Hiragata_ai

The AI model that is used in hiragata.com.

This repo is part of the Hiragata project Links to other parts of the project:

Repo	Description
Hiragata	React based front-end
Hiragata_ai	Source code of the AI model
test-hiragata-ai	AI model testing and visualization react based FE
japanese-fonts-dataset	Source for a small font-based dataset

Dev guide

I want to work with the data only (no model training)

Clone the repo
Change working directory to the repo root (cd Hiragata_ai)
Download some model-ready data from Google Drive
Extract data to data/bin/ (Final result should be something like data/bin/48x48/binary_files_located_here)
Install npm packages npm ci (It's fine if tfjs build fails as you won't need it if you're only working with data. Most of tfjs will still work as it will fallback to the js version if you don't have the GPU stuff working)
Create dir for test-images mkdir test-images
Run npm run init-data to create some images into /test-images and test if everything works
Great success!

I want to work with the model

NOTE: everything will work with the js-node version of tfjs, but if you want to greatly accelerate the learning you could try to get the GPU stuff working

How to get tfjs-gpu to work

Reguirements: The hardware and software requirements are the same as regular tensorflow. You can use this quide to get deps working.

Name	Version
NVIDIA® GPU drivers	>410.x
CUDA® Toolkit	10.0
cuDNN SDK	>=7.4.1 but not 8

Install packages only after ^

If you get C build errors on windows consult tfjs doc

if you get gpu allocation err try running export TF_FORCE_GPU_ALLOW_GROWTH=true

Data

TODO: document this section