This is where i save model used in face-recognition task

1. Setup environments:

conda create --name model_env python=3.6
conda activate model_env
pip install torch==1.8.0+cpu torchvision==0.9.0+cpu -f https://download.pytorch.org/whl/torch_stable.html
pip install -r requirements.txt

2. Run:

To convert pytorch model into onnx, run:

conda activate model_env
python main.py

To test time running on each onnx model, run:

conda activate model_env
python time_check.py

change config in config.py base on what you want to test

3. Parameters number and speed of modify network:

• In experiment with IR-50 network, i replaced last layer of original network with the last layer of Ghostnet, result in the decreasing up to 10 million parameters

IR-50 parameters

IR-GHOST-50 parameters

• In the time Testing of previous models, the modified one (IR-GHOST-50) has the average runtime on 1000 ([1, 3, 112, 112]) numpy tensors lower than the Original network (IR-50) nearly by half.

Time Test:

4. Experiments Training on modify (augmentation, mask) Glint360k:

Some tensorboard images:

More detail is in: Tensorboard

https://pytorch.org/tutorials/advanced/super_resolution_with_onnxruntime.html

https://onnxruntime.ai/

https://github.com/ZhaoJ9014/face.evoLVe