Hi, did you check the gradients of this operators? about torch-ngp HOT 7 CLOSED

No, thanks for mentioning this! Could you tell me which part of code are you testing?

from torch-ngp.

No, thanks for mentioning this! Could you tell me which part of code are you testing?
I test it in the hash-encoder part, the visualized picture is grad_input, but also for the grad_emddings, if you check, you can print what you want in the torch.autograd. gradcheck and i hope you can check it, later i will test the tiny_cuda code.

Thank you very much;

from torch-ngp.

Thanks, I'll check that.

from torch-ngp.

Hi, can you provide the email, i can sent you the test_code, you can have a try;
If you are ok, you can sent to my email, my email is the [email protected];
and I check the tiny_cuda, it is true;

from torch-ngp.

@Dengzhi-USTC Thanks a lot! It's [email protected]; or you can paste it in this issue so it will be automatically sent to my email.

from torch-ngp.

Thank you very much. I rewrite the code with the tiny_cuda, and i solved this problem.
Later, i will report this;

from torch-ngp.

Thanks for the testing code! I have verified the problem.

1. For grad_embeddings, there is a problem if input range is not in [0, 1], and this has been fixed in 853dd51.
2. For grad_inputs, since there is still no use case in this repo, I'll focus on it later.

from torch-ngp.