ADLINK Technology EVA SDK, User Manual, Page: 115 / 117

EVA SDK User's Manual
Convert for TensorRT Model
111
8.3.3.2 FP16
Linux:
/usr/src/tensorrt/bin/trtexec --uff=sample_ssd_relu6.uff --output=NMS --
uffInput=Input,3,300,300 --workspace=3000 --saveEngine=ssdv2_fp16.engine --
verbose --maxBatch=4 --fp16
Windows:
Trtexec.exe --uff=sample_ssd_relu6.uff --output=NMS --uffInput=Input,3,300,300 -
-workspace=3000 --saveEngine=ssdv2_fp16.engine --verbose --maxBatch=4 --fp16
8.3.3.3 Parameters
Parameter Description
--uff Load uff file path
--uffInput=Input,3,300,300 Set input layer name and input layer dimension
--output=NMS Set output layer name
--workspace=3000 Set GPU memory in MB to allow TensorRT to use to build engine.
Optimal setting is around 80% of all memory.
--maxBatch=4 Build engine will allow maximum input batch size = 4
--saveEngine=mobilenet_ssdv2.engine Save the final engine to ssdv2.engine file
--fp16 Set model to fp16 precision
--verbose Display more information
--buildOnly Reduce 10s for testing model, and only build and save engine.
(Optional)