Why does a DeepGPLayer & the SoftMaxLikelihood both output MCMC samples?

[plucaci]

Apparently, I get too many samples than I was hoping for!

with torch.no_grad():
    for x_batch, y_batch in test_loader:
        with gpytorch.settings.num_likelihood_samples(1):
            x_batch, y_batch = x_batch.to(device), y_batch.to(device)
            output = model(x_batch)
            print("Model output:", output)

            preds = likelihood(output)
            print("Likelihood output:", preds)

I get:

Model output: MultitaskMultivariateNormal(mean shape: torch.Size([1, 1024, 5]))
Likelihood output: Categorical(logits: torch.Size([1, 1, 1024, 5]))

But I was expecting:

Model output: MultitaskMultivariateNormal(mean shape: torch.Size([1024, 5]))
Likelihood output: Categorical(logits: torch.Size([1, 1024, 5]))

Essentially, I was expecting only the SoftmaxLikelihood to output MCMC samples. Why does this happen?

[gpleiss]

@plucaci can you post a complete code example?