Awesome

TWA

The code is the official implementation of our ICLR paper Trainable Weight Averaging: Efficient Training by Optimizing Historical Solutions.

We propose to conduct neural network training in a tiny subspace spanned by historical solutions. Such optimization is equivalent to performing weight averaging on these solutions with trainable coefficients (TWA), in contrast with the equal averaging coefficients as in SWA. We show that a good solution can emerge early in DNN's training by properly averaging historical solutions with TWA. In this way, we are able to achieve great training efficiency (e.g. saving over 30% training epochs on CIFAR / ImageNet) by optimizing these historical solutions. We also provide an efficient and scalable framework for multi-node training. Besides, TWA is also able to improve finetune results from multiple training configurations, which we are currently focusing on. This colab provides an exploratory example we adapt from Model Soups.

Dependencies

Install required dependencies:

pip install -r requirements.txt

How to run

TWA in tail stage training

We first show that TWA could improve the performance of SWA in the original SWA setting, where the improvements are more significant when the tail learning rate swa_lr is larger.

cd swa

First, run SWA using original code:

bash run.sh

Then, we could perform TWA using:

bash run_twa.sh

The training configuration is easy to set as you need in the scripts.

TWA in head stage training

In this part, we conduct TWA in the head training stage, where we achieve considerably 30%-40% epochs saving on CIFAR-10/100 and ImageNet, with a comparable or even better performance against regular training. We show sample usages in run.sh.

For the first step, we conduct regular training for generating the historical solutions (for ImageNet training, the dataset need to be prepared at folder path). For example,

datasets=CIFAR100
model=VGG16BN
DST=results/$model\_$datasets

CUDA_VISIBLE_DEVICES=0 python -u train_sgd_cifar.py --datasets $datasets \
        --arch=$model --epochs=200 --lr 0.1 \
        --save-dir=$DST/checkpoints --log-dir=$DST -p 100

Then, we conduct TWA training for quickly composing a good solution utilizing historical solutions (note that here we only utilize the first 100 epoch checkpoints):

CUDA_VISIBLE_DEVICES=0 python -u train_twa.py --epochs 10 --datasets $datasets \
        --opt SGD --extract Schmidt --schedule step \
        --lr 2 --params_start 0 --params_end 101 --train_start -1 --wd 0.00001 \
        --batch-size 128  --arch=$model  \
        --save-dir=$DST/checkpoints  --log-dir=$DST

Citation

If you find this work helpful, please cite:

@inproceedings{
    li2023trainable,
    title={Trainable Weight Averaging: Efficient Training by Optimizing Historical Solutions},
    author={Tao Li and Zhehao Huang and Qinghua Tao and Yingwen Wu and Xiaolin Huang},
    booktitle={The Eleventh International Conference on Learning Representations},
    year={2023},
    url={https://openreview.net/forum?id=8wbnpOJY-f}
}