Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

komi + tensorbatch + squeeze-excitation + ladder/legality/liberty #91

Open
wants to merge 9 commits into
base: komi+tensorbatch-se
Choose a base branch
from

Conversation

alreadydone
Copy link
Owner

No description provided.

Ttl and others added 9 commits August 5, 2018 20:17
Legal plane

Liberty planes

ladders start

Working ladder capture and escape

Ladder planes to network

Fix handling of the last bit

Training support

Fix plane order

Add parametric relu

Training changes

Training changes for colab

Fix add_features, net_to_model

Latest colab training parameters

Separate own/opponent liberty planes, fix add_features

add_features cleaning

Python script for add_features

Adds features to all chunks in folder in parallel

Training changes

Add gammas

Colab: Copy model to drive

Gammas to stddivs

Change legal plane back to before.

SE training changes

CPU SE unit support

SE unit GPU support

Local memory in pooling

Local memory in in_transform

Rebase cleanup

Change legal moves plane to empty but illegal moves

Less correlation with move history planes

Training rebase

Rollback training changes

Multi GPU code needs still work.
Exact amount are faster than rounded.
plus some other small changes
As currently implemented, one leelaz.exe process is spawned to convert only one chunk before exiting. By adding the --cpu-only flag along with using a smaller 32x4 weights file, start-up time for each individual leelaz.exe process is vastly reduced. This saves an enormous amount of time when running "add_features" on many chunks.

Pull request #1
Add Ladder.h and Ladder.cpp to VS2017 project.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

3 participants