Katsuki Ohto
Katsuki Ohto
This is ugly and complicated, I think...
Thank you for your detailed report! Since episodes among agents at the beginning are very short in some environments, it's not strange that used memory increases after the replay buffer...
@nizhihao There is no precision problem with `compress_steps`. It only slows down the speed of making batches. If this speed is not enough to use GPUs without a break, training...
@digitalspecialists Thank you for your helpful report! So, is there a problem around CUDA 11.2? We'll look into it when we get a chance.
@digitalspecialists I understand what you pointed out. Yes, that's necessary. Thanks!
@han-x Massively thank you for your report. You are training a big neural network, aren't you? In current implementation, each Gather process store all models after starting workers. Rewrite worker.py...
@han-x Thanks. That's a strange case... Since the original GeeseNet was about 500KB, it will occupy only 100MB when in 200 epochs. It's nothing wrong 129k episodes occupied 10GB in...
Thank you for your question, @Riad123321. As far as I know, HandyRL can automatically use GPUs for training on Kaggle notebook. We don't support using GPUs even in the generation...
If you would like to use GPUs for episode generation, the implementation of SEED will help you! https://github.com/google-research/seed_rl
fixed and updated