The questions in data_source.py

[SeonghwanSeo]

Hello,

Thanks for your great work!

I work on the gflownet-based project, and I wonder the role of following code block:
https://github.com/recursionpharma/gflownet/blob/trunk/src/gflownet/data/data_source.py [Line 254-256]

# Override the is_valid key in case the task made some objs invalid
for i in valid_idcs:
    trajs[i]["is_valid"] = True

Can I know why this code block does not mask the invalid objectives' trajectories as follows?

# Override the is_valid key in case the task made some objs invalid
for i, traj in enumerate(trajs):
    if i not in valid_idcs:
        traj["is_valid"] = False

Best regards,
Seonghwan Seo

[bengioe]

Hi Seonghwan, I think you may be right, the True should be a False and be iterating over invalid idcs 🤔
Basically there are two sources of invalidity, the sampling method (e.g. an illegal action is taken) and the task itself. In practice, we've mostly managed to create action masks so that this doesn't happen, and the tasks we use typically don't decide that objects are invalid, so this issue may have gone unnoticed!

I don't have time to test & fix this now so leaving this open. If you have the time, a PR is welcome!