search

danijar / dreamer

Dream to Control: Learning Behaviors by Latent Imagination
https://danijar.com/dreamer
MIT License
506 stars 109 forks source link

slow in atari tasks #48

Closed xiangyyyy closed 3 years ago

xiangyyyy commented 3 years ago

I'm running Atari tasks on the nvidia 2080 and A100 respectively, but I only have around 20 FPS, which is too slow. That means it takes tens of days to run 10 million steps. When you run the Atari task, how much time does each task take? Is there a way to speed things up?

{"step": 1446778, "expl_amount": 0.09999962151050568, "model_grad_norm": 12.273571968078613, "value_grad_norm": 2023.7537841796875, "actor_grad_norm": 367.6117248535156, "prior_ent": 73.4467544555664, "post_ent": 70.40204620361328, "image_loss": 11264.0, "reward_loss": 1.7543749809265137, "pcont_loss": 0.5491357445716858, "div": 3.099745988845825, "model_loss": 11264.7998046875, "value_loss": Infinity, "actor_loss": -190.18875122070312, "action_ent": 0.013957500457763672, "fps": 20.45536337258022}

xiangyyyy commented 3 years ago

my command is the same as #33 python dreamer.py --logdir ./logdir/atari_mspacman/dreamer/2 --task atari_ms_pacman --expl epsilon_greedy --horizon 10 --kl_scale 0.1 --action_dist onehot --expl_amount 0.4 --expl_min 0.1 --expl_decay 100000 --pcont True --time_limit 1000000

danijar commented 3 years ago

Hi, please use the DreamerV2 code base for Atari experiments.