Closed sashavor closed 3 years ago
@tianyu-z can you share your results / findings / insights and general updates here?
It helps but needs a little bit of modification (what I called V2.1).
Conclusion after eyeballing:
I think for simple cases, the improvement is significant but it does not perform well on some of the Streetview (complex) cases. P.S. AdventV2 dropped Sim data for the second stage while AdventV2.1 kept a part of / all of the Sim data.
Example for 1: Example for 2 (left is V2.1 and right is V1): Example for 4:
More experiments are on the way. I will post them here when they are ready.
To my surprise, I tried entropy_split = 0.30, 035, 0.40, 0.45, 0.50, 0.55, 0.60, 0.67, 0.75, 0.80, 0.85, 0.90, however the 0.67 one which is proposed by the original adventv2 paper has the best performance.
The train and eval links of these exps are here: https://docs.google.com/spreadsheets/d/1oaXsUeICfhmsbikVOUq0jxLr6th_6j0KpOL2UmmAqpc/edit#gid=0
Create a new spreadsheet and edit with others at the same time -- from your computer, phone or tablet. Get stuff done with or without an internet connection. Use Sheets to edit Excel files. Free from Google.
See if it helps improve results