misc

.vscode/launch.json

@@ -19,8 +19,8 @@
                 "env.train.id=CrafterReward-v1",
                 // # make it start early
                 "training.tokenizer.start_after_epochs=1",
-                "training.world_model.start_after_epochs=1",
-                "training.actor_critic.start_after_epochs=1",
+                "training.world_model.start_after_epochs=2",
+                "training.actor_critic.start_after_epochs=3",
                 "training.tokenizer.steps_per_epoch=10",
                 "training.world_model.steps_per_epoch=10",
                 "training.actor_critic.steps_per_epoch=10",

research_journal.md

@@ -345,9 +345,6 @@ Why is it not learning? It's because the dynamics model is total BS!!!
 # 2023-11-19 06:45:34
 
 So I tried just trainign the world model for 200 epochs. And with a post_embedding layer. It helped the flickering. But not enougth to actually go for the obvious local minima of the next state equals the last
-
-
-
   
 idea
 - bypass embedding?, but wait dreamerv3 needed quant z... 
@@ -357,3 +354,7 @@ idea
 
 
 ok we need to freeze it, and change dtype
+OK it seems slightly better yay! Lets train it overnight and see
+
+next idea is to the delta-IRIS thing where the tokens only have to encode the diff(obs)
+

scripts/play.sh

@@ -27,6 +27,7 @@ while [ "$1" != "" ]; do
         -e | --episode )
             mode="episode_replay"
             ;; # replay train, test, or imagined episodes. shows quality of dynamics model
+            # this is quick low resource way to check the dynamics model and agent while training
         -w | --world-model )
             mode="play_in_world_model"
             ;; # human plays in world model
@@ -37,4 +38,4 @@ while [ "$1" != "" ]; do
     shift
 done
 
-python -m pdb src/play.py hydra.run.dir=. hydra.output_subdir=null +mode="${mode}" +fps="${fps}" +header="${header}" +reconstruction="${reconstruction}" +save_mode="${save_mode}"
+python src/play.py hydra.run.dir=. hydra.output_subdir=null +mode="${mode}" +fps="${fps}" +header="${header}" +reconstruction="${reconstruction}" +save_mode="${save_mode}"

src/trainer.py

@@ -200,7 +200,7 @@ class Trainer:
         if epoch > cfg_world_model.start_after_epochs:
             metrics_world_model = self.eval_component(self.agent.world_model, cfg_world_model.batch_num_samples, sequence_length=self.cfg.common.sequence_length, tokenizer=self.agent.tokenizer)
 
-        if epoch > cfg_actor_critic.start_after_epochs:
+        if epoch > cfg_world_model.start_after_epochs:
             self.inspect_imagination(epoch)
 
         if cfg_tokenizer.save_reconstructions: