Skip to content

Latest commit

 

History

History
39 lines (29 loc) · 1.27 KB

README.md

File metadata and controls

39 lines (29 loc) · 1.27 KB
Raw

Towards Understanding Learning Representations: To What Extent Do Different Neural Networks Learn the Same Representation

Requirements

  • python3
  • anaconda
  • pytorch

A easy way is to run the following commands after installing conda

conda create -n pytorch python3.6 anaconda
conda install pytorch torchvision -c pytorch
source activate pytorch

Project Structure

train_cifar10.py is the main script to train or validate on cifar10.

configs/cifar10 contains configurations for training models, like vgg16.py.

You could run python train_cifar10.py -c vgg16 --rng_seed 0 to train a vgg16 model on cifar10 with random seed 0 with all available gpus.

Results would be saved at output/cifar10/{config}/{rng_seed}.

All the models are defined in models/cifar10.

pred_cifar10.py is the main script to extract features on cifar10

# select all activated features and output a compressed npz on test set
python pred_cifar10.py -m vgg16 --ckpt output/cifar10/vgg16/rnd_0/best.ckpt -t relu

scripts maintains some useful scripts for convenience.

Todo

  • Test training scripts by visualization
  • Test feature extractors
  • Test maximal matching calculation
  • Test minimum matching calculation
  • Add docstrings, copyright, license