The conversion is expected to be run from the root directory of the project. First, we download the official Pytorch checkpoints from CR-FIQA: https://github.com/fdbtrs/CR-FIQA.
We put the checkpoints in a folder checkpoints/ within the root directory.
Our folder structure should then look as follows:
root/
├── checkpoints/
│ ├── onnx/
│ │ ├── ...
│ └── pytorch/
│ ├── CRFIQA-L/
│ │ ├── 181952backbone.pth
│ │ └── training.log
│ └── CRFIQA-S/
│ └── 32572backbone.pth
├── cpp/
├── data/
├── python/
├── results/-
Navigate into
pythonfolder -
Set up conda environment
conda env create -n convert -f environment.yml
-
Install pip dependencies
conda activate convert pip install -r requirements.txt
Now, we can convert the model by activating the conda environment and running from the root directory:
python python/convert_model.pyFor testing the model checkpoints, we use the Labeled Faces in the Wild (LFW) dataset: https://www.kaggle.com/datasets/jessicali9530/lfw-dataset.
We unzip the images into the data directory at data/lfw-deepfunneled/.
Next, we can run the Pytorch script by running:
python python/compute_quality.pyThe results will be stored in a .txt file at results/crfiqa-l_lfw.txt.
Next, we want to test the converted onnx model on the LFW dataset.
Therefore, we run the following python script:
python python/run_onnx.pyThe results will be stored in a .txt file at results/crfiqa-l_lfw_onnx.txt.
To test the onnx model in a CPP framework, we need to setup the corresponding environment.
The project requires cmake >= 3.23 and conan > 2.0.
If you cannot install these dependencies directly on your system, we recommend to use a conda virtual environment.
To set up the conda virtual environment, navigate into the cpp folder and run
conda env create -n conan -f environment.ymlThe required packages are then installed in a conda environment with the name conan.
Fur further information on compiling the C++ code, look into cpp/README.md.
List of useful resources:
- https://github.com/onnx/tutorials
- https://pytorch.org/tutorials/advanced/super_resolution_with_onnxruntime.html
- Use opencv instead of onnxruntime
- Compare results between pytorch model and onnx model
- Add cpp implementation for onnx model
- Replace absolut paths in example.cpp