This is my contribution to the "PGU-net+: Progressive Growing of U-net+ for Automated Cervical Nuclei Segmentation" in MICCAI2019 workshop MMMI.
Paper link: http://arxiv.org/abs/1911.01062 Please cite this paper if you find this project helpful for your research.
python 3.7, CUDA 10.1, numpy 1.17, matplotlib 3.1.1, scikit-image (0.21), scipy (0.3.1), pyparsing (2.4.2), pytorch (1.0) (anaconda is recommended) other packages could be the latest version.
1.Install all dependencies
2.For PGU-net+: Progressive Growing of U-net+, python train_PG.py
For the U-net structure in the final stag, python train.py
For testing, python test.py
In the later stage, several parallel and distributed training methods of multi GPU are added，参考自https://github.com/tczhangzhi/pytorch-distributed
1.nn.DataParallel 简单方便的 nn.DataParallel
2.torch.distributed 使用 torch.distributed 加速并行训练
3.torch.multiprocessing 使用 torch.multiprocessing 取代启动器
4.apex 使用 apex 再加速
5.horovod horovod 的优雅实现
For the U-net structure with 5 distributed training methods above
python train_1024_DataParallel.py for nn.DataParallel
python train_1024_distributed.py for torch.distributed
python train_1024_multiprocessing.py for torch.multiprocessing
python train_1024_apex.py for apex
python train_1024_horovod.py for horovod
实验结果如下图（apex和horovod存在一些bug）：
