This repo is the official implementation of our ICLR2022 paper "NASVIT". It currently includes the training/ eval code and a pretrained supernet checkpoint on ImageNet.
python -m torch.distributed.launch --nproc_per_node=1 --master_port=1024 main.py --cfg configs/cfg.yaml --amp-opt-level O0 --accumulation-steps 1 --batch-size 64
Evolutionary search is done on a subsampled data set. Specifically, we randomly select five images for each category from the original ImageNet training set and treat them as our validation set.
ImageNet Accuracy (val)
| Model | Accuracy top-1 | Accuracy top-5 |
|---|---|---|
| Smallest | 78.34 | 93.46 |
| Largest | 82.79 | 96.00 |
The majority of NASViT is licensed under CC-BY-NC, however portions of the project are available under separate license terms: pytorch-image-models (Timm) is licensed under the Apache 2.0 license; Swin-Transformer is licensed under the MIT license.
We actively welcome your pull requests! Please see CONTRIBUTING and CODE_OF_CONDUCT for more info.