Repository navigation
E53: with camera frames and a CNN, SB3 still ends on byte-identical weights through the bridge - #119
Merged
Merged
Conversation
…image observations and a CNN, does SB3 end on the same weights through the bridge
…identical weights through the bridge; 16 frames a step cost 2.6-3.0 ms SB3 PPO CnnPolicy, 16 AtariWrapper(ALE/Pong-v5) envs, 40,960 steps, three seeds: same weights 3/3, same episodes, the bridge episode log agrees; +2.60-2.99 ms per vector step (10-11%). Every check and prediction holds.
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
The question
E50 found the boundary transparent with state vectors. With camera frames, a uint8 image space, a CNN and frame stacking, does SB3 still end on the same weights through the bridge? And what do 16 frames a step cost?
The answer
Yes, the weights are the same. SB3 PPO with CnnPolicy, 16
AtariWrapper(ALE/Pong-v5)environments, 40,960 steps:82f80c02…both60acba42…both6bf04095…bothEvery check and prediction holds. Each step moves 16 frames of 7 KiB twice, since this client does not reuse observations. That costs about 2.8 ms on the loopback, or 10–11% of this short CPU run.
The run uses plugrl-bridges'
image_key(PlugRL/plugrl-bridges#3). It is too short to learn Pong, and it is not meant to: it compares weights.