issues
search
jbloomAus
/
DecisionTransformerInterpretability
Interpreting how transformers simulate agents performing RL tasks
https://jbloomaus-decisiontransformerinterpretability-app-4edcnc.streamlit.app/
MIT License
61
stars
15
forks
source link
issues
Newest
Newest
Most commented
Recently updated
Oldest
Least commented
Least recently updated
Investigate/Possibly Add Gated MLP Units to Transformer Models.
#59
jbloomAus
closed
1 year ago
3
Investigate the effect of Dropout / Stochastic Depth on Model training/interpretability
#58
jbloomAus
opened
1 year ago
0
Add lr scheduling options to the DT training code
#57
jbloomAus
closed
1 year ago
1
Do an experiment where you turn off the weighted random sampler and/or visualize sampling prob distribution.
#56
jbloomAus
closed
1 year ago
1
Ensure that analysis app is able to fold analysis models that use LNPre
#55
jbloomAus
closed
1 year ago
2
Store checkpoints on wandb during offline training
#54
Felhof
opened
1 year ago
1
Explore Improvements to DT Training Procedure
#53
jbloomAus
closed
1 year ago
6
Train a model using layer norm pre to see if this helps formation of calibrated, performant memory env agents.
#52
jbloomAus
closed
1 year ago
1
Upgrade Collect Demonstrations Workflow
#51
jbloomAus
opened
1 year ago
0
Create an Object-Vector Calculator in the Streamlit App
#50
jbloomAus
opened
1 year ago
0
Fixed run_ppo being broken
#49
victorlf4
closed
1 year ago
1
Fixed run_ppo being broken
#48
victorlf4
closed
1 year ago
0
Investigate the effects of training on data sampled using different strategies created in #46
#47
jbloomAus
closed
1 year ago
1
Write a Rollout Sampling Utility for PPO Agents and add features affect generated distribution.
#46
jbloomAus
closed
1 year ago
1
Investigate whether anyone else does/ just experiment with finetuning of PPO models without entropy at the end of training to remove entropy optimising behaviors.
#45
jbloomAus
opened
1 year ago
1
Mega Card: Improve Analysis App in various ways to facilitate better interpretability analysis of the new models
#44
jbloomAus
opened
1 year ago
3
Set padded RTG in training data to be true RTG until masking is implemented correctly.
#43
jbloomAus
closed
1 year ago
1
Update the app to also work with BC models
#42
jbloomAus
opened
1 year ago
0
Add isort to the pre-commit config yaml without causing an issue with black
#41
jbloomAus
opened
1 year ago
0
Write Up Analysis of Memory Env Solution.
#40
jbloomAus
opened
1 year ago
2
Add checkpoints during Offline Training
#39
jbloomAus
opened
1 year ago
0
Added black formatter
#38
ruphail
closed
1 year ago
1
Major: Convert the codebase into an installable python package.
#37
jbloomAus
opened
1 year ago
0
Write a utility for merging sampled rollouts into a single file
#36
jbloomAus
opened
1 year ago
2
Update ppo checkpoints code to upload each checkpoint in real time rather than all at the end of the workflow
#35
jbloomAus
opened
1 year ago
0
Error while sampling from new trajectories generated by LSTM model
#34
jbloomAus
closed
1 year ago
1
Load and sample from model checkpoints
#33
jbloomAus
closed
1 year ago
2
work out a nice way to manage CLI args across all the new PPO Model
#32
jbloomAus
opened
1 year ago
0
Write a Probe Environment that tests a models ability to look at previous observations
#31
jbloomAus
opened
1 year ago
1
Convert trajectories recorded as dense encoding to one hot partial and vice versa
#30
jbloomAus
closed
1 year ago
1
Make it possible to change the grid size in the TrajectoryLSTM model to something other than 7
#29
jbloomAus
opened
1 year ago
0
(#13) Added support for env config
#28
JayBaileyCS
closed
1 year ago
1
Store model checkpoints during ppo training
#27
Felhof
closed
1 year ago
1
Minigrid: Update DictObservationSpaceWrapper to only modify the observation space "mission" entry or write another wrapper to do so
#26
jbloomAus
closed
1 year ago
1
Debug TransformerPPO model
#25
jbloomAus
opened
1 year ago
0
Reimplement BabyAI Recurrent AC Model for Trajectory Generation
#24
jbloomAus
closed
1 year ago
3
Parallelize evaluate_dt_agent (Closes #16)
#23
Felhof
closed
1 year ago
2
Store Model Checkpoints during PPO Training
#22
jbloomAus
closed
1 year ago
0
Implement another architecture such as GRU or LSTM for the PPO agent to generate trajectories for DT training
#21
jbloomAus
opened
1 year ago
0
Vectorize get trajectory minibatches method of memory class (useful for TrajPPO model)
#19
jbloomAus
opened
1 year ago
0
Implement Wandb sweeps for decision transformer training.
#18
jbloomAus
closed
1 year ago
1
Enhancement: Beautify the Streamlit App
#17
jbloomAus
opened
1 year ago
0
Rewrite evaluate_dt_agent to be parallelized (make use of gym sync vector env)
#16
jbloomAus
closed
1 year ago
0
Find a third party service for hosting the streamlit app and making it way faster to use
#15
jbloomAus
opened
1 year ago
0
Implement IBAC/SNI and measure the effect on model interpretability
#14
jbloomAus
opened
1 year ago
0
Improvement: Rewrite make_env to take an environment config file
#13
jbloomAus
closed
1 year ago
1
Bug: Trajectory Dataset contains pre-emptively truncated trajectories from where PPO get's cut off
#12
jbloomAus
opened
1 year ago
0
Write much better tutorials for the repo.
#11
jbloomAus
opened
1 year ago
0
Set up black formatter pre-commit for the repo.
#10
jbloomAus
closed
1 year ago
5
MiniGrid: Finish the Maze Environments for MiniGrid
#9
jbloomAus
opened
1 year ago
0
Previous
Next