103 Commits

Author SHA1 Message Date
lucidrains
d756d1bb8c addressing issues raised by an independent researcher with llm assistance 2025-10-31 08:37:39 -07:00
lucidrains
60681fce1d fix generation so that one more step is taken to decode agent embeds off the final cleaned set of latents, update readme 2025-10-31 06:48:49 -07:00
lucidrains
3beae186da some more control over whether to normalize advantages 2025-10-30 08:46:03 -07:00
lucidrains
0904e224ab make the reverse kl optional 2025-10-30 08:22:50 -07:00
lucidrains
767789d0ca they decided on 0.3 for the behavioral prior loss weight 2025-10-29 13:24:58 -07:00
lucidrains
35b87c4fa1 oops 2025-10-29 13:04:02 -07:00
lucidrains
c4a3cb09d5 swap for discrete kl div, thanks to Dirk for pointing this out on the discord 2025-10-29 11:54:18 -07:00
lucidrains
cb54121ace sim trainer needs to take care of agent embedding and old actions 2025-10-29 11:15:11 -07:00
lucidrains
586379f2c8 sum the kl div loss across number of actions by default for action embedder .kl_div 2025-10-29 10:46:42 -07:00
lucidrains
a358a44a53 always store old agent embeds and old action parameters when possible 2025-10-29 10:39:15 -07:00
lucidrains
3547344312 take care of storing the old action logits and mean log var, and calculate kl div for pmpo based off that during learn from experience 2025-10-29 10:31:32 -07:00
lucidrains
691d9ca007 add kl div on action embedder, working way towards the kl div loss in pmpo 2025-10-29 10:02:25 -07:00
lucidrains
91d697f8ca fix pmpo 2025-10-28 18:55:22 -07:00
lucidrains
7acaa764f6 evolutionary policy optimization on dreams will be interesting 2025-10-28 10:17:01 -07:00
lucidrains
c0450359f3 allow for evolutionary policy optimization 2025-10-28 10:11:13 -07:00
lucidrains
46f86cd247 fix storing of agent embedding 2025-10-28 09:36:58 -07:00
lucidrains
903c43b770 use the agent embeds off the stored experience if available 2025-10-28 09:14:02 -07:00
lucidrains
d476fa7b14 able to store the agent embeddings during rollouts with imagination or environment, for efficient policy optimization (but will also allow for finetuning world model for the heads) 2025-10-28 09:02:26 -07:00
lucidrains
789f091c63 redo so that max timesteps is treated as truncation at the last timestep, then allow for accepting the truncation signal from the environment and reuse same logic 2025-10-28 08:04:48 -07:00
lucidrains
995b1f64e5 handle environments that return a terminate flag, also make sure episode lens are logged in vectorized env 2025-10-27 10:14:28 -07:00
lucidrains
fe79bfa951 optionally keep track of returns statistics and normalize with them before advantage 2025-10-27 09:02:08 -07:00
lucidrains
f808b1c1d2 oops 2025-10-27 08:34:22 -07:00
lucidrains
349a03acd7 redo so lens is always the episode length, including the bootstrap value timestep, and use is_truncated to mask out the bootstrap node from being learned on 2025-10-27 08:06:21 -07:00
lucidrains
59c458aea3 introduce an is_truncated field on Experience, and mask out rewards and values before calculating gae appropriately 2025-10-27 07:55:00 -07:00
lucidrains
fbfd59e42f handle variable lengthed experiences when doing policy optimization 2025-10-27 06:09:09 -07:00
lucidrains
46432aee9b fix an issue with bc 2025-10-25 12:30:08 -07:00
lucidrains
f97d9adc97 oops, forgot to add the view embedding for robotics 2025-10-25 11:39:06 -07:00
lucidrains
32cf142b4d take another step for variable len experiences 2025-10-25 11:31:41 -07:00
lucidrains
4d8f5613cc start storing the experience lens 2025-10-25 10:55:47 -07:00
lucidrains
3d5617d769 take a step towards variable lengthed experiences during training 2025-10-25 10:45:34 -07:00
lucidrains
4ce82f34df given the VAT paper, add multiple video streams (third person, wrist camera, etc), geared for robotics. need to manage an extra dimension for multiple viewpoints 2025-10-25 09:20:55 -07:00
lucidrains
a9b728c611 incorporate proprioception into the dynamics world model 2025-10-24 11:24:22 -07:00
lucidrains
35c1db4c7d sketch of training from sim env 2025-10-24 09:13:09 -07:00
lucidrains
27ac05efb0 function for combining experiences 2025-10-24 08:00:10 -07:00
lucidrains
d0ffc6bfed with or without signed advantage 2025-10-23 16:24:29 -07:00
lucidrains
fb3e026fe0 handle vectorized env 2025-10-22 11:19:44 -07:00
lucidrains
7ecc5d03e8 wire up the time kv cache when interacting with sim / env 2025-10-22 08:39:11 -07:00
lucidrains
d82debb7a6 first pass through gathering experience with a mock env for online rl 2025-10-22 08:32:46 -07:00
lucidrains
03b16a48f2 sketch out the dream trainer, seems like they only fine tune the heads 2025-10-22 06:41:10 -07:00
lucidrains
40da985c6b tweak bc trainer 2025-10-21 10:55:24 -07:00
lucidrains
2fc3b17149 take a gradient step with behavioral clone trainer, make sure it works with and without actions and rewards 2025-10-21 10:20:08 -07:00
lucidrains
283d59d75a oops 2025-10-21 09:50:07 -07:00
lucidrains
b34128d3d0 make sure time kv cache can be passed back in during generation 2025-10-21 09:15:32 -07:00
lucidrains
7ba3988fb9 prepare a mock for interacting with online env 2025-10-21 09:03:20 -07:00
lucidrains
ea13d4fcab take a gradient step with video tokenizer trainer 2025-10-21 08:52:22 -07:00
lucidrains
15876d34cf more muon prep 2025-10-21 08:23:59 -07:00
lucidrains
b4763caff9 fix rotary embeddings in presence of kv caching 2025-10-21 07:10:21 -07:00
lucidrains
7195bbb196 oops 2025-10-20 12:42:27 -07:00
lucidrains
ca244a290c first pass through the kv cache for the time block in the dynamics model 2025-10-20 12:25:50 -07:00
lucidrains
a7e0c395c3 allow for only rmsnorm for keys in attention 2025-10-20 11:20:49 -07:00