felix-kerkhoff / dqfd Goto Github PK

Hi, I am running this on a personal computer with a 3070ti and 32gb of ram.

I am immediately running out of memory with a error saying it can't allocate 9gb of memory.

What are the system requirements for this or do I need to optimise my setup? I am using the packages and versions in requirements.txt.

Thanks.

I use pytorch to implement the DQfD as well, for pure DQN, the result seems to meet expectations. however, when I use DQfD, after the pretrain process, the score result becomes worse. I only use the Je loss .
My strategy is first pretrain, then set epsilon to 1 to random explore and decrease 0.001 every game step. which means, after pretrain on demonstration data, the subsequent process is the normal DQN.
Can you give me some suggestion about it.