UniLoss

Code in Pytorch for the paper:

A Unified Framework of Surrogate Loss by Refactorization and Interpolation
Lanlan Liu, Mingzhe Wang, Jia Deng
ECCV 2020

Downloading Data

For binary classification and classification tasks, the corresponding MNIST and CIFAR-10/100 datasets are downloaded automatically.

MPII for Pose Estimation

1. Download the images from MPII Human Pose Dataset

2. Create a symbolic link to the images directory of the MPII dataset:

   ln -s PATH_TO_MPII_IMAGES_DIR pose/data/mpii/images
   

Training and Evaluation

Binary Classification

To download the MNIST dataset and train the binary classification task with UniLoss, run

python train_mnist.py --batch-size 16 

Multiclass Classification

To download the CIFAR-10 dataset and train the multi-class classification task with UniLoss, run

python train_cifar10.py --batch-size 128

To download the CIFAR-100 dataset and train the multi-class classification task with UniLoss, run

python train_cifar100.py --batch-size 128

Pose Estimation

After downloading the MPII images, to train the pose estimation task with UniLoss, run

python example/train_mpii.py -a hg --lr 2.5e-4 --schedule 30 40 50