Niftynet meeting 5th february 2018

Attending: Tom V., Wenqi, Irme, Marta, Fernando, Richard, Felix, Zach, Imanol, Eli, Carole, Tom D., Jorge

1. Gradient checkpointing

• Checkpoints Hard coded for each network - For new networks ask for people to checkpoint them. Explicit flag in config files allowing for full / partial / no checkpointing

2. CMIC Lab / Git -

• Bennett Landmann Core funding Part time Development from Vanderbilt - Formalise consortium - Multi institution - 3 Universities - Nifty branding? NiftK need to be kept.
• GitHub transition. Ask for academic status to set up private branch. Multiple developers. Need for access.
• Reverse roles between GitHub and CMICLab. Each user need to have academic status and activate. Internal Niftynet contrib -
• Make sure private branches - Also for new comers with code not ready. Action:Jorge to check and report on academic status and private branches.
• Each university to have private mirror of the GitHub repository.

3. Test servers and continuous integration

• Migration of the CIA continuous integration private? CI with GitHub? No with CPU on GitHub. Action: Tom D (with Dzhoshkun and Allan) to check feasibility of mirroring to CMICLab for testing
• Pb Windows not suitable for GPU so far. Good Linux test server but may break if used by another job simultaneously. Need dedicated two test servers. Need to go ahead anyway. Need for enough GPU power for nightly build NiftyNet.

4. Website

• Need for link to read the docs on the io website - Need good redirection. Need for maintenance.
• Solution: as Git Project under master. Can be pulled from Git.

5. Config file refactorisation

• YAML with Dzhoshkun - Appropriate testing of exit and config. Meeting - Carole / Dzhosh / Wenqi

6. Classification application

• For the moment image input output - Transition from input to database? Google API dataset? More generic on what is an input? Prepare data input layer Action - Meeting to be arranged Imanol / Eli / Wenqi / Jorge / Carole

7. Interpretability of features / Features maps in Tensorboard / Classification activation maps / Debugging

• Tensorboard summary - Activation maps or in Evaluation code ? Turn on flag Activation map on Tensorboard? Simply secondary output. Image summary and add normalising layer. Not possible directly through config file
• Need Tutorial on how to add outputs - Possibility of adding External function / contrib for output?
• Debugging question - Need of assessing compatibility between Eager and checkpointing ? Graph on the fly with parallel engine.
• 2 Birds one stone solution: Packaging classification application with class activation map tutorial / demo. 2D should not be the focus but rationale is to bring people in with easy - Action: To be first discussed at IO meeting Comment: NiftyNet should not become a KITCHEN SINK !!!

8. Educational / Tutorial aspects of Niftynet

• Need for tutorials on transfer learning?
• How to separate feature extraction from output layer
• Need for Very good tutorials? Pb on return on investment? Potentially as Master student projects

9. XNat support

• Comes back to how to handle input data that can be quite messy - Also in IO meeting?
• Solution: XNat in net download. Utilities to download from XNat - trained network. upload on XNat as well.