We’re planning a CCP-SyneRBI/CCPi hackathon on integration of SIRF/STIR/CIL with PyTorch. The aim is to discuss how to efficiently link data structures in SIRF (and its underlying libraries including STIR) and CIL to PyTorch tensors, and hack some prototype implementation together. 

Contact tomography@stfc.ac.uk if you are interested in taking part to the event.

Aim: optimise usage of CUDA in STIR/SIRF/CIL when interfacing to pytorch

Example tasks:

• STIR 

• • Allow use CUDA managed pointers for image objects 
  • Add CUDA numerical operations ? 
  • Adjust STIR’s parallelproj interface to use above 
  • Expose via SWIG to python 

• SIRF 

• • Pass through STIR image objects appropriately to python 

• CIL 

• • Investigate DataContainer agnostic on backend, e.g. numpy, cupy, pytorch… 
  • Efficient data exchange with GPU backend, TIGRE/ASTRA 
  • making a denoising application with torch function 
  • run CIL Algorithm with DL denoising step 

Tentative schedule