How can I train the SfSNet to achieve the paper’s result?

[dldl3D]

Hi,

I have used the 180k CelebA dataset as the real data(the normal, albedo and light are produced by the pretrained model) and the 180k sysnthetic data the author released. I have trained the model at 35000 iterations, but the training loss of albedo, light, normal and reconstruction are 320.308, 0.01084, 387.409 and 280.136. But the testing loss of albedo, light, normal and reconstruction are 715.851, 0.011, 3325.68 and 1513.35. The albedo, normal and reconstruction image are very bad. where is wrong？

The solver file is as follows: `test_iter: 1000

iter_size: 6

test_interval: 1000

base_lr: 0.01 lr_policy: "multistep" gamma: 0.1 stepvalue: 15000 stepvalue: 40000 stepvalue: 45000 display: 100 max_iter: 50000 momentum: 0.9 weight_decay: 0.0005 `

[wujiachen987]

@dldl3D Are you solve the problem?