Question about performance on ScanRefer about eda HOT 5 OPEN

I've encountered exactly the same problem. Has it been resolved?

from eda.

I know the reason now; don't use a single gpu for training when running train_scanrefer_single.sh.

from eda.

Hello! I also encountered the same problem, and I would like to ask if you know:

1. Why using a single GPU results in high loss values?
2. I used 2 GPUs to run the program, but the loss after epoch=10 decreased very slowly compared to the author's results. Can you help analyze the reasons?

from eda.

As training continues, the loss value does not show significant decrease

from eda.

{'num_target': 256, 'sampling': 'kps', 'num_encoder_layers': 3, 'num_decoder_layers': 6, 'self_position_embedding': 'loc_learned', 'self_attend': True, 'query_points_obj_topk': 4, 'use_contrastive_align': True, 'use_soft_token_loss': True, 'detect_intermediate': True, 'joint_det': True, 'batch_size': 12, 'dataset': ['scanrefer'], 'test_dataset': 'scanrefer', 'data_root': '/home/user1009/project/EDA/preparation/', 'use_height': False, 'use_color': True, 'use_multiview': False, 'wo_obj_name': 'None', 'butd': False, 'butd_gt': False, 'butd_cls': False, 'augment_det': True, 'num_workers': 4, 'start_epoch': 1, 'max_epoch': 400, 'optimizer': 'adamW', 'weight_decay': 0.0005, 'lr': 0.0002, 'lr_backbone': 0.002, 'text_encoder_lr': 1e-05, 'lr_scheduler': 'step', 'lr_decay_epochs': [50, 75], 'lr_decay_rate': 0.1, 'clip_norm': 0.1, 'bn_momentum': 0.1, 'syncbn': False, 'warmup_epoch': -1, 'warmup_multiplier': 100, 'checkpoint_path': None, 'log_dir': '/home/user1009/project/EDA/preparation/output/logs/eda/scanrefer/1718353528', 'print_freq': 500, 'save_freq': 3, 'val_freq': 3, 'local_rank': 0, 'ap_iou_thresholds': [0.25, 0.5], 'rng_seed': 0, 'debug': False, 'eval': False, 'eval_train': False, 'pp_checkpoint': '/home/user1009/project/EDA/preparation/gf_detector_l6o256.pth', 'reduce_lr': False}

The only difference from the parameters given by the author is in 'lr_decay_epochs': [50, 75].
But I don't think this is the reason for the worsened results.

from eda.