Error from SMT --> ValueError: setting an array element with a sequence.

[Val4on]

After upgrading to version 2.4.0, I'm seeing the following errors:

TypeError: only length-1 arrays can be converted to Python scalars
krg_based.py", line 1127, in _reduced_likelihood_gradient
    grad_red[i_der] = (
ValueError: setting an array element with a sequence.

Is this a real issue or am I incorrectly using KPLS ?

[relf]

Hi. Thank you for reporting. I think we've got an issue here. How do you call KPLS(), what options are you using? What is the dimension of your training inputs and the value of n_comp?

In 2.4, we've changed the default internal optimizer from COBYLA to TNC. The latter uses gradients, I guess that if you switch back to COBYLA by using KPLS(hyper_opt="Cobyla", ...) you will retrieve a working solution.

[Val4on]

Hi, thanks for getting back, here’s how I’m calling it:

self.surrogate = KPLSK(print_global = False,
                               n_comp=num_params,
                               theta0=t0s,
                               print_prediction=False, corr='squar_exp')

Where I’m setting principal components to the number of input parameters, theta0 is of the same dimension as well. So, I do see a bunch of cobyla failures along with this message:

fmin_cobyla failed but the best value is retained Optimization failed. Try increasing the nugget

Is what gets printed before the error. So it sounds like I should go back to TNC? Correct.

Val.

[relf]

1. the point of using KPLS or KPLSK is to choose n_comp < num_params to get actual dimension reduction otherwise you'd better use KRG. What is the value of num_params? Whats is the shape of your training data (n_samples, n_dim) ? What is the shape/value of t0s?

2. what version of SMT worked for you before?

3. did you try to increase the nugget like it was suggested (option nugget=1e-8) when you test with Cobyla?

4. If you go back to TNC you get an error, correct? At the moment I can not reproduce the error you've got. So without an actual example to reproduce the error, it is difficult to help you more on this.

[Val4on]

Answers below. FYI, I will be out tomorrow and all of next week, so my responses will be delayed.

Val.

1. the point of using KPLS or KPLSK is to choose n_comp < num_params to get actual dimension reduction otherwise you'd better use KRG. What is the value of num_params? Whats is the shape of your training data (n_samples, n_dim) ? What is the shape/value of t0s?

ANS: I figured n_comp is number of principle components to retain, the default now was just to set it to num_params (I will change this), but you’re right we should change it. num_params is the size of the domain, this can vary from 2 to 30. Theta0 is the same as the number of params. We’re currently using discrete value indexes, so this defaults to 1.0.

1. what version of SMT worked for you before?

ANS: 2.0.1

1. did you try to increase the nugget like it was suggested (option nugget=1e-8) when you test with Cobyla?

ANS: I did try, but I still saw errors, I will re-run to verify

1. If you go back to TNC you get an error, correct? At the moment I can not reproduce the error you've got. So without an actual example to reproduce the error, it is difficult to help you more on this.

ANS: If I set hyper_opt to TNC, I still see : Optimization failed. Try increasing the nugget fmin_cobyla failed but the best value is retained

But it doesn’t crash. KPLSK(print_global = False, hyper_opt='TNC', n_comp=num_params, theta0=t0s, print_prediction=False, corr='squar_exp')