Excuse me , why the param training=True in eval? If training=True in eval, the mode in handshake will be softmax instead of argmax/activated_selected.

[WangzcBruce]

MultiAgentPerception/ptsemseg/trainer.py

Line 968 in 4ef3005

outputs, _, action_argmax, _ = self.model(images_val, training=True, MO_flag=self.MO_flag)

[drowning-in-codes]

Actually I find that it uses argmax during training in Who2Com which confuses me. And I why the action_argmax is used as a metrics in validation or test but not used during training.

I'm also confused that why the dim here is 2 which is not the agent-axis.
Have you solved the problem？I hope to get your answer sincerely.
https://github.com/GT-RIPL/MultiAgentPerception/blob/4ef300547a7f7af2676a034f7cf742b009f57d99/ptsemseg/models/agent.py#L625C1-L627C45

[nzhc]

@drowning-in-codes Do you still have the dataset for this paper? The download link provided by the author is no longer valid. If you have it, could you please send me a copy? Thank you very much (my email: 15351413870@163.com ）