v0.3.0

Additions

• Added inception module.
• Added pool downscaling.
• Added kernel, stride and padding to pool layer.
• Added layer_num attribute to Network to select layers to use.
• Added composite layers from .json files.
• Added filters factor parameter to convolutional layers.
• Added inception.json composite layer.
• Added mode to pooling layer.
• Added compatibility for non-square kernels.
• Added shapes attribute to Network.
• Added layer arguments check.
• Added adaptive_pool.
• Added inceptionv4.json example.
• Added output shape warning suppression for composite layers.
• Added defaults.py for defaults for each layer.
• Added overriding defaults as arguments in net.
• Added layer_test_network.json to test layer functionality.
• Added layer information for shortcut shape error.

Changes

• Changed reshape to only ignore first dimension and treat others as free.
• Changed output shape error to warning.

Fixes

• Fixed network construction errors due to incorrect shapes.
• Fixed install not including subdirectories.
• Fixed dictionary overwriting.
• Fixed _kernel_shape only accepting integer padding.
• Fixed concatenate not using dim parameter.
• Fixed concatenation copy bug.
• Fixed sampling shape bug.
• Fixed default padding in pool layer.
• Fixed layers not checking supported parameters if they use other layers.
• Fixed concatenate layer not working for 1D inputs with no channel dimension.
• Fixed shortcut layer converting shapes to arrays.
• Fixed recurrent incorrect shape if using concatenate.
• Fixed pool shape copy override.

Removals

• Removed squeeze layer as reshape achieves the same.
• Removed load_network as this should be implemented per project.
• Removed composite.py.