v0.6.0

• Add new augmentation methods:
  • random image region blurring (BatchCreatorImage.blurry_blobs_probability)
  • random erasing
    • with noise: BatchCreatorImage.noisy_random_erasing_probability)
    • with constant values (BatchCreatorImage.uniform_random_erasing_probability)
• Remove broken --gpu=auto option from elektronn2-train
• Add zxy flag to BatchCreatorImage, making loading data in xyz axis order possible
• Improve version strings with versioneer (git revision and uncommitted modifications are now reflected in elektronn3.__version__, making identification of the current version easy)
• Various small improvements and bug fixes

This is the last planned release of ELEKTRONN2. We are focusing on elektronn3 now.

v0.5.1

• Fix model.modelload() when using the new Conv constructor with optional pool_shape. Loading models from .mdl files was broken in 0.5.0 (#13, c2a7b8a)
• Better progressbar for model.predict_dense() (thanks @mahsabh, 6ee6d18)
• Extend maximum runtime and iterations of example networks to make sure they can be fully trained (970bfa4, 86fea4c)

v0.5.0

• The new, more flexible AutoMerge operation replaces UpConvMerge (UpConvMerge is still available as an alias to AutoMerge for backwards compatibility) (da5b648, 45c507e, e45082c, 660c945).
  • New merge_mode argument, which specifies the Node type to perform the merging ('concat' or 'add'. This simplifies usage of the new Add node (using merge_mode='add').
  • Automatic uponvolutions can now be disabled (disable_upconv=True), making it a mere autocrop-merge wrapper.
  • Better log output that tells how the parents are merged and what operations are automatically performed for merging. If an UpConv is inserted, it is explicitly logged.
  • kwargs for the possible UpConv call are now bundled to upconv_kwargs to clean up the signature (potentially breaking change).
  • Renamed some parameters (breaking, see new signature of AutoMerge)
  • If used for two differently-sized input nodes, the larger one doesn't need to be the first one any more. Sizes are detected automatically now.
  • The concept of high-res and low-res inputs is now explicitly optional (this still needs some refactoring to be more clear).
• Better exception handling by the trainer (71914ce). Errors messages are now displayed correctly when preview predictions fail.
• Reduce annoying malis import warning to "debug" log level (de8f414). Note: Fixing malis support is not on my short-term TODO-list and malis is not properly built in or documented, so these warnings were just distracting. Related to #3. (If there is any real demand for malis support, please tell me, then I will try to fix it soon)
• (WIP) Support SAME-mode convolutions (border_mode='same') via Theano's 'half' mode (3dd35df). There is an issue with the fov calculation of Conv nodes that use border_mode='same'. The invalidate_fov=True hack (0691711) can be used as a temporary fix until the fov calculation works as intended.
• (WIP) Experimental new Pad node for value-padding along spatial a image tensors (1d27a87).
• More nodes print their __repr__ upon initialisation, making logs more clear.
• Support (path, key) tuples as hdf5 data path for preview predictions (81483b7).
• Fix preview predictions of unet3d example by using validation data for previews (preview data of neuro_data_zxy was too small). (05adabe)
• Fix sphinx CSS issue. (c7ba3ae)

v0.4.0

• (1,1,1)-maxpoolings are now implied in Conv Node constructors → less redundant code in network models that don't use conv-pool layer stacks (917bdea, c81294e)
• New Add Node type for merging two network branches (cd87b78)
• Use the short form nm for the elektronn2.neuromancer module throughout examples and docs (8a412d0)

v0.3.0

• Support ELU and SELU activation functions for Self-Normalizing Networks (#12, e27276a)
• Log everything to stdout instead of stderr to make terminal output recording less confusing (5214e30)
• Fix file system permission problems on multi-user systems by changing output file names (8a436ef, 7661fe7)
• Documentation fixes, tables are now displayed correctly

v0.2.0

• New elektronn2-profile command for profiling and finding optimal input sizes of already trained neural networks
• Default save directory of examples is now ~/elektronn2_training
• Disable initial debug plots on elektronn2-train startup (3eec2c9)
• Many improvements to U-Net-based config examples (8866e1b, 25cd87c)
• Move examples that are yet not ready to use to examples/wip
• Many documentation updates (New section on predictions, new detailed installation instructions, better index page, improvements to some docstrings)
• Correct many typos in code and docs
• Minor bug fixes

v0.1.0 "NEUROMANCER"

First pre-release version of ELEKTRONN2. There are no major known bugs, but breaking changes in the next versions are expected.