tensorrt_l2norm_helper

TensorRT plugin that addresses issue with two unsupported oprations within l2_normalize TensorFlow operation.

l2_normalize/Maximum: Unsupported binary op max with constant right
l2_normalize/Rsqrt: Unary not supported for other non-constant node

NOTE: as per NVIDIA (r)sqrt operation should be fixed in next release (was not fixed in TensorRT 5.0.6).

Usage

Prerequisites:

• TensorFlow freezed (and optimized) .pb graph.
• CUDA and TensorRT (tested with Jetpack 4.2 on Xavier)
• TensorFlow (tested with 1.13.1)
• cmake >= 3.8

Steps:

• Clone this repo to Nvidia device and put your .pb file as sample.pb.

• Edit USER DEFINED VALUES values in step01_pb_to_uff.py and step02_uff_to_plan.cpp.

• Compile plugin and scripts

mkdir build
cd build
cmake ..
make
cd -

• Convert .pb to UFF ./step01_pb_to_uff.py. This should produce sample.uff
• Convert UFF to TRT plan ./build/step02_uff_to_plan. This should produce sample.engine

Example network

This plugin was tested with following Keras based network with patch.

Notes

• FP16 (half-precision) still not supported (TODO)
• op_type's: 0 - Maximum, 1 - Rsqrt, 2 - Sqrt