This repository has been archived by the owner on Mar 17, 2021. It is now read-only.
-
Notifications
You must be signed in to change notification settings - Fork 404
Niftynet meeting 6th november 2017
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:
- Support for both python 2 and python3? For now keep both: running CI for python 3 and do the nightly build with both.
- 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.
- 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
- 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
- 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
- Versioning: v0.2.1 ready after Wenqi testing of model zoo. v0.3 for Christmas
- 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.
- Resampling - Transformation applied before loading. For the moment single thread. To be included in v0.4