Spred: Solving L1 with SGD

This is the repository for the paper Spred: Solving L1 with SGD.

The subtasks are contained into different folders

Environment

The conda environment can be found in enironment.yml

We note that there will be some compatibility issue to load the cancer datasets. An adhoc fix is to change the line 322 of GEOparse.py into line 323-324, as indicated in the following picture.

Task1: Linear regression

The first task is described in folder linear. Please check the linear/readme.md

Task2: Non-linear classification in high-dimensional dataset

The second task is non-linear classification in high-dimensional dataset, where the labels of genes are predicted. Please check the non-linear/readme.md

Task3: Image classification on CIFAR10 and CIFAR100

The third task is to leverage the training protocal of STR. The original repository can be found in this link. We modify the model implementation.

Please check scripts with run_{STR/spred/L1}_{cifar10/cifar100}.sh for the usage on CIFAR datasets. The experiments records can be found in this online form.

This repo can be also used to run experiments on Imagenet. The raw experiment results and commands to run the experiments is found in this online form.

(Legancy) previous implementation of CIFAR dataset.

The CIFAR results in earlier version is implemented in folder CIFAR (old).