1. Multiple opset support: Enables user to generate model based on a specified ONNX opset
2. ONNX opset 8: Adds support for ONNX opset 8 in all converters
3. ONNX model optimization: Reduces redundant operators like transpose and identity in the converted model
4. Convert channel_last to channel_first: Enables converted model to run on WinML even if the original model is channel_last, since the WinML API only supports channel_first models
5. onnxruntime as backend test: Uses onnxruntime as the backend to test the converted model; improves code quality and compatibility between converters and the inference engine
6. Separate LightGBM from scikit-learn converter: Improves code organization