maybe some bugs?

[Dinngger]

1. In grid_encode.py, log2 in Python is not exactly same with std::log2 in C++. This can lead to a mismatch in scale in some cases.
2. The jt.nn.softplus is not the same as the softplus in PyTorch. I don't know if this will cause any problems.
3. In GridEncode, JNeRF implementation doesn't have the loss scale while tcnn has it (with loss_scale=128). This will cause gradient disappearance.
4. When computing alpha in sdf, JNeRF uses safe_clip(0.0, 1.0). I think clamp_(0.0, 1.0) should be used.
5. The cumprod function in jittor is unsafe! Have a look at PyTorch implementation or mine in Jeuralangelo.
6. JNeRF doesn't have the weight norm but I found it in Jittor's source. I don't know why.

[qianqiuzzz]

I met No.3 error，loss increases sharply during training. do you have any solution to fix it?

[Dinngger]

I met No.3 error，loss increases sharply during training. do you have any solution to fix it?

refer to my code, it might help you.