torch.linalg_cholesky error

[hjf1997]

Thanks for your kind work

I'm running you code at nonstat_gp_cat. However, the code constantly throws an error:

torch.linalg_cholesky: The factorization could not be completed because the input is not positive-definite (the leading minor of order 314 is not positive-definite).

It seems that you have code to reinitialize all the parameters at gp_train.py:

        # except Exception as e:
        #     loss_val = np.inf
        #     torch.cuda.manual_seed(rand)
        #     rand+=1
        #     init_vars(model)
            pprint('Iter', i, 'Batch', bi, 'Failed with', e, 're inited the params')
            # break

I also tried this for many time, but the model still cannot get the right initializations. This reinitialization of all the parameters is the only way to solve this problem? Do you have other suggestions?

Thanks in advance :)

[hjf1997]

Hi, are you here?

[patel-zeel]

Hi @hjf1997. Sorry for not replying earlier as I missed the notification. Are you trying to use gp_train.py for your own dataset?

[hjf1997]

Hi @patel-zeel No, I'm using the same Beijing Airquality dataset processed by your .jupyter file.
Besides the 'not positive-definite' error, I also met some nan errors during backpropagation.
Did you also meet this error?

[patel-zeel]

Yes, we got those errors earlier but they depend on way too many things. To help you better (and to solve potential bugs in our code), would you like to explain step by step the bare minimum process you tried so that I can reproduce the error (I'd appreciate a direct code that I can execute and hopefully see the same error as you are seeing)? please also share the torch and gpytorch versions you have used.

[hjf1997]

OK, thanks for your kind help. I will first try to fix the problem and will give you feedback in a few days if there is any problem.

[hjf1997]

BTW, will you release the processing code for the London dataset?

[tiantian0612]

hello，do you meet "No module named 'nsgp'"error?
running lib.py,there is a bug:
from nsgp.utils.inducing_functions import f_kmeans, f_random
ModuleNotFoundError: No module named 'nsgp'

Can you tell me how to solve it?Thank you