PyTorchx

This is a brother project with wang-xinyu/tensorrtx.

Popular deep learning networks are implemented with pytorch in this project. And then weights files are exported for tensorrt implementation.

Test Environments

1. Python 3.7.3
2. cuda 10.0
3. PyTorch 1.3.0
4. torchvision 0.4.1

prepare pytorch-summary

pytorch-summary is a very useful tool for understanding the model structure, for example it can output the dimensions of each layer.

Clone, and cd into the repo directory.

git clone https://github.com/sksq96/pytorch-summary
python setup.py build
python setup.py install

Run

Most of the models are from torchvision, exception for yolov3, which has a readme inside.

A file named xxxnet.py can do inference and save model into .pth. And a file named inference.py can do inference and save weights into .wts, which is used for tensorrt.

For example, googlenet,

cd googlenet
python googlenet.py  // do inference and save model into .pth firstly.
python inference.py // then do inference and save weights file