uchidalab / vision-conformer

Vision Conformer: Incorporating Convolutions into Vision Transformer Layers
MIT License
5 stars 1 forks source link

vision-conformer

Implemention of Vision Conformer, a vision classification model that conbines Vision Transformer and CNN.

The contents are under developing.
If you have any questions, please touch me. akihiro.kusuda@human.ait.kyushu-u.ac.jp

Usage

You can create a conda environment with envirionment.yaml.

conda env create -f envirionment.yaml
conda activate vision-conformer

You can train the model by running main.py.

python main.py

You can also check the results on mlflow dashboard.

mlflow ui

Parameters

In main.py, you can change the parameters for this model.

model = VisionConformer(
    image_size=28,
    patch_size=4,
    num_classes=10,
    dim=256,
    depth=3,
    heads=4,
    mlp_dim=256,
    dropout=0.1,
    emb_dropout=0.1,
    pool='cls',
    channels=1,
    hidden_channels=32,
    cnn_depth=2,
)