NiftyNet Dev Meeting 10th May 2019

NiftyNet meeting 10th May 2019

Attendance

Eric, Marta, Samuel, Stian, Felix, Wenqi, Dan, Pedro, Tom Ve. Marc, Jorge, Ben, Tom Va., Jorge, Carole

Introductions

Progress since last time

• Felix:

1. Refactoring into latest version of NN
2. Probability layer Network - Privilege Modularity Decouple as much as possible

• Ben:

1. Gaussian sampler separated from VAE in progress (ready after NIPS)
2. Uniform sampler documentation - Bug fix
3. Input with missing data with placeholders instead

• Irme

1. Whole volume validation in progress

• Carole:

1. Loss functions
   • Regression (Smooth L1, Cosine regression)
   • Classification with multi rater
   • Segmentation Volume enforcement

• Tom:

1. Pull CSV reader Look at modularity -> Too tailored Keeping it more general

• Stian:

1. Input changes: Decouple application from input
2. Creation of wrapper class for ImageSource and Sink NiftyNetIO. Not Streaming

Refactoring discussion for NN (Ben presenting suggestions/ideas)

• More homogeneity needed
• Decoupling ini files from API capabilities - open for potentiality of new features to lite users. Compatibility with TFServing
• Toolkit from framework Model Zoo / App Zoo
• Example of new possible application: Inference on single image Training / Validation

Decision - Make NN an API

Transition to Tensorflow 2.0 (Tom Va. presenting)

Changes

• // Pytorch
• Removal of session f still a graph numpy array input and output
• Only tf.pyfunc
• Possibility to keep better track of variables
• Need to use tf.keras.layers or tf.layers - Difference to be further investigated - tf.keras.layers serialisable Potentially more useful

Meaning for NN

• May need to change layer abstraction?
• Application driver to change
• All applications to be rewritten
• Try to stay closer to keywords of ML processing (fit, predict, compile)
• Continue using current handlers
• Change variable collectors

Points of discussion

• Should we move now or wait for version beta?
• How in the context of all recent features integration?
• Possibility - First transfer NN as an API with the Eager tf.13 framework and then transition
• Need to maintain current working state and continue for a while to support it

NN Hackathon

• Tuesday 4th June - Thursday 6th June
• Marc to find room
• 9.30 - 4pm
• Goal integration of all recent features
• Last half day - with Wenqi - More in depth discussion on refactorisation as API

Until then: everyone to indicate features to be included under Project board for NN Hackathon