DeepGO (DGM) 🤖

Author . Abdelkader Chihab Benamara

Several two headed neural networks with ~ less 1M parameters that performs pretty good on Go game , trained on 1M games dataset from the KataGo dataset. I have used the Grid5000 clusters and trained my models using 8x Nvidia Tesla V100 (32 GiB) GPU's and 512GB of RAM. And my current best model has 57.75% (top 1 during the weekly tournament organized by Mr.Tristan Cazenave ) validation accuracy for the policy head.

Getting Started 🛠

Create the docker container ( Python(3.9) + TF(2.7.0) )

Lunch: docker build -t deep-go -f Dockerfile .

see requirements.txt for more details

Now you need to run this container :

Run : docker run -it deep-go /bin/bash

Note. The MacOS version isn't working currently there will be updates soon !

Or more simply you can use the following deployment file through :

sh ./path_to_deploy.sh

Note. If you are using an AMD GPU you can switch the deploy file and use the second one.

Train your own models 🏋️‍♂️

Actually i managed to make it easier for any one using this project to build and train his own model , first just you need to specify your architecture by creating a new folder inside ./src/models and DGVx for example and then you can specify your model body (input and outputs blocks are already fine 😁) . Then you can change the self.name for your version so you could have some pretty logs , and then you are all ready to go. Use the following command to run and train your powerful model !

   python?3 ./path_to_train.py -gpu 8 -s 1 -e 2000 -b 1024

This command train the model DGVx on 8 GPU's for 2000 epochs and with a batch size 1024. Here are the details for options that you can use for training :

---------------------------------------------------------------
	Command Line Guide (args to use)
---------------------------------------------------------------
-h [–-help]	          :  help with command line
-gpu [–-num-gpu]      :  \# of GPUs to be used 
						 for training [default : 2]

-s [--start-epoch]    : the epoch to porsuit the training 
					    from [default : 1]

-e [--end-epoch]      : the last epoch of training [default : 500]

-b [--batch-size]     : the batch size used to
                        train model [default : 512]

-n [--num-samples]    : \# of samples in the 
                        get_batch per epoch [default : 25k]

-w [--show-warnings]  : the warnings show level [from 0 to 3]

-l [--load][PATH]     : load Model to continue 
                        training from [PATH]

Note . Logs are made with the beautiful library rich in order to get more readable results.

Why not some charts ? 📊

Yes ! you can do it u simply need to recover the history of your models which is done automatically through the earlier command for training then you can simply call the ./src/utils function :

def train_plots(epochs,histories,styles)
-- styles can be of the format (color-style)
   Ex. mo means magenta and dots ...

Further Details

For more details and examples on how to use this charts see ./src/train_dgm.ipynb And to see my results during the training and testing of my models you can take a look on my paper [DeepGo] via this link.

Contibute

If you encountered a problem or want to contibute , don't hesitate to open an issue and/or reach out on LinkedIn : Chihab Benamara

TODO --

implement chess version (stockfish's nnue)

ChihabEddine98/DeepGo