Are all agents involved in calculating the loss? How long one epoch takes in training?

[fengsky401]

I have three question:
1、I guess you only use target without other agents in calculating loss and propagating backward because you only generate one theta value in a sample scene, if not please give me more detail......If you use all agents in loss, you regard every agent as target, then your data preprocess code need to be modified for training
2、How long one epoch takes in training, and how many GPU did you use in experiment, which type GPU did you use.
3、How much improvement did the data augmentation give?