Gradients as Features for Deep Representation Learning

This code repository is under construction.

Overview

This repository contains code for reproducing the results in Gradients as Features for Deep Representation Learning published as a conference paper at ICLR 2020. The code has been tested in an conda environment with Python 3 and PyTorch >= 1.3.

Quick Start

Download the base networks here. We currently support BiGAN/ALI encoder pre-trained on CIFAR-10/-100 or SVHN as the base network. In the download link, "ali" stands for ALI trained on Jenson-Shannon divergence, and "wali" stands for ALI trained on Wasserstein distance. See (and please star :) ) our repository on Wasserstein BiGAN.

• File names with a trailing zero correspond to randomly initialized networks (e.g., fnet0.pt, std_hnet0.pt, etc.).
• File names with a trailing one correspond to networks pre-trained with generative modeling (e.g., fnet1.pt, std_hnet1.pt, etc.).
• File names with the prefix "std" correspond to networks under standard parametrization.
• File names with the prefix "ntk" correspond to networks under NTK parametrization.

Update the loading and saving paths in the configuration files before you try out the sample commands.

• Activation baseline (i.e., the standard multi-class logistic regressor)

python ./src/benchmark.py -c ./configs/cifar10/ali/actv.config

• Full model (i.e., the proposed linear model)

python ./src/benchmark.py -c ./configs/cifar10/ali/linear_conv3.config

• Gradient baseline (i.e., the gradient term alone in the proposed model)

python ./src/benchmark.py -c ./configs/cifar10/ali/grad_conv3.config

• Network fine-tuning

python ./src/benchmark.py -c ./configs/cifar10/ali/finetune_conv3.config

Please note that we use a slightly different set of hyperparameters for training compared with what was originally used in the paper. In particular, we apply stochastic gradient descent (SGD) instead of ADAM as the default optimizer to repect the convention. Accordingly, we modify the learning rate scheduling since we found that it leads to faster convergence.

Contact

Fangzhou Mu ([email protected])

Bibtex

@inproceedings{mu2020gradfeat,
  title={Gradients as Features for Deep Representation Learning},
  author={Mu, Fangzhou and Liang, Yingyu and Li, Yin},
  booktitle={International Conference on Learning Representations (ICLR)},
  year={2020}
}