search

EPFL-VILAB / omnidata

A Scalable Pipeline for Making Steerable Multi-Task Mid-Level Vision Datasets from 3D Scans [ICCV 2021]
Other
418 stars 53 forks source link

Doubt about the cosine_angular_loss #47

Open ireneMsm2020 opened 1 year ago

ireneMsm2020 commented 1 year ago

Hello, I have doubts about cosine_angular_loss. After permute operation, the channel dim is in dim 3, but in normalize operation, you set dim=1. So is there a mistake? Can you please explain it ?

def masked_cosine_angular_loss(preds, target, mask_valid): preds = (2 preds - 1).clamp(-1, 1) target = (2 target - 1).clamp(-1, 1) mask_valid = mask_valid[:,0,:,:].bool().squeeze(1) preds = preds.permute(0,2,3,1)[mask_valid, :] target = target.permute(0,2,3,1)[mask_valid, :] preds_norm = torch.nn.functional.normalize(preds, p=2, dim=1) target_norm = torch.nn.functional.normalize(target, p=2, dim=1) loss = torch.mean(-torch.sum(preds_norm * target_norm, dim = 1)) return loss

alexsax commented 1 year ago

Hi your concern looks right to me, but the models are quite good. So I think the released code might just be the problem (and the evaluation code we borrowed right from the OASIS paper). @Ainaz99 what do you think?

limacv commented 1 year ago

isn't the mask operation flatten the two tensor into [N, 3] shape? In that case dim=1 should be correct.