NiftyNet Hackathon Debrief 06th June 2019

NiftyNet Hackathon Debriefing

2.30 - 4.30 Thursday 6th June

Attendance: Zach, Jorge, Kerstin, Wenqi, Dan, Maurizio, Reuben, Lucas, Marta, Ben, Pedro, Richard, Tom Va, Marc, Irme, Eric, Pritesh, Carole

Padding Use volume padding to pick minimum size to padd things to->efficient inference / Useful for inference memory ZER
Convolutional layer feature normalisation (batch instance group) Design architecture ZER
Demo for learning rate scheduler self contained Pedro
NiftyReg image sampling - replace one resampling GPU implementation of NR faster - Pb compilation - Using docker released by TF - Lucas
Investigation of 1.13 / 1.14 Change in CRF Change also in misc_io Dan
Performance handler - Tom Irme
Early stopping Different strategies implemented Robust 95% Smooth historyError improvement -> makes sure you set num_classes to the number of unique labels. - Irme Tom
Deep Boosted Regression in the contrib. Kerstin
Multiple output and csv output - Carole Marta
Fixed mismatch between the input and output sizes in regression application - Pritesh
Resampler interpolation - Samuel
⅓ of networks commented - Marta

Multi step application Documentation demo -> Handler markdown file to do that 2 days of work - Pedro Richard
Memory based image as input Application API Pickle 1 more week npy array - Wenqi
Write demo for registration with NN - Lucas TomV.
1.14.0 1.13 0 Tensorflow Test class not working anymore - Need to change the naming. Waiting for all features to be included and then changing all tests - Dan
Epoch / shuffle Needs to go through tests - Dan
Whole volume validation Design done 1 week - Tom Irme Wenqi
Implement csv reader regression application - Pritesh
Wiki pages for ini file info - Ben tomorrow
CSV reader preparation - Carole 1 week
CSV sampler - Carole 2 weeks
Sparse label missing data - Ben next week
TODO: Reuben to merge multi-input application
TODO: Marta to do more documentation on networks.

Config file / Yaml - Ben / Tom / Eric 15’ Need to assess how much work to pull dev into old YAML branch
Towards API - Wenqi 15’
- Image Reader and Image Samplers done
- TODO: Implement numpy array as input.
- TODO: Application as modules - Will enable a user to run inference straight from a Jupyter Notebook.
Towards TF 1.14 / Move to 2.0 - Eric / Dan 15’
- Testing needs to be adjusted.
- Porting is easier than developing in the first place.
- TF 2.0 demo in the style of Pytorch demo. Ben, Tom, Eric, Wenqi, Dan use Hippocampus segmentation