How long did it take to train Prime-agent and Mesh-agent?

[yudopan]

Hi Lin,

thanks for sharing your interesting and impressive work!
I am trying to implement it on my own machine. Could you please tell me how long it took you to train both agents and hardware specifications of the machine you used?

Thanks in advance!

Best,
yudopan

[clinplayer]

Hi, for one category, it takes about 4 days for Prim-Agent and about 2 days for Mesh-Agent, on a single GTX2080 card.

[yudopan]

thanks so much!

[yudopan]

Hi Lin, Happy New Year!
I tried you model on my machine. (RTX Titan and AMD Ryzen Threadripper 3990X).
For one category, it takes 8 days already to train Prim-Agent and it is still working and it has heavy computation on CPU. I guess something is wrong. For me it took one hour to train one RL Epoch, in default setting there are 2000 RL epochs.
Do you use the default setting in config.py?
And can you share me more details about training?
Thanks in advance!

[clinplayer]

Hey, sorry for not making this point clear. The epoch 2000 is just a random number to indicate the maximal training epochs. In practice, I keep observing the reward curve until it converges, and then stop the training manually. The other training details are the same as the file "config.py".