finetune pretrained PoseNet using NeRF
(args, model, feat_model, optimizer, i_split, hwf, near, far, device, early_stopping, images=None, poses_train=None, train_dl=None, val_dl=None, test_dl=None)
| 410 | return total_loss_mean, total_psnr_mean |
| 411 | |
| 412 | def train_feature_matching(args, model, feat_model, optimizer, i_split, hwf, near, far, device, early_stopping, images=None, poses_train=None, train_dl=None, val_dl=None, test_dl=None): |
| 413 | ''' finetune pretrained PoseNet using NeRF ''' |
| 414 | # half_res = False # direct-pn paper settings |
| 415 | half_res = True # debug |
| 416 | |
| 417 | # load NeRF model |
| 418 | _, render_kwargs_test, start, grad_vars, _ = create_nerf(args) |
| 419 | global_step = start |
| 420 | if args.reduce_embedding==2: |
| 421 | render_kwargs_test['i_epoch'] = global_step |
| 422 | |
| 423 | data_loaders = [train_dl, val_dl, test_dl] |
| 424 | bds_dict = { |
| 425 | 'near' : near, |
| 426 | 'far' : far, |
| 427 | } |
| 428 | # render_kwargs_train.update(bds_dict) |
| 429 | render_kwargs_test.update(bds_dict) |
| 430 | i_train, i_val, i_test = i_split |
| 431 | |
| 432 | render_kwargs_test['embedding_a'] = disable_model_grad(render_kwargs_test['embedding_a']) |
| 433 | render_kwargs_test['embedding_t'] = disable_model_grad(render_kwargs_test['embedding_t']) |
| 434 | render_kwargs_test['network_fn'] = disable_model_grad(render_kwargs_test['network_fn']) |
| 435 | render_kwargs_test['network_fine'] = disable_model_grad(render_kwargs_test['network_fine']) |
| 436 | |
| 437 | N_epoch = 2001 |
| 438 | print('Begin') |
| 439 | print('TRAIN views are', i_train) |
| 440 | print('TEST views are', i_test) |
| 441 | print('VAL views are', i_val) |
| 442 | |
| 443 | world_setup_dict = { |
| 444 | 'pose_scale' : train_dl.dataset.pose_scale, |
| 445 | 'pose_scale2' : train_dl.dataset.pose_scale2, |
| 446 | 'move_all_cam_vec' : train_dl.dataset.move_all_cam_vec, |
| 447 | } |
| 448 | |
| 449 | time0 = time.time() |
| 450 | |
| 451 | model_log = tqdm(total=0, position = 1, bar_format='{desc}') |
| 452 | for epoch in tqdm(range(N_epoch), desc='epochs'): |
| 453 | #train 1 epoch with batch_size = 1, 15% speed up for DFNet_s |
| 454 | loss, psnr = train_on_epoch(args, data_loaders, model, feat_model, hwf, optimizer, half_res, device, world_setup_dict, **render_kwargs_test) |
| 455 | |
| 456 | # 26% speed up for DFNet_s |
| 457 | val_loss, val_psnr = eval_on_epoch(args, data_loaders, model, feat_model, hwf, half_res, device, world_setup_dict, **render_kwargs_test) |
| 458 | |
| 459 | |
| 460 | tqdm.write('At epoch {0:4d} : train loss: {1:.4f}, train psnr: {2:.4f}, val loss: {3:.4f}, val psnr: {4:.4f}'.format(epoch, loss, psnr, val_loss, val_psnr)) |
| 461 | |
| 462 | # check wether to early stop |
| 463 | early_stopping(val_loss, model, epoch=epoch, save_multiple=(not args.no_save_multiple), save_all=args.save_all_ckpt, val_psnr=val_psnr) |
| 464 | if early_stopping.early_stop: |
| 465 | print("Early stopping") |
| 466 | break |
| 467 | model_log.set_description_str(f'Best val loss: {early_stopping.val_loss_min:.4f}') |
| 468 | |
| 469 | if epoch % args.i_eval == 0: |
no test coverage detected