Question about cache_warmup.ipynb

[yjb6]

When I run the cache_warmup.ipynb script, the loss becomes NaN during the second epoch. What could be causing this issue? I haven't made any modifications to the code after cloning the repository.

[yjb6]

I believe there's an issue when calculating quaternion_loss. After replacing it with L1 loss, the code runs without issues, but the NTC performance doesn't match the official results. The scenario is flame_steak.

[SJoJoK]

Please refer to https://github.com/SJoJoK/3DGStream/issues/8 which has the same issue as you. The tiny-cuda-nn may preform wrongly in different environments.

[yjb6]

I think I have found the issue. My environment is:

python                    3.10.13        
tinycudann                1.7                    
torch                     2.2.0+cu118    

"As mentioned in #8 , different versions can cause problems. In my version, when calculating cos_theta = F.cosine_similarity(q1, q2, dim=1) in def quaternion_loss, the gradient backpropagation computes the 3/2 power of q1, which exceeds the precision of torch.float16, causing NaNs. The following code can verify this situation:

for ntc_conf_path in ntc_conf_paths:    
    with open(ntc_conf_path) as ntc_conf_file:
        ntc_conf = ctjs.load(ntc_conf_file)
    ntc=tcnn.NetworkWithInputEncoding(n_input_dims=3, n_output_dims=4, encoding_config=ntc_conf["encoding"], network_config=ntc_conf["network"]).to(torch.device("cuda"))
    ntc_optimizer = torch.optim.Adam(ntc.parameters(), lr=1e-4)

    xyz = torch.rand(2000,3).cuda()
    ntc_output=ntc(xyz)
    # .to(torch.float64)
    gt = torch.tensor([1,0,0,0],dtype=torch.float64).cuda()

    cos_theta1 = F.cosine_similarity(ntc_output, gt, dim=1)

    cos_theta2 = torch.clamp(cos_theta1, -1+1e-7, 1-1e-7)

    loss = 1-torch.pow(cos_theta2, 2).mean()
    ntc_grad = autograd.grad(loss, ntc_output, retain_graph=True)[0]
    print(ntc_grad)
    print(ntc_grad.isnan().sum())
    print(ntc_output[ntc_grad.isnan().sum(dim=-1)!=0])
    test = ntc_grad[ntc_grad.isnan().sum(dim=-1)!=0]
    print(torch.pow(test,1.5))

Therefore, increasing the precision of q1 can solve this problem. The code can be modified as shown below: ntc_output=ntc(ntc_inputs_w_noisy).to(torch.float64)

[SJoJoK]

Nice job, thank you! I'll mention this modification in README in next few days.

[SJoJoK]

The README is updated.