Why is the length of the reward limited to 3 or more? about ml-compiler-opt HOT 3 OPEN

I assume this is here (could you please use the permalink feature like I did - easier to follow)

Do you mean that you have cases where for every module in a particular corpus sample, there are at most 2 decisions? If there are hangs, I believe the right fix would be to handle that graciously, because there could well be modules where there are no decisions, so we should skip over and retry (i.e. resample). Probably in local_data_collector.py.

The value "2" there IIRC was an optimization - i.e. there was little to gain from such short trajectories.

from ml-compiler-opt.

Thank you! As you say, this happens when there are at most two decisions on each object.
At the behavioral cloning stage, if all the size of trajectories in the default trace are less than 3, the dataset is empty. As a result, hang occurs here.
On the other hand, I understand that such a small dataset is not practical.
If you don't mind, there may be something I can do to fix it, so I'll leave this issue open.

from ml-compiler-opt.

Sounds good, and patches are very welcome!

from ml-compiler-opt.