57 Commits

Author SHA1 Message Date
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
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
6f1a7a24ed try to fix ci 2025-10-21 11:47:39 -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
4a5465eeb6 fix ci 2025-10-21 09:17:53 -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
ea13d4fcab take a gradient step with video tokenizer trainer 2025-10-21 08:52:22 -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
374667d8a9 take care of the loss normalization mentioned at the end of the first paragraph of section 3 2025-10-19 08:24:41 -07:00
lucidrains
b6aa19f31e complete multi-token prediction for actions, tackle loss balancing another day 2025-10-18 10:23:14 -07:00
lucidrains
5fc0022bbf the function for generating the MTP targets, as well as the mask for the losses 2025-10-18 08:04:51 -07:00
lucidrains
22e13c45fc rename 2025-10-17 14:44:25 -07:00
lucidrains
cb416c0d44 handle the entropies during policy optimization 2025-10-17 08:47:26 -07:00
lucidrains
0dba734280 start the learning in dreams portion 2025-10-17 08:00:47 -07:00
lucidrains
a0161760a0 extract the log probs and predicted values (symexp two hot encoded) for the phase 3 RL training 2025-10-16 10:40:59 -07:00
lucidrains
2d20d0a6c1 able to roll out actions from one agent within the dreams of a world model 2025-10-16 10:15:43 -07:00
lucidrains
d28251e9f9 another consideration before knocking out the RL logic 2025-10-14 11:10:26 -07:00
lucidrains
9c78962736 sampling actions 2025-10-12 11:27:12 -07:00
lucidrains
8a73a27fc7 add nested tensor way for getting log prob of multiple discrete actions 2025-10-11 10:53:24 -07:00
lucidrains
b2725d9b6e complete behavior cloning for one agent 2025-10-11 09:24:49 -07:00
lucidrains
563b269f8a bring in hyper connections 2025-10-11 06:52:57 -07:00
lucidrains
5df3e69583 last commit for the day 2025-10-10 11:59:18 -07:00
lucidrains
9230267d34 handle subset of discrete action unembedding 2025-10-10 11:27:05 -07:00
lucidrains
32aa355e37 prepare unembedding parameters in ActionEmbedder as well as the policy head, to allow for behavioral cloning before RL 2025-10-10 10:41:48 -07:00
lucidrains
9101a49cdd handle continuous value normalization if stats passed in 2025-10-09 08:59:54 -07:00
lucidrains
31f4363be7 must be able to do phase1 and phase2 training 2025-10-09 08:04:36 -07:00
lucidrains
e2d86a4543 add a complete action embedder that can accept any number of discrete actions with variable bins as well as any number of continuous actions, pooled and added to the agent token as described in the paper (seems like they fixed that horrendous hack in dreamer v3 with sticky action) 2025-10-09 07:53:42 -07:00
lucidrains
c4e0f46528 for the value head, we will go for symexp encoding as well (following the "stop regressing" paper from Farebrother et al), also use layernormed mlp given recent papers 2025-10-08 07:37:34 -07:00
lucidrains
187edc1414 all set for generating the perceived rewards once the RL components fall into place 2025-10-08 06:33:28 -07:00
lucidrains
36ccb08500 allow for step_sizes to be passed in, log2 is not that intuitive 2025-10-07 08:36:46 -07:00
lucidrains
a8e14f4b7c oops 2025-10-07 08:09:33 -07:00
lucidrains
c6bef85984 generating video with raw teacher forcing 2025-10-07 07:22:57 -07:00
lucidrains
83ba9a285a reorganize tokenizer to generate video from the dynamics model 2025-10-06 11:37:45 -07:00
lucidrains
7180a8cf43 start carving into the reinforcement learning portion, starting with reward prediction head (single for now) 2025-10-06 11:17:25 -07:00
lucidrains
25b8de91cc handle spatial tokens less than latent tokens in dynamics model 2025-10-06 09:19:27 -07:00
lucidrains
f507afa0d3 last commit for the day - take care of the task embed 2025-10-05 11:40:48 -07:00
lucidrains
fe99efecba make a first pass through the shortcut training logic (Frans et al from Berkeley) maintaining both v-space and x-space 2025-10-05 11:17:36 -07:00
lucidrains
971637673b complete all the types of attention masking patterns as proposed in the paper 2025-10-04 12:45:54 -07:00
lucidrains
5c6be4d979 take care of blocked causal in video tokenizer, still need the special attention pattern for latents to and from though 2025-10-04 12:03:50 -07:00
lucidrains
6c994db341 first nail down the attention masking for the dynamics transformer model using a factory function 2025-10-04 11:20:57 -07:00
lucidrains
895a867a66 able to accept raw video for dynamics model, if tokenizer passed in 2025-10-04 06:57:54 -07:00
lucidrains
8373cb13ec grouped query attention is necessary 2025-10-04 06:31:32 -07:00
lucidrains
046f8927d1 complete the symexp two hot proposed by Hafner from the previous versions of Dreamer, but will also bring in hl gauss 2025-10-03 08:08:44 -07:00
lucidrains
8d1cd311bb Revert "address https://github.com/lucidrains/dreamer4/issues/1"
This reverts commit e23a5294ec2f49d58d3ccb936c498eb86939fa96.
2025-10-02 12:25:05 -07:00
lucidrains
e23a5294ec address https://github.com/lucidrains/dreamer4/issues/1 2025-10-02 11:49:22 -07:00