FastSCNN 训练
the real-time image segmentation FastSCNN
Fast segmentation convolutional neural network (Fast-SCNN), an above real-time semantic segmentation model on high resolution image data (1024x2048px) suits to efficient computation on embedded devices with low memory. Building on existing two-branch methods for fast segmentation, the 'learning to downsample' module computes low-level features for multiple resolution branches simultaneously. FastSCNN combines spatial detail at high resolution with deep features extracted at lower resolution, yielding an accuracy of 68.0% mean intersection over union at 123.5 frames per second on Cityscapes.
For more detail：https://arxiv.org/abs/1902.04502
Requirements
use pytorch, you can use pip or conda to install the requirements
# for pip
cd $project
pip3.7 install -r requirements.txt
Note: pillow recommends installing a newer version. If the corresponding torchvision version cannot be installed directly, you can use the source code to install the corresponding version. The source code reference link: https://github.com/pytorch/vision，
Suggestion the pillow is 9.1.0 and the torchvision is 0.6.0
CANN 20210617_5.0.T205
torchvision
数据集准备
1.从以下网址获取leftImg8bit_trainvaltest.zip作为训练集
https://www.cityscapes-dataset.com/downloads/
2.从以往网址获取gtFine_trainvaltest.zip作为标签
https://www.cityscapes-dataset.com/downloads/
文件结构如下：
FastSCNN
|-- configs
|-- datasets
| |-- cityscapes
| | |-- gtFine
| | | |-- test
| | | |-- train
| | | `-- val
| | `-- leftImg8bit
| | |-- test
| | |-- train
| | `-- val
|-- docs
|-- test
|-- segmentron
|-- tools
将数据集按照以上结构放在代码目录下
安装
请注意，本模型使用了新版本的pytorch以及CANN包，具体版本为：20210617_5.0.T205；
.png)
source 环境变量
source ./test/env.sh
安装
python3 setup.py develop
或使用sh脚本安装
bash ./test/setup.sh
TRAIN
单p训练
source 环境变量
source ./test/env.sh
运行单p脚本
bash ./test/run1p.sh
多p训练
source 环境变量
source ./test/env.sh
运行8p脚本
bash ./test/run8p.sh
模型保存在./runs/checkpoints目录下，以数字命名的pth文件是当前epoch训练得到的权重文件，可用来恢复训练，best_model.pth是当前训练出的最优模型；
运行日志保存至./runs/logs目录下
TEST
测试精度
使用sh文件
bash test/eval.sh
精度对比
GPU8p loss scale使用O1 128混合精度获得的结果为：mIoU:64.46
NPU8p loss scale使用O1 128混合精度获得的结果为: mIoU:63.914
注意事项
由于在./FastSCNN/segmentron/modules/csrc/vision.cpp中添加了Licence，可能会导致程序在调用此文件时报错，只需要删除Licence就可以使用
公网地址说明
代码涉及公网地址参考 public_address_statement.md