Synthetic-only training Issue

[htuann2712]

Hi @guipotje,

I'm currently trying to train the XFeat model using only synthetic data. I tried running your implementation with both my own dataset and the MSCOCO 20K dataset, but I noticed a phenomenon where the refinementMLP outputs seem to be random.

Loss: 6.1655 acc_c0 0.391 acc_c1 0.227 acc_f: 0.069 loss_c: 8.066 loss_f: 7.264 loss_kp: 0.467 #matches_c: 2042 loss_kp_pos: 10.147 acc_kp_pos: 0.284: 5% 7999/160000 [2:18:57<294:02:21, 6.96s/it]saving iter 8000
Loss: 6.3344 acc_c0 0.239 acc_c1 0.193 acc_f: 0.078 loss_c: 6.790 loss_f: 7.115 loss_kp: 0.252 #matches_c: 3726 loss_kp_pos: 12.934 acc_kp_pos: 0.147: 5% 8499/160000 [2:27:34<43:11:25, 1.03s/it]saving iter 8500

Have you experienced this issue before? Or are there any specific considerations when training the model solely on synthetic data?

Thank you.