Skip to content
This repository has been archived by the owner on Mar 17, 2021. It is now read-only.

Niftynet meeting 2nd october 2017

Jump to bottom
Wenqi Li edited this page Jun 20, 2018 · 1 revision

MICCAI 2017: Release of v0.2

  1. Discussion of tasks/development attributed to v0.2.1 and v0.3 v0.2.1

    • v0.2.1:
      1. net_regress
      2. selective sampler (constraints on label distribution)
      3. loss functions to be re-updated - bugs still there
      4. initial definition of model zoo format
    • v0.3:
      1. evaluation
      2. documentation
      3. 3D resampler (Marc)
      4. hyperparameters optimisation (inclusion of external library?)
      5. modification of input augmentation -> move to more tensor flow operations - avoid GPU memory usage
      6. modification of command line interpretation for application to individual images
      7. uncertainty estimation: format to be discussed. Dropout at inference generating multiple solutions

  2. Discussion around model zoo: Currently zip file with README. Possibility to use tar instead to be able to access directly specific files. Requires additional utility to be able to create the model package (net_export?). Objective would be to save only relevant files at current stage of training. Question of model zoo related to the transformation from bash script to python in demos. Put on hold for now. Will be retackled for v0.3

  3. Discussion around UI: Need to think about end-user. Application dependent. Need some thought to make it generalisable enough. Integration with Slicer? Use of Deep Infer?

  4. File matching: Exact vs fuzzy matching. Default should be exact but fuzzy should be allowed but specified as optional

  5. Need to assess back logs and identify burning issues

  6. Tensorflow version? What version to require? What version to support? For now compatibility 1.1 and 1.3 but 3D resampler from Marc will require 1.3 at least. Any development requiring compiling is committed back to Tensorflow with pull request.

  7. How to support the development? Use of forms? Integration with StackOverFlow with nifty net tag? Zach to look into people able to create tag What format to use? Slack lacks in traceability + history erased. Need to be able to refer to older answers. Need for a FAQ on the Wiki. Potential solution: Discourse (used also by VTK and ITK). In between Slack and Stackoverflow. Possibility to link it to the website, potentially less intimidating.

  8. Pull request from GitHub? To be accepted or not? Specification to contributions? Pb of solving conflicts? Merge requests?

  9. Pb of private branches in repository - Preparation of new code

  10. Pb of not tested elements. Move things to contrib if not thoroughly tested. Contrib should have all nifty net structure + additional features still in development.

Clone this wiki locally