hellotem/classmix

Code used for the results in the paper "ClassMix: Segmentation-Based Data Augmentation for Semi-Supervised Learning"

Getting started

Prerequisites

• CUDA/CUDNN
• Python3
• Packages found in requirements.txt

Datasets

Cityscapes

Download the dataset from the Cityscapes dataset server(Link). Download the files named 'gtFine_trainvaltest.zip', 'leftImg8bit_trainvaltest.zip' and extract in ../data/CityScapes/

Pascal VOC 2012

Download the dataset from here. Download the file 'training/validation data' under 'Development kit' and extract in ../data/VOC2012/. For training, you will also need to download additional labels from this link, extract this directory into ../data/VOC2012.

Input arguments

Arguments related to running the script are specified from terminal and include; number of gpus to use (if >1 torch.nn.DataParalell is used), path to configuration file (see below), path to .pth file if resuming training, name of the experiment, and whether to save images during training. More details can be found in the relevant scripts.

Arguments related to the algoritms are specified in the configuration files. These include model, data, hyperparameters related to the training, and what methods to apply on unlabeled data. A full description is provided further below.

Examples

Training a model with semi-supervised learning with example config on a single gpu

python3 trainSSL.py --config ./configs/configCityscapes.json --name name_of_training

Resuming training of a model with semi-supervised learning

python3 trainSSL.py --resume path/to/checkpoint.pth --name name_of_training

Evaluating a trained model

python3 evaluateSSL.py --model-path path/to/checkpoint.pth

Pretrained model

Here is a model trained with SSL with 1/8 (372) labeled samples for Cityscapes.