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This repository has been archived by the owner on Mar 17, 2021.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

Niftynet meeting 6th november 2017

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

Closed issues:

  • Loss metrics
  • Git Hub pull request
  • Continuous integration service (CI running)
  • Configure for any location for NiftyNet
  • CI on cpu
  • Image dimensionality inconsistencies - Differences in rounding between python 2 and 3.
  • Initialisation factory

New issues/ Discussion:

  1. Support for both python 2 and python3? For now keep both: running CI for python 3 and do the nightly build with both.
  2. Config file - Complexification - Hierarchical structure
    • Need for more structure in the relationship between options
    • Suggested options: Python file - Multi level config file - Hierarchical - Proposed format protobuf / json / yml. With protobuf possibility of backward compatibility vs YML more visual (tree, more user friendly) -> Dzhoshkun to try a few things out
    • Hierarchy of help and parameters (Carole) Need for double layer of config? Need for a light weight config file and a larger that contains everything. Only parameters specified in config file will overwrite default ones.
  3. Offline validation workflow - evaluation Internal (using config file to access inferred images) or external. Iteration subject by subject or label. Possibility to do threshold iterations for ROC curve - Post training - Need to extend for multiple checkpoints
  4. Online volume level for validation in training - For the moment at the window level - Progressive steps to be taken:
    • Validation / Test Loss
    • Visualisation in Tensorboard of volumes
    • Used for weight maps creation Need for validation assessment over loss function for learning
  5. Model zoo fully working via dropbox download of models. Need to switch to git backend: One model on each branch, avoid cloning everything. Possibility to download a specific branch. Hosting could be done in git repository. Advantages of git: Centralised version, version control, listing. Need to avoid change for the user in terms of requirement
  6. Versioning: v0.2.1 ready after Wenqi testing of model zoo. v0.3 for Christmas
  7. Support for user - Discourse or Stack Overflow? Bug tracking already there through Git. Need for something else? Discourse could enable Mailing list for announcement / Long discussion. Discourse set aside for now.
  8. Resampling - Transformation applied before loading. For the moment single thread. To be included in v0.4
Clone this wiki locally