jbloomAus/DecisionTransformerInterpretability

Issues

• Over resource limits on Streamlit Cloud

  #110 opened 2 months ago by subratpp
  0

• Over resource limits on Streamlit Cloud

  #109 opened 5 months ago by eggsyntax
  1

• Over resource limits on Streamlit Cloud

  #108 opened 7 months ago by mycpuorg
  0

• Over resource limits on Streamlit Cloud

  #107 opened 9 months ago by hamzaali98
  0

• Cuda cannot be disabled

  #106 opened a year ago by jackmiller2003
  3

• Folding Layer Norm in Model Loading

  #71 opened 2 years ago by jbloomAus
  9

• "Algebraic value editing" raises exception

  #82 opened a year ago by alexander-turner
  1

• Complete Embedding visualizations

  #78 opened 2 years ago by jbloomAus
  0

• Shapley Values on Attention Heads or Causal Edges Via Ablation

  #79 opened 2 years ago by jbloomAus
  0

• Complete QK/OV Circuit visualizations

  #81 opened a year ago by jbloomAus
  0

• Fix Ablation Tool

  #80 opened a year ago by jbloomAus
  0

• Write Up Analysis of Memory Env Solution.

  #40 opened 2 years ago by jbloomAus
  2

• Write a post before EAG London

  #74 opened 2 years ago by jbloomAus
  2

• Reverse Logit Lense

  #77 opened 2 years ago by jbloomAus
  0

• Mega Card: Improve Analysis App in various ways to facilitate better interpretability analysis of the new models

  #44 opened 2 years ago by jbloomAus
  3

• Look into why MemoryDT appears to have no bias on the value terms.

  #76 opened 2 years ago by jbloomAus
  1

• Expand analytical AVEC

  #75 opened 2 years ago by jbloomAus
  0

• Implement AVEC in the interpretability app

  #72 opened 2 years ago by jbloomAus
  2

• Streamlit app requires mujoco installation

  #73 opened 2 years ago by DalasNoin
  0

• Make it possible to track the preferences of the PPO in the app.

  #70 opened 2 years ago by jbloomAus
  1

• SVD Decomp / Explore ways to use dimensionality reduction to quickly understand what heads are doing.

  #69 opened 2 years ago by jbloomAus
  1

• Improve history panel in streamlit app

  #68 opened 2 years ago by jbloomAus
  1

• Facelift of the RTG Scan in the streamlit app

  #67 opened 2 years ago by jbloomAus
  1

• Ensure that analysis app is able to fold analysis models that use LNPre

  #55 opened 2 years ago by jbloomAus
  2

• Train a BC on PCT traj = 1 with two different agents mixed in and see if we can tell which one it thinks it is.

  #66 opened 2 years ago by jbloomAus
  0

• Add static interpretability visualizations to wandb dashboard.

  #60 opened 2 years ago by jbloomAus
  0

• Explore Improvements to DT Training Procedure

  #53 opened 2 years ago by jbloomAus
  6

• Do an experiment where you turn off the weighted random sampler and/or visualize sampling prob distribution.

  #56 opened 2 years ago by jbloomAus
  1

• Write a check to look at layer weight norms at initialization on the architecture, maybe visualize in a bar chart.

  #63 opened 2 years ago by jbloomAus
  0

• Verify Initialization of Transformer Model Components is good/appropriate.

  #65 opened 2 years ago by jbloomAus
  6

• Better encode/embed MiniGrid State to speed up training in DT's.

  #61 opened 2 years ago by jbloomAus
  7

• Add model export option using ONNX to facilitate better Netron visualization

  #64 opened 2 years ago by jbloomAus
  0

• Check how LSTM model BOW init is being done and whether it needs a fix

  #62 opened 2 years ago by jbloomAus
  0

• Investigate/Possibly Add Gated MLP Units to Transformer Models.

  #59 opened 2 years ago by jbloomAus
  3

• Investigate the effect of Dropout / Stochastic Depth on Model training/interpretability

  #58 opened 2 years ago by jbloomAus
  0

• Add lr scheduling options to the DT training code

  #57 opened 2 years ago by jbloomAus
  1

• Train a model using layer norm pre to see if this helps formation of calibrated, performant memory env agents.

  #52 opened 2 years ago by jbloomAus
  1

• Write a utility for merging sampled rollouts into a single file

  #36 opened 2 years ago by jbloomAus
  2

• Investigate the effects of training on data sampled using different strategies created in #46

  #47 opened 2 years ago by jbloomAus
  1

• Write a Rollout Sampling Utility for PPO Agents and add features affect generated distribution.

  #46 opened 2 years ago by jbloomAus
  1

• Upgrade Collect Demonstrations Workflow

  #51 opened 2 years ago by jbloomAus
  0

• Create an Object-Vector Calculator in the Streamlit App

  #50 opened 2 years ago by jbloomAus
  0

• 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 opened 2 years ago by jbloomAus
  1

• Set padded RTG in training data to be true RTG until masking is implemented correctly.

  #43 opened 2 years ago by jbloomAus
  1

• Update the app to also work with BC models

  #42 opened 2 years ago by jbloomAus
  0

• Add isort to the pre-commit config yaml without causing an issue with black

  #41 opened 2 years ago by jbloomAus
  0

• Add checkpoints during Offline Training

  #39 opened 2 years ago by jbloomAus
  0

• Error while sampling from new trajectories generated by LSTM model

  #34 opened 2 years ago by jbloomAus
  1

• Major: Convert the codebase into an installable python package.

  #37 opened 2 years ago by jbloomAus
  0

• Update ppo checkpoints code to upload each checkpoint in real time rather than all at the end of the workflow

  #35 opened 2 years ago by jbloomAus
  0