a basic question about custom layers for Tensorflow UFF model

AI & Data Science Deep Learning (Training & Inference) TensorRT
1

Hi, I’d like to ask a basic question about custom layers for Tensorflow UFF model. Let’s say a network has a part not supported by TensorRT, and I am going to replace the part with a custom layer.

To my understanding, when using custom layers for Caffe,
Caffe is supposed to give you a “model file” and “deploy.txt”, and you can manually touch the model definition (i.e., deploy.txt) so that it includes proper custom layer definitions, and you can provide the custom layers’ implementation with C++ API.

However, I am not sure how this can work with Tensorflow.
Tensorflow conversion requires “Tensorflow model → frozen binary .pb graph → binary-ish UFF format”. In this procedure, intermediate files are all binary. How can I manually modify the part and replace?

2

Look for sampleUFFSSD in the samples folder for TRT 4.0.1.6

3

@dhingratul, thanks for your answer! I didn’t know the new version 4.0.1.6. It’s helpful but, could you also take a look at a follow-up question?

4

Look at the sample, it has everything. It uses graphsurgeon to create standalone nodes you want to replace which you can define in your program.