pytorch-deeplab-resnet

Why we included this project

This PyTorch port of DeepLab-ResNet v2 reproduces the multi-scale training scheme from the original Caffe release, computing losses at 1x, 0.75x, and 0.5x scales and merging the score maps. That makes it a useful reference for seeing how the classic architecture behaves in PyTorch rather than Caffe. Training and evaluation scripts cover Pascal VOC and accept custom datasets through label-count flags, and the SGD with poly learning-rate decay follows the original authors' setup. The documented mean-IOU results and the caffemodel-to-pth converter are genuinely useful for researchers and students comparing segmentation baselines, though the code targets Python 2.7 and an older PyTorch era, so expect some friction on current setups.