This repository has been archived by the owner on Jun 26, 2021. It is now read-only.
Releases: delira-dev/delira
Releases · delira-dev/delira
Release v0.6.0
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
andval
- 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
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
dict
s during prediction and training [commit] - Kfold: The old
kfold
- function has now merged with the newkfold
,stratified_kfold
andstratified_kfold_predict
to a newkfold
function supporting all these behaviors. [commit] - Rename tf Backend: The
tf
backend has been renamed totensorflow
during installation. The new command to installdelira
with TensorFlow backend is now:pip install delira[tensorflow]
[commit]
Non-Breaking Changes
- Unify Trainer: Merging Code from Backend Trainers (
TfNetworkTrainer
andPyTorchNetworkTrainer
) intoBaseNetworkTrainer
; 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 ofmultiprocessing
[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
Bugfixes
- Random sampler can now be seeded via numpy for completely determinitic behavior (a7c10ee)
Release v0.3.1
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