Neural network weights prune in a static LoRA–like way -- via SVD.

⚠ Not correct, need more investigation

The test dataset is 1000 images picked from the validset of ImageNet-1k.

• download it from https://pan.quark.cn/s/cf8688f5bca1
• unzip under /data folder; the password is ]@vD4S%ojjH6UC#I!b37, just to avoid content censorship :(

data
└── imagenet-1k
    ├── image_name_to_class_id_and_name.json
    └── val
        └── ILSVRC2012_val_*.png

• torch.svd https://pytorch.org/docs/stable/generated/torch.svd.html
• LoRA: https://arxiv.org/abs/2106.09685

by Armit 2023/09/04