More Op and bugfix
Pre-release
Pre-release
Features:
- Rewrite GRU Op for Onnx
- Support deconvolution for OpenCL Buffer mode
Bugfix / Refractor:
- Move NN from MNN_Express to MNN_Train
- Fix bug for LSTM model crash
- Add HardSwish Fuse
- Add Onnx::Not, Onnx::GatherElement
- Support TF::FusedBatchNormal when is_training = True
- Fix compute bug for Pooling3D
- Fix bug for CUDA backend convolution depthwise with multi-input crash