Vision Transformer in JAX/Flax
This repository implements Vision Trasnformer(ViT) in Flax, introduced in an ICLR paper 2021 submission, with further explanation by Yannic Kilcher. This repository is heavily inspired from lucidrain's implementation.
Install
pip install vit-flax
Usage
import jax
from jax import numpy as jnp
from flax import nn
from vit_flax import ViT
rng = jax.random.PRNGKey(0)
module = ViT.partial(patch_size=32, dim=1024, depth=6, num_heads=8, dense_dims=(2048, 2048), img_size=256, num_classes=10)
_, initial_params = module.init_by_shape(
rng, [((1, 256, 256, 3), jnp.float32)]
)
model = nn.Model(module, initial_params)
img = jax.random.uniform(rng, (1,256,256,3))
output = model(img)examples directory contains code to train ViT on CIFAR datasets.
Docs and references
Documentation for all the modules can be viewed here.
Note
This repository is still in initial stages. Feel free to Contact me or raise issues/PR for suggestions, improvements or bugs.
Help needed
A recent commit introduces code for training CIFAR models in the examples directory. If you're using this code and have the resources to run, I'd be happy to include those reports here and give appropriate credits for the same.