ChangeLog of NNHelferlein package

todo

• use CUDA.CuIterator in train?
• padding no longer imported from NNlib (incompatibility wirh AutoGrad)

1.3.2

• tidy-up dependency jungle
• Padding added to emebdding layer

v1.3.1

• l1 and l2 decay always parallel to learning rate decay
• severeal bioinformatics tools (aminoacid embedding, blosum, vhse8)
• dataframe_minibatch default "y" changed to nothing.
• Bioinformatics: Aminoacid tokenisation added
• GPU selection added (not yet exported)
• grouped convolutions fixed

v1.3

• Transformer API added for Bert-like architectures
• Transformer example
• ramp-up of beta added to VAE
• disambiguate vae signature

v1.2

• imagenet preprocessing fixed for vgg and resnet
• ResNetBlock added
• ResNet added
• Padding layer added
• print_network changed to summary
• Pretrained nets saved at zenodo and simplified constructors added
• AbstractNN and AbstractLayer added
• copy model and save/load as JLD2 added

v1.1.2

• Depthwise conv-layer added (experimental)
• focal loss functions added to classifier
• FeatureSelection layer added
• explicit signature added for 3d-convolution
• train: possibility to disable tensorboard logs
• train: possibility to return losses and accs for plotting after training

v1.1.1

• some docstring cosmetics
• Activation Layers added
• layer GlobalAveragePoling added
• pre-trained vgg example fixed for changed "import-HDF"-interface
• hdf5 import with all kwargs possible
• added: Layer + Layer = Chain
• changelog added to docu

v1.1.0

• documentation for release added
• split_minibatches() made stable (never returns an empty iterator)
• docs slightly re-organised
• Gaussian Layer added
• minibatch iterator for masking added

v1.0.0

• initial release