Semantic segmentation with dynamic upsamplers, based on mmsegmentation

For example, to train UPerNet-R50 with CARAFE in FPN:

bash dist_train.sh configs/dynamic_upsampling/upernet_r50_4xb4_carafe-80k_ade20k-512x512.py 4

We find that the performance on ADE20K is unstable and may fluctuate about (-0.5, +0.5) mIoU.

The code of upsampler application on SegFormer (Semantic Segmentation) and DepthFormer (Monocular Depth Estimation) can be found here.