Extremely large gradient and vanishing images when using jax.precision=float32 about dreamerv3 HOT 3 OPEN

May I ask if you have successfully trained the dreamerv3 agent. I'm curious what the final loss of each component looks like，such as image, reward or cont. When the reconstructed images are very similar, I find that the reward's prediction is not as good as it should be. I'm not sure if it also affects subsequent strategy training. Thanks for sharing your thoughts.

from dreamerv3.

Unfortunately not, at least not when training on images in the walker environment

from dreamerv3.

Walker always worked for me from images, regardless of precision. I've just updated the paper and code, which has a better optimizer now. Curious if this is still an issue on your end.

from dreamerv3.