sample

Sample of analysis with katawrap

Jupyter notebooks

You may like to skip this document and open katawrap_sample.ipynb directly for a quick look at what it looks like. You can try it with prepared analysis results immediately.

Or, you can also upload google_colab/katawrap_sample_colab.ipynb to Google Colaboratory and run KataGo + katawrap there. (Do not forget to change runtime type to GPU.)

About this directory

• README.md: instructions (this file)
• katawrap_sample.ipynb: usege of analysis results with Python (Jupyter notebook)
• analyze_go_styles.md: another sample to analyze players' Go styles
• google_colab/: sample for katawrap on Google Colaboratory
• sgf/: sample SGF files
• sample_result.jsonl: results of analysis by KataGo with katawrap

Note

This document assumes that you can run katago on the command line as follows.

$ katago analysis -config analysis.cfg -model model.bin.gz

Change this part as appropriate for your case in the following examples.

Prepare analysis results

Run katawrap to dump the results.

$ ls sgf/*.sgf \
  | ../katawrap/katawrap.py -visits 400 \
      katago analysis -config analysis.cfg -model model.bin.gz \
  > result.jsonl

If this is too slow, try decreasing visits and/or analyzing every N turns, for example.

$ ls sgf/*.sgf \
  | ../katawrap/katawrap.py -visits 100 -every 25 \
      katago analysis -config analysis.cfg -model model.bin.gz \
  > result.jsonl

Run Jupyter notebook

$ jupyter notebook katawrap_sample.ipynb

Rewrite 'sample_result.jsonl' with 'result.jsonl' at the beginning of the notebook. Then run all!

Bonus: jq version

Find the top 5 exciting games in your collection:

$ cat result.jsonl \
  | jq -s 'map(select(0.2 < .winrate and .winrate < 0.8))
    | group_by(.sgfFile) | map(max_by(.unsettledness)) | sort_by(- .unsettledness)
    | limit(5; .[])
    | {sgfFile, turnNumber, winrate, scoreLead, unsettledness}'

Calculate the match rates with KataGo's top 3 suggestions in first 50 moves:

$ cat result.jsonl \
  | jq -sc 'map(select(0 <= .turnNumber and .turnNumber < 50 and .nextMoveColor != null))
    | group_by(.sgfFile)[] | group_by(.nextMoveColor)[]
    | [(.[0] | .sgfFile, .PB, .PW, .nextMoveColor),
       (map(.nextMoveRank) | (map(select(. < 3 and . != null))|length) / length)]'