Summarize changes for upcoming 0.4.0 release.

Author: sharvil

CHANGELOG.md

@@ -1,5 +1,25 @@
 # ChangeLog
 
+## 0.4.0 (2020-04-13)
+### 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`.
+
 ## 0.3.0 (2020-03-09)
 ### Added
 - PyTorch support.