratch from: {}' .format(args.verb_module)) if len(args.verb_module) == 0: raise Exception('[pretrained verb module] not specified') utils.load_net(args.verb_module, [model.conv, model.verb], ['conv', 'verb']) #print(model) if args.gpuid >= 0: print('GPU enabled') model.cuda() # optimizer = torch.optim.Adam(model.parameters(), lr=lr, weight_decay=weight_decay) utils.set_trainable(model, False) utils.set_trainable_param(model.classifier.parameters(), True) utils.set_trainable_param(model.verb.parameters(), True) optimizer = torch.optim.Adam([{ 'params': model.classifier.parameters(), 'lr': 1e-3 }, { 'params': model.verb.parameters(), 'lr': 5e-5 }]) # scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=lr_step, gamma=lr_gamma) scheduler = torch.optim.lr_scheduler.ExponentialLR(optimizer, gamma=0.9) '''optimizer = utils.CosineAnnealingWR(0.01,1200000 , 50, torch.optim.Adam(model.parameters(), lr=0, betas=(0.9, 0.98), eps=1e-9))''' #gradient clipping, grad check print('Model training started!') train(model, train_loader, dev_loader, traindev_loader, optimizer, scheduler, n_epoch, model_dir, encoder, args.gpuid, clip_norm, lr_max)