Cannot Reproduce grounded-end-to-end score

[minuenergy]

there is no pretrain weight for GVSR, so I followed your training but I cannot reproduce score. this is my output after training

CIDEr : 0.44,

{'loss_SRL': tensor(2.4280), 'loss_vb': tensor(4.9050), 'loss_role': tensor(0.3863)}██████████████████| 100.00% [83/83 03:13<00:00] {'cider': tensor(0.4461, device='cuda:0', dtype=torch.float64), 'rouge': tensor(0.3432, device='cuda:0', dtype=torch.float64), 'MacroVb_cider': tensor(0.5970, device='cuda:0', dtype=torch.float64), 'MacroArg_cider': tensor(0.3542, device='cuda:0', dtype=torch.float64), 'lea': tensor(0.3569, device='cuda:0'), 'Per_Ev_Top_1': tensor(0.3831, device='cuda:0'), 'Per_Ev_Top_5': tensor(0.7021, device='cuda:0'), 'recall_macro_1_th_9': tensor(0.2445, device='cuda:0'), 'IoU': tensor(0.1871, device='cuda:0'), 'IoU_30': tensor(0.2935, device='cuda:0'), 'IoU_50': tensor(0.1391, device='cuda:0')}

[jinhseo]

This might be off-topic, but does training done with the provided vsitu_all_11_frames_feats.zip? I would like to point out if the vsitu_all_11_frames_feats.zip used for training is indeed 30GB

[zeeshank95]

I will re-upload the object features, you should be able to reproduce the results with the new features.