Skip to content
This repository has been archived by the owner on Mar 11, 2021. It is now read-only.

Minigo training using Coral Dev Board #998

Open
stephenlamis opened this issue Oct 16, 2020 · 1 comment
Open

Minigo training using Coral Dev Board #998

stephenlamis opened this issue Oct 16, 2020 · 1 comment

Comments

@stephenlamis
Copy link

Hi, I would like to try using Coral Dev Board and do a Minigo training from scratch.
Is it possible?

I have a working Coral Dev Board, and able to start the game by following this steps https://coral.ai/projects/minigo/#step-5-start-a-game

But afterwards, I have no glue on how to do bootstrap and start the selfplay run inside the board.
Anyone try using the board and run the complete reinforcement loop?

P.S. do I have to install tensorflow? (or need to install tensorflow lite?)

@tommadams
Copy link
Contributor

Edge TPUs such as the Coral Dev Board are inference-only by design, you cannot use them to train a model.

Training a Minigo model from scratch is non-trivial and requires a significant amount of computing resources. The process is complicated further if you want to run on an Edge TPU rather than a desktop GPU. The model used in the coral.ai Minigo project was retrained from existing data in the (minigo-pub)[https://console.cloud.google.com/storage/browser/minigo-pub/] GCS bucket.

Before attempting this, I suggest familiarizing yourself with how to train a simpler model by following a tutorial link this:
https://coral.ai/docs/edgetpu/retrain-detection/

Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants