Continual Inference Versions Save

Fixed

• co.Conv onnx export for kernel_size=1.

Added

• Ability to access onnx from root, i.e. co.onnx.

Added

• ONNX compatibility to all library modules!
• co.onnx.export function mirroring torch.onnx.export.
• purely functional _forward_step function to all modules.
• _state_shape and _dynamic_state_inds properties to modules.
• Add about info to package root.

Changed

• Change call_mode internally from enum to tensor.
• Change internal state_indexes to tensors.
• Change stride to tuple.
• Change padding to tuple.

Fixed

• Fix assertion bug in co.Lambda.

Removed

• TensorPlaceholder in favour of None.

Added

• Missing part on Continual Transformers in README.

Removed

• Conv cpp impl.

Added

• Citations for Continual Inference lib paper.
• Docs.
• Automatic conversion for RNN modules.
• Continual Transformer modules, including:
  • RecyclingPositionalEncoding
  • RetroactiveMultiheadAttention
  • SingleOutputMultiheadAttention
  • SingleOutputTransformerEncoderLayer
  • RetroactiveTransformerEncoderLayer
  • TransformerEncoderLayerFactory
  • TransformerEncoder

Added

• "lagging" option for shrink in co.Delay and co.Residual.
• co.RNN.
• co.LSTM.
• co.GRU.

Changed

• phantom_padding renamed to residual_shrink.

Fix missing cpp in package

Added

• CoConv step impl in C++.

Fixed

• FLOPs module registration compatibility with ptflops >=v0.6.8.