Guided self-play #142
Draft
Guided self-play #142
Conversation
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
* Add human ll loss with random actions [placeholder]. * Change clip coef * Typo * Integrate the actual human data into pipeline. * Back at 194 K FPS. Save data to disk, only load small batch to gpu when needed. * Ignore human data. * Revert small unrelated change. * Remove print statement * Return the continuous actions for sanity checks. * Log the weighted loss objectives instead. * Always compute human log prob for systematic analysis. * Bug fix: do not use human entropy. * Small fixes * Experimental settings. * Simplify expert data generation logic. * Small improvement * Store amount of human data used in checkpoints (each seq is 3 seconds of human data). * Cleanup/
* WIP guidance reward * remove irrelevant const * heading fix * minor refactor * fix normalize function defn * removed master weight * guidance reward sweeps
…#112) * start on obs space cfg * rmv magic numbers * Pass ini -> drive env. * Working on fuse linear * Integrate 700k SPS kernel. * Dirty * Profile fix * linear max kernel * Add linearmax kernels * LinearMax with layernorm: 600/700K SPS, score of 0.85 in < 5 min. * Use 1000 maps * Best run gets score of .95 in 25 minutes. * Advantage weighting exps * Sweep setup. * Increase steps. * Remove exploration logging. * Delete Triton kernel. * Best score is 0.98 with 600K FPS. 800 K FPS with 128 road points. * Better hparams * Replace default learning rate and ent_coef. * Minor * start on obs space cfg * rmv magic numbers * Pass ini -> drive env. * Working on fuse linear * Integrate 700k SPS kernel. * Dirty * Profile fix * linear max kernel * Add linearmax kernels * LinearMax with layernorm: 600/700K SPS, score of 0.85 in < 5 min. * Best run gets score of .95 in 25 minutes. * Advantage weighting exps * Sweep setup. * Increase steps. * Remove exploration logging. * Delete Triton kernel. * Best score is 0.98 with 600K FPS. 800 K FPS with 128 road points. * Better hparams * Clean up. * Minor * Update profiling script. * Adapt drive c network. * Temp settings * minor * Set max agents to 32 * Add linear -> relu -> max kernel. * Score is .98 after 70 minutes. * Minor --------- Co-authored-by: Joseph Suarez <jsuarez@puffer.ai>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
No description provided.