Missing loss mask with hierarchical action space in behavior cloning script?

[phython96]

Hi, I found that finetuning foundation-model-1x starts with a high loss (14.0) using find-cave contractor data.
As described in section E.3, when using hierarchical action space, loss for the secondary mouse movement should be masked during training.
However, I didn't find the implementation in behavior_cloning.py, which may be the reason why the fintuning loss is large.
Looking forward to your reply, thanks.

[Miffyli]

Hey! Sorry for the delay; I have been OOF :)

Hmm can you pinpoint the phrase about "secondary mouse movement"? The action space should capture mouse movement regardless of the task/setup.

The high loss is also probably due to to different training setup; due to memory constraints, the BC example trains with very small sequences lengths (and one sample at a time). The behaviour of the players in the find-cave (or other BASALT task) is also significantly different from general Minecraft gameplay, so I am not that surprised the loss starts high.