denisyarats / pytorch_sac_ae Goto Github PK

I wonder if everyone else has the same problem. Environments like cheetah run or walker need way more RAM than other environments like the ball in the cup. Perhaps three times more ram consumption.

I am interested in running benchmarks on noisy observations as described in section 5.2 of https://arxiv.org/abs/1910.01741.
The code for this part would be greatly appreciated as it apparently is missing from the repo at the moment.

hello:
line 247 in sac_ae.py:self.actor.encoder.copy_conv_weights_from(self.critic.encoder) just copy parameters from critic encoder to actor encoder ,when you initialize sac agent.
Why not synchronize the critic encoder parameters to the actor's encoder after each critic parameters update ?

What is the recommended way to recreate the sac:pixel from the results? Modify Pytorch_sac, or sac_ae, or is code available? (I know changes are trivial in some sense but I'd just like to keep the implementation as similar as possible)

Hi! @denisyarats

It's been a long time since this repo was published. Still hoping you'll respond to this issue.

Have you ever tried this method with a larger image size? In your code, the height and width of 84 is the used as the biggest size.

I tried to increase the size to 128 but the memory cost reached almost 100GB.

For the tested dm_control tasks, the image size of 84 seems to be adequate, but may not be enough for other tasks.

Would you mind sharing some of your opinions? Thanks a lot!