A mirror repository of MXNetSeg based on the HUAWEI MindSpore framework.
In this project, we adopt:
- Python 3.7.5
- CUDA 10.1
- cuDNN 7.6.5
- Mindspore-gpu 1.5.0
Refer to MindSpore Installation Guide for details.
- Convert the dataset to .mindrecord files
python build_seg_data.py --data-name Cityscapes --dst-path tmp_data/train.mindrecord --num-shard 4 --shuffle- Run the
train.pyscript
# for multiple GPUs
mpirun -n 4 python train.py --train-dir experiment --data-file tmp_data/train.mindrecord0 --batch-size 8 --crop-size 1024 --ignore-label 19 --num-classes 19 --epochs 240 --lr-type poly --base-lr 0.04 --wd 4.e-4 --model eprnet --device-target GPU --distributed- Run the
eval.pyfor performance evaluation
python eval.py --data-name Cityscapes --batch-size 16 --crop-size 1024 --scales 0.5 0.75 1.0 1.25 1.75 2.0 --flip --ignore-label 19 --num-classes 19 --eval-split val --model eprnet --checkpoint experiment/ckpt/eprnet-10_27.ckpt- This repository uses
BatchNorm2das theSyncBatchNormoperator in MindSpore does not support NVIDIA GPU for now, which may lead to poor segmentation accuracy when training on multiple devices. - When evaluating model performance, the
ignore-labelindicator is temporarily not supported, which will be improved in a future version.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent