// the find
bubbliiiing/deeplabv3-plus-pytorch
这是一个deeplabv3-plus-pytorch的源码，可以用于训练自己的模型。
A PyTorch implementation of DeepLabv3+ for semantic segmentation, aimed at Chinese-speaking developers who want to train on custom VOC-format datasets. It supports MobileNetV2 and Xception backbones with configurable downsampling factors. The repo is a teaching aid first, a production tool second.
Pretrained weights are provided for both backbones with mIOU benchmarks on VOC12+SBD, so you can verify the implementation is actually correct before training. Multi-GPU training support was added in 2022. The VOC-format custom dataset pipeline is straightforward — label your images, run voc_annotation.py, set num_classes, go. The inclusion of FPS testing and video inference modes in predict.py means you can sanity-check performance on real hardware without writing extra code.
Pinned to torch==1.2.0 — a version from 2019 that predates half of what modern PyTorch offers and will not install cleanly on any CUDA toolkit newer than 10.1. The repo has been dormant since October 2023 with no indication it will be updated. All documentation is in Chinese with no English translation, which limits its audience considerably. Weights are hosted on Baidu Netdisk, which requires a Baidu account and is effectively inaccessible outside China without a workaround.