[models] Add model compression utils

[fg-mindee]

Add a doctr.models.utils module to compress existing models and improve their latency / memory load for inference purposes on CPU. Some interesting leads to investigate:

• FP conversion (feat: Added TFLite, FP16 and quantization utilities #10)
• Quantization (feat: Added TFLite, FP16 and quantization utilities #10)
• Pruning (cf. https://www.tensorflow.org/model_optimization/guide/pruning/comprehensive_guide)
• TF Lite export (feat: Added TFLite, FP16 and quantization utilities #10)
• ONNX export (cf. https://github.com/onnx/keras-onnx & https://github.com/onnx/tensorflow-onnx)
• Export to SaveModel (docs: Added SavedModel export instructions in doc #246)

Optional: TensorRT export (cf. https://developer.nvidia.com/blog/speeding-up-deep-learning-inference-using-tensorflow-onnx-and-tensorrt/)

[fg-mindee]

ONNX conversion seems to be incompatible with TF 2.4.* as per onnx/keras-onnx#662. I tried on my end and encountered the same problem.
Moving this to the next release until this gets fixed!

[fg-mindee]

A good lead for ONNX support would be to use https://github.com/onnx/tensorflow-onnx (might have to create a savemodel to use it but it's worth a look)

[felixdittrich92]

@frgfm i think we can remove the tensorrt point If we support onnx wdyt ?

[frgfm]

Yes sure! We'll need to take a look at pruning at some point

[felixdittrich92]

yeah pruning is fine but tensorrt is a bit to much (should do the user on his own side if we can provide onnx this should be not so tricky)