Beyond Categorical Label Representations for Image Classification

Boyuan Chen, Yu Li, Sunand Raghupathi, Hod Lipson
Columbia University
International Conference on Learning Representations (ICLR 2021)

Project Website | Video | Paper | Arxiv

Overview

This repo contains the PyTorch implementation for paper "Beyond Categorical Label Representations for Image Classification".

Content

• Installation
• Data Preparation
• About the labels
• Training
• Attacks
• Training with Less Data
• Shared Utilities

Installation

Create a python virtual environment and install the dependencies.

virtualenv -p /usr/bin/python3.6 env3.6
source env3.6/bin/activate
pip install -r requirements.txt

Data Preparation

Download the CIFAR10 and CIFAR100 datasets by running:

mkdir ./data
cd ./data
wget https://www.cs.toronto.edu/\~kriz/cifar-10-python.tar.gz
wget https://www.cs.toronto.edu/\~kriz/cifar-100-python.tar.gz
cd ..

About the labels

All labels are pre-generated in the labels folders and ready to be loaded directly for training. The notebook labels/labels.ipynb contains code for generating these labels.

Label types (--label):

• Category: category
• High-dimensional: speech, uniform, shuffle, composite, bert, and random
• Low-dimensional: lowdim and glove

Base model types (--model): vgg19, resnet32, and resnet110

Datasets (--dataset): cifar10 and cifar100 for cifar10 and cifar100

Seed (--seed): an int value for seeding data loading sequence

Data Level (--level): an int percentage (<90) for training with level% of all data (defaults to 100)

Base directory (--base_dir): location to save training/attacking results (required)

Label directory (--label_dir): location where label files are located (defaults to ./labels/label_files)

The labels/label_files folder contains the labels stored in .npy files.

• cifar10 high-dim labels:

• shape (10, 64, 64)

• dtype float32

• cifar100 high-dim labels:

• shape (100, 64, 64)

• dtype float32

You can find the original audio used to generate the speech labels in labels/cifar10_wav/ and labels/cifar100_wav/. You can view the grayscale images of all composite labels (rescaled to 0-255) in labels/composite/.

Training

Now to train the models use the following.

• Category: use train.py and specify label --label category

python3 train.py --model resnet110 --dataset cifar10 --seed 7 --label category

• High-dimensional: use train.py and specify a particular high-dimensional label

python3 train.py --model vgg19 --dataset cifar100 --seed 77 --label speech

Attacks

Run both targeted and untargeted FGSM and iterative attacks against trained models.

• Category: use attack.py

python3 attack.py --model resnet110 --dataset cifar10 --seed 7 --label category

• High-dimensional: use attack.py and specify a particular high-dimensional label

python3 attack.py --label speech --model vgg19 --dataset cifar100 --seed 77

Training with Less Data

Use the same training files as before, but specify a data level.

• Category: use train.py to train with 2% data

python3 train.py --model resnet110 --dataset cifar10 --seed 7 --label category --level 2

• High-dimensional: use train.py to train with 8% data

python3 train.py --model vgg19 --dataset cifar100 --seed 77 --label speech --level 8

Shared Utilities

• architecture.py: where the category and high-dimensional models are defined
• cifar.py: for loading cifar10 and cifar100 dataset for all label types
• utils/trainutil.py: other training helpers

BibTex

@inproceedings{chen2021beyond,
  title={Beyond Categorical Label Representations for Image Classification},
  author={Chen, Boyuan and Li, Yu and Raghupathi, Sunand and Lipson, Hod},
  booktitle={The International Conference on Learning Representations},
  year={2021}
}

License

This repository is released under the MIT license. See LICENSE for additional details.