Tensor shape when predicting from Clustered Multitask GP Regression

[tcristea]

Hi,

I've run this example:
https://github.com/cornellius-gp/gpytorch/blob/master/examples/07_Pyro_Integration/Clustered_Multitask_GP_Regression.ipynb
When in the 6th code cell, the fragment:

model.eval()  
with torch.no_grad():  
   test_x = torch.linspace(0, 1, 51)
   predictions = model.likelihood(model(test_x)).rsample()
print(predictions.shape)  # num_samples x num_data x num_tasks

the shape of prediction tensor is [10, 51, 4], i.e., num_data = 51 (size of test_x), num_tasks = 4 (set to the number of train_y features), while num_samples is always comes out as 10.
The value of num_samples (= 10) is set by this statement model.likelihood(model(test_x)).
Where is the "10" coming from? I've read the code but I'm not any wiser.

Best regards,
TC

[wjmaddox]

It's this setting: https://docs.gpytorch.ai/en/stable/settings.html#gpytorch.settings.num_likelihood_samples

Basically when latent GPs are evaluated, gpytorch automatically draws several samples for usage in MC estimation over the latent GP.

[tcristea]

Thank you!