Large overhead for log p computation in DOMIAS

[JimAchterbergLUMC]

Description

When evaluating log probabilities for the DOMIAS metric, gradient computation is not disabled, causing massive overhead when the test set is large.

During training, log probabilities are computed in batches. However, during inference this is not the case, causing massive overhead since gradients are still being computed.

How to Reproduce

1. Go to 'synthcity/metrics/eval_privacy.py', line 594. Here we see that log probabilities are computed for the entire test set at once, without disabling gradients. When the test set is made to be reasonably large, this (unnecessarily) takes up a large amount of memory.

Expected Behavior

Disable gradient computation during inference when computing log probabilities for DOMIAS.