pytorch上实现语义分割网络bisenet
语义分割方向新近提出来的网络大概是deeplabv3+和bisenet，在18年2月和8月先后被提出。
无条件相信google，于是直觉上认为deeplabv3+更靠谱。?
看paper的话，bisenet准确率更低，速度更快。
deeplabv3+之前已经实现，现在来对bisenet进行实现。
我的环境：
anaconda3
pytorch-gpu 1.0
文章目录
- 源
- 模型
- 数据集
- 训练
- 验证
源
分别找到tensorflow和pytorch的版本：
tf: https://github.com/GeorgeSeif/Semantic-Segmentation-Suite
pt: https://github.com/CoinCheung/BiSeNet
But George said “This repo has been depricated and will no longer be handling issues. Feel free to use as is ?” so我这边选择pt版本。
模型
下载模型文件：
git clone https://github.com/CoinCheung/BiSeNet.git
数据集
实现过程使用Cityscapes Dataset作为数据集。
从www.cityscapes-dataset.com上下载Cityscapes Dataset——数据集leftImg8bit_trainvaltest.zip (11GB)和对应的标注集gtFine_trainvaltest.zip (241MB)。下载完成后解压到路径BiSeNet/data/，目录结构：
BiSeNet/data/
- gtFine
- leftImg8bit
训练
在路径BiSeNet/下：
CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.launch --nproc_per_node=2 train.py
会出现：
it: 79500/80000, lr: 0.000105, loss: 2.3915, eta: 0:06:58, time: 41.0036
it: 79550/80000, lr: 0.000096, loss: 2.4082, eta: 0:06:16, time: 40.9691
it: 79600/80000, lr: 0.000086, loss: 2.3882, eta: 0:05:34, time: 40.9198
it: 79650/80000, lr: 0.000076, loss: 2.4236, eta: 0:04:53, time: 43.7076
it: 79700/80000, lr: 0.000067, loss: 2.3925, eta: 0:04:11, time: 40.7097
it: 79750/80000, lr: 0.000056, loss: 2.4639, eta: 0:03:29, time: 40.9966
it: 79800/80000, lr: 0.000046, loss: 2.4212, eta: 0:02:47, time: 43.6417
it: 79850/80000, lr: 0.000036, loss: 2.4144, eta: 0:02:06, time: 40.9328
it: 79900/80000, lr: 0.000025, loss: 2.4603, eta: 0:01:24, time: 40.9232
it: 79950/80000, lr: 0.000013, loss: 2.3965, eta: 0:00:42, time: 40.9462
it: 80000/80000, lr: 0.000000, loss: 2.4537, eta: 0:00:00, time: 43.6481
training done, model saved to: ./res/model_final.pth
================================================================================
evaluating the model ...
setup and restore model
0%| | 0/100 [00:00<?, ?it/s]
compute the mIOU
100%|█████████████████████████████████████████| 100/100 [18:32<00:00, 10.92s/it]
mIOU is: 0.786584
验证
在路径BiSeNet/下：
python evaluate.py
会出现：
================================================================================
evaluating the model ...
setup and restore model
compute the mIOU
100%|█████████████████████████████████████████| 100/100 [17:14<00:00, 10.30s/it]
mIOU is: 0.784525