wing3s / pysc2-rl-mini Goto Github PK

In worker.py, we need to sample experience by using local model and then backpropagate the gradients of total loss. The total loss is composed of loss_of_actor, loss_of_critic and entropy.

In line 132, you have already multiply 0.5 to value_loss_vb, but in line 152 you to that again. Does it mean the factor of critic loss in total loss is 0.25? Is it what you want to do?

Thank you. :)

[Question1]
In model.py, line 107 - 117:
The following layer's weight was initalized by xavier init but at line 111 they are also initialized by weight_norm. Why do we initialize their weights twice?

• ns_fc3
• nsa_fc4
• nsc_fc4

[Question2]
In the same file, line 119:
Why do we need to add this line: self.train(). According to pytorch official document, it says:

This has any effect only on modules such as Dropout or BatchNorm.

Does this mean we can remove this line?

Thank you, this repo is awesome and with nice coding structure.