pfnet-research / chainer-trt Goto Github PK

TensorRT supports fp16 input/output. It is better to set nvinfer1::DataType::kHALF here when building fp16 mode engine.

It is necessary to expose getBindingDataType with this change because it is needed to know input tensor data type for inference engine.

Currently plugins have to be included in chainer-trt code base.
However it would be great if user can insert both dump function (python side) and plugin implementation from outside chainer-trt.

https://docs.nvidia.com/deeplearning/sdk/tensorrt-developer-guide/index.html#unary-layer
https://github.com/pfnet-research/chainer-trt/blob/master/src/model.cpp#L445

Current chainer-trt cannot handle layers with multiple outputs.

layer->setPrecision(nvinfer1::DataType::kINT8)

Probably hook would work.

Will there be a jetson tx2 support, by making this repository compatible with tensorrt 4.0?

I know that there will be jetpack module which will be declared for jetson tx2 and tensorrt 5.0 but for now I would like to continue with that.