Add Linear and improve Lambda

Added

• co.Linear module.
• Improved repr in co.Lambda.
• Option to skip unsqueeze in co.Lambda.forward_step.

Multistream support

Changed

• Renamed co.Parallel to co.BroadcastReduce.

Added

• co.Broadcast module.
• new co.Parallel module.
• co.Reduce module.
• Automatic inference of co.Broadcast.num_streams in co.Sequential.

Closure and Conditional modules

Added

• co.Lambda module.
• co.Add module.
• co.Multiply module.
• co.Unity module.
• co.Conditional module.

Bug-fixes

Fixed

• Bug in forward_stepping.
• Bug in clean_state.

Another minor API change, bug-fixes and implementation restructure

Fixed

• Bugs in forward_step(s) with update_state=False.

Changed

• forward_steps interface to always include pad_end argument.
• Name of "interface.py" to "module.py".
• Implementations of forward_step(s) to be consolidated in CoModule.

Removed

• Padded interface

Minor API change, bug-fixes and improved test-coverage

Added

• Independent state_dict and load_state_dict functions.
• Added nonempty check for aggregation functions in Parallel.
• update_state argument to all forward_step(s) methods.
• Additional tests for edge-cases

Changed

• Changed default pad_end value to False.

Fixed

• Continual interface and conversion to support both class and module.
• Replicate padding in co._ConvNd

Remove modules names in co.Residual

Changed

• co.Residual modules to be unnamed. This allows the module state dicts to be flattened.

Improved state interoperability with complex 3rd-party nets

Added

• Flattened state dict export and loading via a flatten argument. This feature improves interoperability complex modules, that were not originally constructed with the co.Sequential and co.Parallel building blocks.
• Context manager for triggering flattened state_dict export and loading.

_PoolNd attribute changes and Parallel improvements

Added

• Support for zero-delay in co.Delay
• Support for broadcasting in co.Parallel
• Mul (hadamark product) aggregation in co.Parallel
• Example of Squeeze and Excitation block

Changed

• co._PoolNd attribute naming: "temporal_" removed as prefix for kernel_size, stride, dilation, and padding.

Improved handling of padding, strides and initialisation

Added

• co.Delay handling for padding.
• Handling of initialisation and strides in containers

Changed

• co.Conv build_from behavior to not change dilation and stride. Argument overload supported instead.
• pad_start and pad_end args to convolution and pooling modules forward_steps.
• Behavior of modules while they initialise. Now, a TensorPlaceholder is passed for initialising steps.

Removed

• Automatic unsqueeze in pooling.