Stitched-Up RNNPool Face Detection Model and Quantization Code Performance Refactor

[ShikharJ]

There's a lot to be done here:

• Additional utility functions and refactoring some existing activation functions such as TanH, ReLU6 and Sigmoid
• Adding support for variable demotion and usage of 8-bit and 16-bit quantized mixed operations
• Implementing full pipeline as a model
• Implementing quantized and floating-point numerical test-cases for the full pipeline
• Adding the stripped down pipeline as a model
• Adding complete documentation
• Porting runtime-optimized code
• Rewrite tests for runtime-optimized code and remove deprecated operators

[ShikharJ]

@harsha-simhadri This is ready for reviewing.

[harsha-simhadri]

@ShikharJ Is there a way to move the model params to a place outside github. I am concerned about bloat in the repo. How large are all the model files you are planning to check in?

[ShikharJ]

@harsha-simhadri The collective size of the model params is around 2.2 MBs (1.6 MBs for the larger model and 0.6 MBs for the smaller model). I'm not aware if a modular method exists, apart from not committing the weight files into the master.

[harsha-simhadri]

@ShikharJ I am able to checkout git lfs and compile, but unclear what are the exec files and their commandline args. can you please add some notes to README, and print out a help line when the execs are run?