Shintarou Okada

read_image's alpha should be 'blend_with_white' as default

1

For back compatibility, `read_image` 's `alpha` parameter should be 'blend_with_white'. Current `None` fails classifier train example, because imagenet has images with alpha.

This PR adds TensorRT backend. It enables fast inference by using NVIDIA's TensorRT. - [x] Modify CMakeLists.txt for merging - [x] Add a license - [ ] Testing

fix install command for menoh_onnx and menoh_onnx_proto

1

Modification for #215

There is no example which shows how to call run() multiple times. So many users get confused that they have to construct the model for each run(). There is no...

This PR contains ONNXIFI backend. Backend loads external `libonnxifi-xxxx.so` Pytorch's glow is used for testing.

The official `libonnx` contains some useful functions: - dtype/shape inference - Fusing optimization We should utilize them.

see #145 We should make it taking dynamic parameters or checking the prameter static.

ONNX format has [metadata_props](https://github.com/onnx/onnx/blob/master/docs/IR.md#models) which can be used for notating pre-process and post-process of the model. `const char* get_metadata_prop(model_data_handle, char *key)` seems suitable.

Currently mkldnn backend reorders fixed parameters (e.g. weights of Conv) at runtime. It should be done once at build time.

Currently `attribute_completion_and_shape_inference.hpp` is generated at coding time but it should be gnerated at building time.