Release v0.6.0

This is a major release since it changes the license to a dual licensing strategy. This may cause some troubles if you want to use the latest version, so make sure to check the README and the LICENSE.

Release 0.5.0

Bugfixes

• removed many mutable default arguments [Commit]
• Correctly unload TF in Tests [Commit]
• Split TF-Tests in Graph and Eager Backend [Commit]
• Backend-Checks for Tests (avoid duplicate testing) [Commit]
• Move to new APEX API [Commit]
• Fix Graph-Based Logging [Commit]
• Fix Utils Imports [Commit]
• Move to new Slack API [Commit]
• Fix Travis CI/CD [Commit]
• Fix Config overwrites [Commit]
• Minor type fixes inside the trainer [Commit]](e7294fd)
• Fix Typecheck for Transforms [Commit]
• Config Traversal [Commit]
• Fix logging frequency and timeout [Commit]
• Fix Logging Steps [Commit]

Breaking Changes

• Logging is now done outside the closure
• Metrics are now based on numpy and are calculated the same way for train and val
• Remove trixi dependency [Commit]
• Remove deprecated codeparts (TorchvisionClassificationDataset , train_test_split , Nii3DLazyDataset , Nii3DCacheDataset , [delira.utils.imageops.py](http://delira.utils.imageops.py) , DefaultOptimWrapperTorch , DebugMode Context Manager and the following requirements: scikit-image, pillow, SimpleITK, psutil [Commit]

New Features

• Slack Notificator [Commit]
• Function to seed all backends [Commit]
• Add Callbacks after and before epoch, after and before training and after and before every step (regardless of train and val) [Commit]
• New Custom Data Pipeline (completely replaces the batchgenerators pipeline) for more customization and better integration of our sampling [Commit]
• Register Logger in Callbacks [Commit]
• Onecycle Scheduler Callback [Commit]
• Allow creation of config from system args [Commit]

Bugfix Release v0.4.1

Bugfix Release v0.4.1

Bugfixes

• Several improvements to reduce memory usage (and fixed an actual memory leak)
• Fix the Search for previous trainings [commit]
• Convert Metric Values to Numpy [commit]
• Epochs correctly saved in checkpoints [commit]
• Refactoring to clean up global namespace [commit]

Improvements

• Add numba jitting for transforms [commit]
• Bump TensorFlow to 1.14 [commit]
• Versioning via versioneer for more detailed versions [commit]

Release v0.4.0

Release v0.4.0

New Features

• Slack: We now have a community chat on slack [commit]
• Paper: We wrote and submitted a Paper to JOSS [commit]
• Dataset Integration Guide: New tutorial to switch from old dataset API to new dataset API [commit]
• Predictor: Introduce new Predictor class to predict from network without having to initialize Trainer with dummy values [commit]
• PEP-8: Complete Codebase is now PEP-8 compatible and CI/CD now checks for it [commit]
• Debug Mode: New debug mode disables all multiprocessing (currently multiprocessing is only used during data loading); Might slow down the program but increases debug capabilities [commit]

Breaking Changes

• Return dicts: Models must now return dicts during prediction and training [commit]
• Kfold: The old kfold- function has now merged with the new kfold, stratified_kfold and stratified_kfold_predict to a new kfold function supporting all these behaviors. [commit]
• Rename tf Backend: The tf backend has been renamed to tensorflow during installation. The new command to install delira with TensorFlow backend is now: pip install delira[tensorflow] [commit]

Non-Breaking Changes

• Unify Trainer: Merging Code from Backend Trainers (TfNetworkTrainer and PyTorchNetworkTrainer) into BaseNetworkTrainer; Will increase maintainability and ease of adding new backends [commit]
• Unify Experiment: Merging Code from Backend Experiments to BaseExperiment [commit]
• Sampler StopIteration: Refactor samplers to use a common function to check if StopIteration must be raise (where possible) [commit]

Bugfixes

• TF Resnet18: ResNet18 in TF now matches the official paper [commit]
• Parameters: Correct Permutations in Parameters class [commit]
• Infinite Recursion: Infinite recursion in LookUpConfig due to wrong typechecks is now fixed [commit]
• DataLoading num_batches: The number of batches is now calculated without any influence of multiprocessing [commit]
• DataLoading: Instabilities due to inter-process communication [commit]

Release v0.3.2

Release v0.3.2

Added

• WeightedPrevalenceSampler (433cccb)
• _DatasetIterator allowing iteration over a dataset with a simple loop without having to iterate over indices explicitly (70d6232)
• Support for TF1.13 (5e28172)
• Support for Python 3.5 (5db1a17)

Breaking Changes

• New Dataset API (#88 )
• Changed test framework from pytest to unittest (a709809)

Bugfixes

• Random sampler can now be seeded via numpy for completely determinitic behavior (a7c10ee)

Release v0.3.1

Bugfix-Release 0.3.1

Bugfixes:

• loading behavior with weights_only=True for torch backend (fff41cf)
• Behavior of BlankDataset (353648e)
• Refactoring of backend choosing (01f25de)

Release v0.3.0

Release 0.3.0

Added

• TensorFlow support including

  • AbstractTfNetwork
  • TfNetworkTrainer
  • TfExperiment
  • ClassificationNetworkBaseTf

• Basic Test-API in Experiment (commit)

• More Unittests

• Methods to get a subset of each Dataset

• Stratified Kfold

Bugfixes

• dropping last batch in PyTorchNetworkTrainer.predict (commit)
• Check for installed backends

First Beta Release

Based on PyTorch, batchgenerators and trixi it provides a framework for

• Dataset loading
• Dataset sampling
• Augmentation (multi-threaded) including 3D images with any number of channels
• A generic trainer class that implements the training process
• Already implemented models used in medical image processing and exemplaric implementations of most used models in general (like Resnet)
• Web-based monitoring using Visdom
• Model save and load functions