sethtroisi / cloudygo Goto Github PK

Could you add a top 10 models rating table on the minigo page?

e.g.

https://cloudygo.com/v16-19x19/joseki/full/1548398561-tpu-player-deployment-d74dbc759-27cfr-7.sgf

https://www.reddit.com/r/cbaduk/comments/8n4j7f/cloudygo/dzt39mu/

Setting up CloudyGo with all of MiniGo V7 takes ~20 hours of just copying / import data.

It would be nice to have an zipped instance folder. I imagine it would have 2 training runs with 100 models each and maybe 100 game per model + some eval games. This would make testing some changes much simpler for newcomers

https://cloudygo.com/v16-19x19/eval-graphs updated to m1156, but
https://cloudygo.com/ updated to m1145

I'm writing SPRT calculator and I'm confused by the sprt math.

I wrote up the three methods I know in this gist
LLF = my optimized version
wiki = my take on likelihood-ratio_test
LLR = the code from leela-zero-server

Other links:
https://www.chessprogramming.org/Match_Statistics#SPRT
^ uses similar math to this repo but with draws
^ draws it's justification from http://hardy.uhasselt.be/Toga/GSPRT_approximation.pdf

https://nowak.ece.wisc.edu/ece830/ece830_fall11_lecture9.pdf
^ maybe I'm misunderstanding p0(x) and p1(x)

plz can haz?

This is work against https://github.com/tensorflow/minigo/blob/master/oneoffs/training_curve.py

The pickle file is used in the Colab Figure 3 Investigation

http://cloudygo.com/v7-19x19/model_comparison/pv/100/195?group=policy
and
http://cloudygo.com/v7-19x19/model_comparison/pv/195/100?group=policy

show the same thing

All eval graphs
Game length

Context: tensorflow/minigo#722 (comment)

This issue tracks adding an endpoint for the cross-run-eval bucket that makes accessing
https://cloudygo.com/all-eval-graphs?bucket=cross-run-eval possible through the UI.

I started work on a collaborative Go playing page like Kasparov versus the World but for lack of time I've never got it to any sate where it could be presented. I should remove it all or fix it.

A couple of people (AMJ, KillerDucky, TheOperation) know about the site but I hope to make a splash with that high quality first post :)

I'd love to make an FAQ / links page that maps the CloudyGo SiteNav to the related template and code blocks

is such a thing possible? Making a combined fig3, and making the per-run fig3 only the zoom in at the tail?

• Drop 19x19 in model names.
• Add help icon / readme in Header bar
• Add "Zero to MG in five minutes" guide (colab for converting weights?)
• links to weigths from cloudy (https://cloudygo.com/converted_model/v16-19x19/models/)

i think it involves supporting a custom GTP command; might involve being in minigo

https://cloudygo.com/v14-19x19/puzzles/ is failing.
https://cloudygo.com/v14-19x19/puzzles/1 and https://cloudygo.com/v14-19x19/puzzles/2 work
http://localhost:5002/v14-19x19/puzzles/ also works.

I am working to create file format signatures for use in file identifications tools such as DROID and Siegfried. Can you please let me know what format the files on https://pminus1.cloudygo.com/ are? For example, https://pminus1.cloudygo.com/ download/m6918649

They have the initial four bytes of "K9z1", which is a decent enough magic to identify the files, but I like to have some additional information such as file format specifications, etc.

Thanks.

cd rl_loop
python
import shipname
print('\n'.join(["google-chrome https://www.google.com/search?q=british+ship+%s &" % n for n in shipname.NAMES]))

copy and paste the first set to a new terminal
(This should open a google search tab for each one

function photo { echo "$2" > $1.meta; curl "$3" > $1.jpg; }

mkdir -p photos; cd photos
photo <NAME> <wikipedia image page> <wikipedia image url>

i.e.

photo warspite "https://en.wikipedia.org/wiki/HMS_Warspite_(1666)#/media/File:HMS_Edinburgh-IMG_7758.jpg" "https://upload.wikimedia.org/wikipedia/commons/8/89/HMS_Edinburgh-IMG_7758.jpg"

Many of the name represent multiple ships (WWI, WWII, older ships, ...) feel free to choose any one of them (I like models from museums but choose as you please.

I'm working alphabetically backwards. If someone (cough Andrew cough) wanted to work forwards that would be nice.

vs

Asdf

I'm not sure exactly what I'll add but I should consider it

Expected behavior:

When checking both 'filter bad' and 'filter old', the moving-average at the leftmost-edge should correspond to the moving average at that point when 'filter old' is unchecked.

Actual behavior
moving average of zoomed in window begins at artificially inflated number

Result: Upward trend obfuscated.

easier shown in pictures :)

I'd love for the site to be a little less white and a little more modern suggestions (and pulls) very welcome along this line

https://cloudygo.com/v16-19x19/eval-graphs
it shows 114380 games for a long time, and doesnt show new weights such as 1050

Using the new joseki page here https://cloudygo.com/v15-19x19/openings.html, I see all the openings for v15 are on a decline at the end of training. Is it somehow calculated differently or is there a dominating joseki that isn't showing up? Love the new tool btw.

leela-zero/leela-zero#2212

Reported to me: Person finds an interesting go game, but can't see a way to download the SGF from the page.
Idea: a download link on the sgf-player page.

https://cloudygo.com/all-eval-graphs?bucket=cross-run-eval
Is there any way to show the cross-over graphic without only using color? For example, you can choose which line you want to show. For color blindness like me, I can hardly tell which is the corresponding one.

I did some prework

https://docs.google.com/spreadsheets/d/19fNtHb--WdstW5Kf5Dkw4ylBBirPfp2ddY3UiE0C3tA/edit?ouid=103721162867446834414&usp=sheets_home&ths=true

and with this branch

https://github.com/sethtroisi/leela-zero/tree/zz__pro_analysis/scripts/pro_game_analysis

but more work needs to be done

I have an ~ELO number hidden under each model's eval page it would be nice to use this more familiar number on the evaluation page

guessing which games i looked at is sad :(

https://github.com/tensorflow/minigo/blob/master/oneoffs/training_curve.py#L138

Then do the work to display better graphs on my side

Some initial LZ data in imported in http://cloudygo.com/leela-zero-v1/eval-graphs but I should try to import all training games, and their evaluation games

I have some hardcoded assumption about game participants based which will need to be changed.

Also unluckiness will probably not work (and debug games will not exist)

tensorflow/minigo#695 (comment)