Re-ran some experiments from yesterday using longer max step sizes and multistep E2C. I also changed the sampling scheme to pick trajectories and times without replacement (but it doesn't have much of an effect for this robot). Here are the latent space reconstructions:
Single-step E2C, max_step=1
Single-step E2C, max_step=10
Multi-step E2C, max_step=1
I tried re-running the multistep training initialized to the single-step max_step=1 weights, but the gradients immediately blew up.
No comments:
Post a Comment