The NeoOnnx library lets you load third-party neural networks serialized in ONNX format.
#include <NeoOnnx/NeoOnnx.h>
NEOONNX_API void LoadFromOnnx( const char* fileName, NeoML::CDnn& dnn, CArray<const char*>& inputs,
CArray<const char*>& outputs, CMap<CString, CString>& metadata );
NEOONNX_API void LoadFromOnnx( const void* buffer, int bufferSize, NeoML::CDnn& dnn, CArray<const char*>& inputs,
CArray<const char*>& outputs, CMap<CString, CString>& metadata );Loads a network from a file or a buffer.
For each network input the dnn network will have a CSourceLayer with the same name. For each source layer a blob of the size specified in the ONNX model will be allocated. Also input names will be added to the inputs array. The inputs with initializers will be ignored and the initializer values will be loaded directly.
For each network output the dnn network will have a CSinkLayer with the same name. Also output names will be added to the outputs array.
Model's metadata_props will be written to the metadata.
Both inputs and outputs name pointers are attached to the names of the corresponding layers in dnn and will become invalid after changing name or deleting corresponding layer.
The library will be built automatically together with NeoML.
We use the ONNX opset version 9, supporting the main convolutional neural network operations, LSTM, and most activation functions.
See the methods that load ONNX models in Objective-C and Java interfaces.