ADLINK Technology EVA SDK, User Manual, Page: 111 / 117

EVA SDK User's Manual
Convert for TensorRT Model
107
8 Convert for TensorRT Model
Basic understanding of Deep Learning Neural networks is required.
It is very important that each GPU architecture can only use the converted engine of the corresponding Compute
Capability (CC) (
https://en.wikipedia.org/wiki/CUDA ). For example, the TX2 CC is 6.2, so an engine converted on
TX2 can only run on other devices with CC 6.2, they cannot run on another PASCAL GPU like the P1000 (CC 6.1),
or Tesla P100 (CC 6.0).
Download and install the TensorRT command-line wrapper tool “trtexec” from:
https://github.com/NVIDIA/TensorRT/tree/master/samples/opensource/trtexec
If TensorRT is installed, trtexec is prebuilt and installed in the following path.
On Linux: /usr/src/tensorrt/bin/trtexec
On Windows: [TensorRT install Path]/bin/trtexec.exe
8.1 Build TensorRT Engine Parameters
TensorRT supports the ONNX, Caffe, and UFF models.
8.1.1 ONNX Model
Parameter Description
--onnx=<file> Path to ONNX model
--output=<name>[,<name>]* Output layer name, ONNX not required
8.1.2 Caffe Model
Parameter Description
--model=<file> Path to Caffe weight
--deploy=<file> Path to Caffe prototxt
--output=<name>[,<name>]* Output layer name (required)
8.1.3 UFF Model
Parameter Description
--uff=<file> Path to UFF model file
--uffInput=<name>,X,Y,Z Input layer and dimension (required)
--output=<name>[,<name>]* Output layer name (required)
--uffNHWC If specified, the input X,Y,Z will be used as H,W,C.
Default is C,H,W