This repo documnet how to convert Tensorflow / Keras model to TRT engine using ONNX.
Note: If tf2onnx didn't work on Keras model try to use keras2onnx

• Try keras2onnx on Facenet model and compare results
• Try tf2onnx with --fold_const option. which may benefit the optimization as stated in repo's README.md :

When set, TensorFlow fold_constants transformation is applied before conversion. This benefits features including Transpose optimization (e.g. Transpose operations introduced during tf-graph-to-onnx-graph conversion will be removed), and RNN unit conversion (for example LSTM). Older TensorFlow version might run into issues with this option depending on the model.

• Try to freeze the graph as using freeze_graph.py in Facenet repo.