This repository contains image parallel training and testing codes for DeepLab ResNet 101.
You can train on train_aug and evaluate on val of PASCAL VOC2012 on multiple GPUs by
python deeplab_main.py 0,1,2,3 train
python deeplab_main.py 0,1,2,3 eval
During evaluation you will also need to run EvalSegResults.m in the matlab folder to compute the mean IOU.
A large part of the code is borrowed from https://github.com/chenxi116/pytorch-deeplab. (see commit history)
Current parallel code doesn't work perfectly for very small images, but it can be easily modified to use only one GPU for small images, and still use multiple GPUs for large ones.