This repo ports YudeWang's excellent repo for deeplabv3plus in Pytorch to Pytorch-Lightning. We reproduce performance near the reported performance in the original paper on Pascal VOC 2012 and Cityscapes.
The major changes are:
- Remove the usage of SyncBatchNorm, pytorch lightning handles this now.
- Adding logging with the WandB logger.
- Added support to run on a single GPU.
Run setup.sh to build a virtual environment and install the required packages. By default, this will also download the Pascal Dataset. If you'd like to download cityscapes uncomment the appropriate lines in setup.sh.
Make sure to set the wandb logger details in train_deeplabv3plus_pascal.py or comment it out if you are not using WandB.
Then execute the run_pascal.sh script with the desired number of gpus, batch size and epochs.
Our Cityscapes result is for the Resnet-101 backbone with atrous convolutions, unfortunately in the paper we only have the result for the Xception backbone.
| Dataset | Backbone | val mIoU | val mIoU (paper) |
|---|---|---|---|
| Pascal VOC 2012 | ResNet-101 | 78.57 | 78.85 |
| Cityscapes | ResNet-101 | 76.8 | 79.55* |
The WandB training log csv for Pascal VOC 2012 is here.
)