A Python module for exporting pre-trained CFD models to ONNX, making them interoperable with other ML frameworks and compatible with browsers.
It currently supports U-Net architecture and PyTorch models, but it will be soon extended to other frameworks and architectures.
Reproducible examples can be found at openfoam-cfd-rom usign DeepCFD.
The module can be installed with:
pip3 install cfdonnx
Usage: python3 -m cfdonnx [OPTIONS]
Options:
-n, --net TEXT network architecture: UNetEx or AutoEncoder (default: UNetEx)
-i, --input PATH checkpoint (default: checkpoint.pt)
-o, --output PATH ONNX output file (default: checkpoint.onnx)
-k, --kernel-size INT kernel size (optional, read from state_dict['kernel_size] by default )
-f, --filters TEXT filter size, e.g. 8,16,32,32 (optional, read from state_dict['filters'] by default)
-c --channels INT number of channels (optional, read from state_dict['input_shape'] by default)
-x --nx INT X dimension (optional, read from state_dict['input_shape'] by default)
-y --ny INT Y dimension (optional, read from state_dict['input_shape'] by default )
-o, --output PATH Save model path (default: mymodel.pt)
Example:
python3 -m cfdonnx \
--net UNetEx \
--input flowAroundObstacles.pt \
--output flowAroundObstacles.onnx
You can use your CFD ONNX models on runtime in Babylon.js as showcased at https://play.simzero.com/#D3SFTH#6 for the flowAroundObstacles example.
A generic template for using ONNX is also available at https://play.simzero.com/#WIB297#1.