After much experimentation, VesselNet’s Keras model has been successfully converted into a Tensorflow-TensorRT optimized model via the following transformation:
- A custom function called freeze_session() that freezes the graph session of a trained TF model and saved to a ProtoBuf (.pb) format:
- Utilizing the Python API of TF-TRT, trt.TrtGraphConverter() method to convert the previous ProtoBuf format file into a TRT-Optimized frozen graph:
The coming week we’ll move towards uploading to AIAA server and run several tests on it with training data.
Commands to upload to AIAA Server:
Here are the samples of the testing datasets provided by the VesselNet’s author:
The VesselNet’s Keras model was successfully converted into a TRT-Optimized frozen graph that now can be uploaded to AIAA server to integrate with existing pipeline. Next, we’ll upload and run through the workflow with testing datasets.