czq142857
commented
2 years ago Hi,
During the noisy input experiment, we use SIREN to overfit the SDF of each shape, so the noise amount is actually small. According to your numbers, your input noise must be much larger. In that case, you could try to increase the receptive field of the network to resolve the ambiguity introduced by the noisy inputs, i.e., try using the network for voxels. Also, you may have noticed that if the inputs are SDF grids, the network will not predict the signs of the grid vertices, but use the signs of the input signed distances as the ground truth. This works well when the input is clean or has small noises, but may cause issues when the noise is larger. In that case, you may need to modify the code a bit to treat the input more like voxels.
Best, Zhiqin
Howardyangyixuan
Hi, zhiqin
I am trying to use the noisy inputs to train the NMC model from scratch to get similar results from paper but found the loss decrease far too slow compared to clean inputs. for bool part: [0/400] time: 1366 loss: 0.25179240 loss_bool: 0.91642386 loss_float: 0.00000000 [1/400] time: 2282 loss: 0.25031856 loss_bool: 0.91567051 loss_float: 0.00000000 [2/400] time: 3002 loss: 0.25059867 loss_bool: 0.91625684 loss_float: 0.00000000 [3/400] time: 3594 loss: 0.25166979 loss_bool: 0.91583282 loss_float: 0.00000000 [4/400] time: 4078 loss: 0.25226465 loss_bool: 0.91645384 loss_float: 0.00000000 ... [275/400] time: 45009 loss: 0.24075049 loss_bool: 0.91662174 loss_float: 0.00000000 [276/400] time: 45091 loss: 0.24094345 loss_bool: 0.91680372 loss_float: 0.00000000 [277/400] time: 45175 loss: 0.23898503 loss_bool: 0.91793585 loss_float: 0.00000000 [278/400] time: 45257 loss: 0.23988940 loss_bool: 0.91663897 loss_float: 0.00000000 [279/400] time: 45340 loss: 0.24025825 loss_bool: 0.91807079 loss_float: 0.00000000
Does it also occur to you when you train it? How many epochs did you train for both the bool and float part? and would you like to share noisy training weights with me to see more results from it? Great thanks!