wassname/iris_bigvae
1
0
Fork 0
mirror of https://github.com/wassname/iris_bigvae.git synced 2026-08-21 11:16:18 +08:00
Code Issues Packages Projects Releases Wiki Activity
48 Commits 7 Branches 0 Tags
main
Commit Graph
33 Commits
Author SHA1 Message Date
wassname a4234425af checkin 2024-05-11 10:11:10 +08:00
wassname 991f157fdd Merge branch 'full_ft' 2024-04-27 07:04:40 +08:00
wassname 8ecdf35bd3 learns the best yet 2023-11-25 05:48:17 +08:00
wassname 9155ecca95 Fix embedding path in Transformer model 2023-11-24 09:24:12 +08:00
wassname c3878e4a4c full finetune? 2023-11-23 09:08:30 +08:00
wassname 72ee806601 try training tokenizer and world model together: result 💩 2023-11-23 09:08:10 +08:00
wassname 2763449f64 bug fix 2023-11-20 10:01:56 +08:00
wassname 56daac9761 rank, and unfreeze head 2023-11-20 07:31:31 +08:00
wassname 6308397bb7 it works better but still flickers when imagining. hmm I do think I need the delta-iris stuff 2023-11-20 07:14:50 +08:00
wassname 6eb25883f6 misc 2023-11-19 14:50:54 +08:00
wassname ecee11fc33 notes 2023-11-19 10:58:45 +08:00
wassname 328d281651 using frozen transformer embedding: runs 
- use frozen transformer encoder where possible (obs, worldmodel, tokenizer, etc)
- it's frozen
- a huge vocab
 2023-11-19 10:50:48 +08:00
wassname 7098b67955 1 2023-11-19 08:48:36 +08:00
wassname 36ea66d0ae misc 2023-11-19 08:17:38 +08:00
wassname 9b399031ca document play.sh options 2023-11-18 16:32:59 +08:00
wassname a341e7aaad abs imports, fix play for crafter 2023-11-18 07:44:05 +08:00
wassname 1dabd79251 debugging model size and speed 2023-11-17 18:17:17 +08:00
wassname a778296e40 relative imports 2023-11-17 12:30:16 +08:00
wassname 1af7aa74fa 3min epoch breakout 
https://wandb.ai/wassname/iris/runs/1zjsneyu?workspace=user-wassname
 2023-11-17 06:25:09 +08:00
wassname 1b6462991b smaller horizon, smaller lstm. 100x faster actor_critic. does it learn though? 2023-11-16 17:17:57 +08:00
wassname 106027a7f9 add 1b model, use contextlib, 2023-11-14 10:42:17 +08:00
wassname e495db47e4 Fix num_passes calculation in WorldModelEnv.step() 2023-11-13 17:17:03 +08:00
wassname e417790413 fix kv_cache faking? 2023-11-13 08:11:13 +08:00
wassname b97a499a19 Fix bugs and improve performance in training 
process
 2023-11-13 07:00:17 +08:00
wassname a3ded691e1 misc 2023-11-12 17:27:17 +08:00
wassname 1454e5c053 training 2023-11-12 16:13:23 +08:00
wassname a6b850bd96 fix debug launch 2023-11-12 14:44:10 +08:00
wassname 28910cfd63 poetry 2023-11-12 13:57:33 +08:00
wassname 075de416c8 adding bigvae 2023-11-12 12:56:55 +08:00
Eloi Alonso ac6be401fe Remove unused sampling_weights. 2023-09-18 16:12:04 +02:00
Eloi Alonso faa2d343d2 Add mechanism to save frames when visualizing. Add option to visualize agent playing in the world model. 2022-10-03 16:36:12 +02:00
Eloi Alonso 03290c820e Replace lambda with partial in env instantiation. 
Following hexaflexa suggestion in Issue #4.
 2022-09-19 10:31:03 +02:00
Eloi Alonso 75e5ab57e8 Code release. 2022-09-01 18:02:19 +02:00