TensorRT for keras

Step to run

• Convert your model files (h5) into a frozen graph using the freeze graph code
• Run the optimize script to get a new graph with tensorrt operations

Steps to setup Tensorflow with TensorRT support (Ubuntu 16.04 LTS)

• Install necessary cuda and cudnn libararies. (I got success with cuda-9.0 and cudnn-7.0)
• Install TensorRT from official repos (Building from source works too)
• Build tensorflow from source and enable TensorRT support in configuration script.

Results

• Results available for a denesnet model trained on a custom dataset
• Inference not done in batch mode
• Time in seconds