Yuanzhao Zhai

Thanks for your reply. I also noticed that X = X.cpu().detach().numpy() and y = y.cpu().detach().numpy() in the OLELoss class. How can the gradient backdrop?

Thanks very much. I just believe that the backward() function can not optimize the parameter of the network which generate X and y. BTW, maybe torch.svd can compute SVD on...

More than 800 steps shown in the tensorboard, where I find the EvalAvgReturn is converge to 10.8. By the way, dicg_de_smac_runner.py --map 8m_vs_9m can achieve 0.9 win rate in the...