something about run this code

[zhuchuanleiqq]

something about run this code

[sanghoooon]

Could you provide shapes of feat, feat_recon_pos, feat_recon_neg ?

[zhuchuanleiqq]

this three is all [64, 1024, 18, 9],i also print the shape of self.criterion(feat, feat_recon_pos, feat_recon_neg),it is [64, 1024, 18],the last dimension is disappear,i think the reason is the setting of the nn.TripletMarginLoss(margin=0.3, p=2.0, reduce=False),also,i print the shape of comask_pos,it is [64, 1, 18, 9]

[sanghoooon]

Could you specify your pytorch version?

[zhuchuanleiqq]

python 3.7

[sanghoooon]

PyTorch version?

[zhuchuanleiqq]

pytorch 1.10.1

[sanghoooon]

Please check out the README.md file: our implementation is based on PyTorch 1.7.1
The behaviour within nn.TripletMarginLoss may differ across PyTorch versions.
You might want to reshape tensors prior to passing them through nn.TripletMarginLoss.

[zhuchuanleiqq]

ok ,thank you for your help,may i ask what is the shape of your implementation about the self.criterion(feat, feat_recon_pos, feat_recon_neg)?

[sanghoooon]

All of them have to be [batch size, fdim, H, W].
Check out soft_warping in model.py

[cq0907]

All of them have to be [batch size, fdim, H, W]. Check out soft_warping in model.py

I have the same problem, so may i want to ask what is the shape of your implementation about the comask_pos?

[sanghoooon]

Below are i/o dimensions for compute_comask that outputs comask_pos:

• matching_pr [bs, hw, hw]
• mask_q [bs, 1, h, w]
• mask_k [bs, 1, h, w]
• comask [bs, 1, h, w]

where h and w are height and width of feature maps, respectively.

[cq0907]

Below are i/o dimensions for compute_comask that outputs comask_pos:

• matching_pr [bs, hw, hw]
• mask_q [bs, 1, h, w]
• mask_k [bs, 1, h, w]
• comask [bs, 1, h, w]

where h and w are height and width of feature maps, respectively.

OK, thank you!