This fixes https://github.com/GFNOrg/torchgfn/issues/168.
The idea is to remove the arguments we had before off_policy and sample_off_policy, and be explicit about what we're evaluating and storing when sampling.
When being on_policy, we should store the logprobs. This is the default.
When being off_policy, with a tempered/modified PF, we should only store estimator_outputs.
When we use a replay buffer, we don't need to store anything - we should recalculate the logprobs.
Additionally, this fixes FM + ReplayBuffer, that was broken before, because states extension didn't take into account the _log_probs attribute.
This fixes https://github.com/GFNOrg/torchgfn/issues/168. The idea is to remove the arguments we had before
off_policy
andsample_off_policy
, and be explicit about what we're evaluating and storing when sampling. When being on_policy, we should store the logprobs. This is the default. When being off_policy, with a tempered/modified PF, we should only storeestimator_outputs
. When we use a replay buffer, we don't need to store anything - we should recalculate the logprobs.Additionally, this fixes FM + ReplayBuffer, that was broken before, because states extension didn't take into account the
_log_probs
attribute.