Haste 0.5.0-rc0

v0.5.0-rc0

Bump version to 0.5.0-rc0 in preparation for PyPI release.

Haste 0.4.0

Added

• New layer normalized GRU layer (LayerNormGRU).
• New IndRNN layer.
• CPU support for all PyTorch layers.
• Support for building PyTorch API on Windows.
• Added state argument to PyTorch layers to specify initial state.
• Added weight transforms to TensorFlow API (see docs for details).
• Added get_weights method to extract weights from RNN layers (TensorFlow).
• Added to_native_weights and from_native_weights to PyTorch API for LSTM and GRU layers.
• Validation tests to check for correctness.

Changed

• Performance improvements to GRU layer.
• BREAKING CHANGE: PyTorch layers default to CPU instead of GPU.
• BREAKING CHANGE: h must not be transposed before passing it to gru::BackwardPass::Iterate.

Fixed

• Multi-GPU training with TensorFlow caused by invalid sharing of cublasHandle_t.

Haste 0.3.0

Added

• PyTorch support.
• New layer normalized LSTM layer (LayerNormLSTM).
• New fused layer normalization layer.

Fixed

• Occasional uninitialized memory use in TensorFlow LSTM implementation.

Haste 0.2.0

This release focuses on LSTM performance.

Added

• New time-fused API for LSTM (lstm::ForwardPass::Run, lstm::BackwardPass::Run).
• Benchmarking code to evaluate the performance of an implementation.

Changed

• Performance improvements to existing iterative LSTM API.
• BREAKING CHANGE: h must not be transposed before passing it to lstm::BackwardPass::Iterate.
• BREAKING CHANGE: dv does not need to be allocated and v must be passed instead to lstm::BackwardPass::Iterate.

Haste 0.1.0

Initial release.