E40: DPPO's Gaussian MLP learns square under DPPO's own Gaussian PPO - #94
Merged
Merged
Conversation
DPPO fine-tunes a Gaussian MLP with PPO as its robomimic baseline, from pretrained checkpoints it releases. This adds that policy, built on the dppo package's ResidualMLP and CriticObs so the released square checkpoint loads by its own keys (2,152,485 parameters, as the paper states), and the options ppo needs to run DPPO's square config: critic_learning_rate a second parameter group (DPPO: actor 1e-4, critic 1e-3) n_critic_warmup_itrs iterations in which the actor gets no gradient max_grad_norm = None no clipping adam_eps CleanRL's 1e-5 by default; DPPO keeps 1e-8 reward_scaling_gamma the running return's own discount (DPPO: 0.99) `ppo dppo-square` carries ft_ppo_gaussian_mlp.yaml's values at irom-lab/dppo cc7234ad. The released checkpoint's logvar_max (a deviation of 1) replaces the config's 0.2 on load, as in DPPO; the policy keeps that and says so.
…n setting (before the pilot)
…ssian PPO, 3 of 3
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Stacked on #92 (dppo-gaussian-policy and ppo dppo-square), whose commit this branch contains: merge #92 first.
E40 was pre-registered at 456da16, after a two-iteration pilot. It runs DPPO's Gaussian MLP on robomimic square, started from DPPO's released checkpoint (its
modelweights), under every value of DPPO'sft_ppo_gaussian_mlp.yaml. Three seeds, 40 iterations of 80,000 steps, onguangzhao.modelweights with σ ≤ 1, the dppo-square config) hold.The Gaussian · PPO row's empty square cell becomes learns. That makes it the third pair to learn all four tasks. These are training-rollout rates, not DPPO's deterministic evaluations, and E40 does not claim to reproduce DPPO's curve.