History

BaofengZan 37f03a77c5 fixed some issues (#400 ) * 1 动态输入 2 图像预处理和后处理和pytorch同步 3 代码格式统一 * 重新梳理代码 * tab 替换为4 space * HRNet-Semantic-Segmentation * update result img * update readme * Synchronize * Modify some bugs		2021-02-07 17:41:16 +08:00
..
CMakeLists.txt	fix and optimize	2021-01-29 16:20:48 +08:00
common.hpp	rename files, update readme	2020-12-28 10:08:16 +08:00
demo.py	rename files, update readme	2020-12-28 10:08:16 +08:00
hrnetseg.cpp	rename files, update readme	2020-12-28 10:08:16 +08:00
logging.h	rename files, update readme	2020-12-28 10:08:16 +08:00
README.md	fixed some issues (#400 )	2021-02-07 17:41:16 +08:00

README.md

HRNet-Semantic-Segmentation

The Pytorch implementation is HRNet-Semantic-Segmentation-v1.1. The implemented model is HRNetV2-W18-Small-v2

How to Run

1. generate .wts
Download code and model from HRNet-Semantic-Segmentation-v1.1 and config your environments.

Put demo.py in the YOUR_ROOT_DIR\HRNet-Semantic-Segmentation\tools folder, set savewts in main() as True, and run, the .wts will be generated.

cmake and make

mkdir build
cd build
cmake ..
make
sudo ./hrnetseg -s             // serialize model to plan file i.e. 'hrnetseg.engine'
sudo ./hrnetseg -d  ../samples // deserialize plan file and run inference, the images in samples will be processed.

Result

TRT Result:

pytorch result:

Note

Some source codes are changed for simplicity. But the original model can still be used.

All "upsample" op in source code are changed to mode='bilinear', align_corners=True
Image preprocessing operation and postprocessing operation are put into Trt Engine.
Zero-copy technology (CPU/GPU memory copy) is used.