Also includes code to scale PWM values appropriately based on the freq. specified.

Fix PartProfiler showing wrong number for KerasLinear

Fix PartProfiler to take KeyboardInterrupts into account as well as initialisations of parts. It seems like on_part_finished() is not called when a KeyboardInterrupt occurs leaving the time value in the time difference which messes up the timing table max and avg values. Here, we simply chop off first and last entry form the timing arrays. We also add a median to the table to see if the distribution of timings is skewed.

Pass input shape args in the right order

Rahul actually figured this out - I'm just stealing his thunder.

https://docs.nvidia.com/deeplearning/sdk/tensorrt-api/python_api/parsers/Uff/pyUff.html

Document TensorRT Workflow.

PS3 controller is not compatible with Jetson Nano

Improvements to TensorRT Pipeline.

* Reverse throttle and steering.
* Reduce the number of nodes in the frozen graph further.

Test: Tested with donkey tubplot.

crop and normalize outside of model

Support for TensorRTLinear model.

changed Raspbian link from "Latest" to "Raspbian Stretch Lite"

Update setup_jetson_nano.md docs with the correct unversioned directory names

* This PR adds support for Tensor RT.
* There are probably more minor improvements I can make to
`tensorrt.py`, but I think this is a good starting point.
* `freeze_model.py` exports metadata about the model which helps us
reconstruct the TensorRT model.
* Make sure `pycuda.autoinit` is never loaded during training, as it
interferes with TF-GPU (when training).

Test: Tested this on my Jetson Nano.

Update supported_cars.md

Improvements to the ML Pipeline

Add support for freezing Keras Models. * Add freeze_model.py to convert to a frozen TF graph. * Use `convert-to-uff` to convert the frozen graph to UFF. * Profit.