Nan Loss

[qqc222111]

why is the number 0.9 in ”_bilinear_interpolate(img_masks, u_2_flat, v_2_flat) * img_masks >= 0.9“. Can it be replaced it in order to avoid causing the nan loss value, and what should it be replaced with?

When I replace it with 0.24, it can run but the loss_depth_consistency is too small compared to loss_sparse_flow. Is it normal?

and in the function "torch.nn.functional.grid_sample(input=im.permute(0, 3, 1, 2), grid=grid, mode='bilinear',
align_corners=True, padding_mode=padding_mode).permute(0, 2, 3, 1)" in _bilinear_interpolate(),
why the result is out of (0,1), as the the value of im is either 0 or 1.

[qqc222111]

Now the problem above is nearly solved by adjusting pre_process_data. I remain the number 0.9. The problem is that the loss_depth_consistency is still too small compared to loss_sparse_flow. Is it normal?

[qqc222111]

and the image is not trained completely.

[joelive]

Now the problem above is nearly solved by adjusting pre_process_data. I remain the number 0.9. The problem is that the loss_depth_consistency is still too small compared to loss_sparse_flow. Is it normal?

how do you adjusting pre_process_data to make loss_depth_consistency normal?