1.3.1

Changed

• Add reference to gcc 12.3 compatibility in README

Fixed

• Fix attempt to assign size of random seed in random_setup
• Fix attempt to assign size of random seed in test_network
• Fix attempt to assing size of random seed in example/simple
• Fix abstract interface use in base_layer_type
• Add appropriate references to neural-fortran in CONTRIBUTING.md and associated procedures, and modules

Full Changelog: 1.3.0...1.3.1

1.3.0

Added

• batchnorm1d_layer_type (handles both 1D spatial channel data and rank 0 channel data)
• base_layer submodule file
• network submodule file
• Added comments to network and base_layer procedure interfaces

Changed

• Improved comment and headers in all source files
• Improved comments in all test source files
• Added interfaces within base_layer submodules for ease of readability
• Added interfaces within network submodules for ease of readability
• Moved methods of base_layer into submodule
• Moved methods of network into submodule

Fixed

• Remove reference to doc/ directory
• Fixed typo in example library
• Fixed example/mnist/src/main.f90 using wrong constants module
• Fixed cmake isntall path prefix duplication issue

Removed

• Deprecate and remove step_decay and lr_reduce_on_plateau procedures

Full Changelog: 1.2.4...1.3.0

1.2.4

Added

• fpm continuous integration workflow

Fixed

• move location of output rank test in maxpool3d test to fix memory access issue

Full Changelog: 1.2.3...1.2.4

1.2.3

Added

• More unit tests
• Code coverage reporting

Changed

• Unify shuffle and split procedure names
• Change handling of magnitude optional argument in renormalise_sum
• Convert copy and reduce to child procedures of network_type
• Set already_initialised to False by default for random_setup
• Update list of file descriptions in README

Fixed

• Index reference in initialiser_ident
• avgpool3d forward pass from max to sum
• avgpool3d output shape initialisation
• Correct handling of "valid" padding method
• Random seed for shuffle
• conv2d input gradient calculation
• conv3d input gradient calculation
• Incorrect set_gradient spread dimension
• MAE total loss using MSE instead of MAE
• metric_dict allocation handling error
• icount field separator
• Optional argument handling in renormalise_norm

Removed

• Remove assign_list procedure from mod_tools_infile
• Remove assign_listvec procedure from mod_tools_infile
• Remove cat procedure from mod_tools_infile

Full Changelog: 1.2.2...1.2.3

1.2.2

Added

• Add fpm compilation handling
• Add status badges to README

Changed

• Move example-dependent library from /example/include to example/example_library
• Update README to include fpm documentation
• Update README to include dependencies
• Migrate repository from GitLab: https://git.exeter.ac.uk/hepplestone/athena
• Migrate repository to GitHub: https://github.com/nedtaylor/athena

Fixed

• Fix typos in README
• Fix typos in CHANGELOG

Removed

• Remove parallel reference from examples

Full Changelog: 1.2.1...1.2.2

1.2.1

Ported from GitLab on 2024-02-18

Released on GitLab on 2024-02-08

https://git.exeter.ac.uk/hepplestone/athena/-/releases/1.2.1

Changes

• Tidy up TODO

Fixed

• Fix typos in README

Removed

• Remove parallel build option
• Remove src/athena_omp.f90
• Remove superfluous comments and urls

1.2.0

Ported from GitLab on 2024-02-18

Released on GitLab on 2023-12-18

https://git.exeter.ac.uk/hepplestone/athena/-/releases/1.2.0

Added

• Add inference to dropout layers
• Add 2D and 3D batch normalisation layers
• Add 1D and 4D flatten layers
• Add predict procedure to network_type
• Add network reset procedure
• Add Adam optimiser
• Add procedures to get and set network parameters
• Add procedures to get and set network gradients
• Add procedure to get layer output
• Add new test directory
  • Test jobs test the working of the library
• Add test jobs
• Add more example jobs
  • Example jobs show the functionality of the library
  • batchnorm, dropblock, 3D, sine, and simple function
• Add average pooling 2D and 3D layers
• Add RMSprop, Adam, and Adagrad optimisers
• Add learning rate decay
• Allow accuracy scoring methods to be imported

Changed

• Handle batch size in each layer
• As such, increase rank of io data (and gradients) for all layers by 1
• Handle loss and metric network initialisation in separate procedures
• Convert optimiser to parent and derived types
• Change optimise procedure to minimise
• Handle network learning in network, rather than individual layers
• Rename test/ to example/
• network_tyep%forward now handles assumed rank
• input_layer_type%set now handles assumed rank
• Add abstract conv, pool, and batch layer types
• All layer forward and backward passes take defined shape
  • Previously, rank was defined and shape was assumed
• Flatten layers now use output_shape as well as num_outputs
• Make accuracy score functions available to import

Fixed

• Fix drop inference
• Fix backpropagation for overlapping windows in maxpool layers
• Fix dropblock 2D and 3D generate_mask to check for .lt. gamma instead of .gt.
  • This was already correctly handled in dropout
• Fix R2 score function
  • Add zero division check

Removed

• Remove mod_batch_norm.f90
• Remove update procedure from learnable_layer_types
• Remove parameter velocities from learnable_layer_types