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

Added visualisation improvements for post #99

Closed JayBaileyCS closed 11 months ago

JayBaileyCS commented 11 months ago

Did these as requested:

I've made a way to sample trajectories that's useful for simple experiments and diagrams (see "get_all_memory_env_scenarios"). I was wondering if you could make a version of this function that adds the left/right action + forward to show the "correct" trajectories associated with an optimal model. Currently I have a gif showing models going 4 steps forward but not the last 2 steps.

Heatmap/Bar Chart: Reorder so it goes original, then 05, 42, 0542 then 26. Round to 2 decimal places in text for bars

JayBaileyCS commented 11 months ago

Note, less confident about this than most changes since notebooks are a bit messier. But also, since this is a dev notebook, I am not holding myself to the same production standards, we can easily fix something if we have to without anyone else relying on it.