In one_hot_actions() the 15th dim of the returned output is described in the code as cameraX, and the 16th dim as cameraY.
However, the actual convention of the VPT contractor dataset, as well as what it seems like the model was trained with, cameraY comes first (down is positive) and cameraX is second (right is positive). I.e. if you set dimension 15 of the actions to positive, the video generated by Oasis will have the POV look down.
So, both instances of cameraX (in ACTION_KEYS and in one_hot_actions) should be replaced with cameraY in the code, and similarly both instances of cameraY should be replaced withcameraX.
In
one_hot_actions()the 15th dim of the returned output is described in the code as cameraX, and the 16th dim as cameraY.However, the actual convention of the VPT contractor dataset, as well as what it seems like the model was trained with, cameraY comes first (down is positive) and cameraX is second (right is positive). I.e. if you set dimension 15 of the actions to positive, the video generated by Oasis will have the POV look down.
So, both instances of
cameraX(inACTION_KEYSand inone_hot_actions) should be replaced withcameraYin the code, and similarly both instances ofcameraYshould be replaced withcameraX.