Optimize model with non-tensor inputs.

In this example, we show how to optimize a model with non-tensor inputs. We differentiate two categories of non-tensor inputs:

• tuples, lists and dictionaries - those are fully supported by the Model Navigator, but might not be supported by the underlying technologies (e.g. torchscript).

• primitive types (integers, floats, strings and booleans) - those are currently only supported when they are consistent across all samples in the dataloader. For example, if the first sample is {"input_1": <some tensor>, "input_2": False} all other samples also must have input_2 set to False. This restriction will be lifted in the future.

In the optimize.py both categories are used.

We recommend running this example in NVIDIA NGC PyTorch container. To run the example, simply run the optimize.py script:

./optimize.py