Skip to content

Latest commit

 

History

History
40 lines (30 loc) · 1.6 KB

README.md

File metadata and controls

40 lines (30 loc) · 1.6 KB

DeiT example

preparation

The DeiT pretrained model is the checkpoint from https://github.com/facebookresearch/deit . The example will automatically download the checkpoint using torch.hub.load.

The datasets used in example are calibration dataset and validation dataset.

  • For calibration dataset, follow the statements in https://github.com/megvii-research/Sparsebit/blob/homeworks/homeworks/quant_homework.md#resource and download imagenet-1k dataset. Use ./calibration_data/calibration/ as the path to calibration dataset.

    tar -xvf imagenet-1k-images.tar -C path_to_calibration_data/calibration
ln -s path_to_calibration_data calibration_data

  • For validation dataset, download from https://image-net.org/ and move validation images to labeled subfolders, using this script. Use ./validation/ as the path to validation dataset.

    tar -xvf ILSVRC2012_img_val.tar -C path_to_validation_data
cd path_to_validation_data
wget https://raw.githubusercontent.com/soumith/imagenetloader.torch/master/valprep.sh
chmod +x valprep.sh
./valprep.sh
cd path_to_example
ln -s path_to_validation_data validation_data

The datasets are loaded with torchvision.datasets.ImageFolder. Using custom datasets for calibration and validation is OK.

Usage

python3 main.py qconfig.yaml

Use argument -b batch_size to assign batch_size if the default batch_size(=128) is too large.

Metrics

model float32 acc 8w8f acc
DeiT-tiny 72.026 70.778
DeiT-base 81.742 81.152