QKeras to QONNX converter redux

[jmduarte]

This is the PR for the QKeras to QONNX converter that has fixed some of the issues in the previous PR draft #7.
Quick Overview:

• Original strategy: (1) “Strip” QKeras model of quantization attributes and store in the dictionary; (2) convert (as if plain Keras model) using tf2onnx; (3) Insert “Quant” nodes at appropriate locations based on a dictionary of quantization attributes
• Works for fully-connected models; however stumbling blocks when applied to convolutional models (related to interference between how tf2onnx handles channels-last/channels-first transformations and inserted Quant nodes)

This branch addresses the problem with the convolution layers and provides new tests for checking the stability of the converter. Even with the current fixes, there are still some issues with the converter. Some have workarounds that are outlined below:

• Quantized-Relu inserts redundant quant node when used as output activation of Dense/Conv2D layer.
  Workaround: To use Quantized-Relu as separate QActivation layers.

• Quantized-Bits Node is not added when used in QActivation layers
  Workaround: Use Quantized bits only at the output of Dense/Conv2D layers

• Threshold of 0.5 must be used when using Ternary Quantization (This is sometimes unstable even with t=0.5)

[maltanar]

Could you also update this with the latest main? I took the fail-fast changes from this PR plus a pinned a few more versions to keep things building smoothly.

[selwyn96]

Could you also update this with the latest main? I took the fail-fast changes from this PR plus a pinned a few more versions to keep things building smoothly.

tf2onnx does not support NumPy 1.24.1 (NVIDIA/TensorRT#2557), can we revert the dependency to NumPy 1.23.5 ?
Edit: Looks like this was fixed in tf2onnx 1.12.1. Will update accordingly

[jmduarte]

Note, some additional diffs got introduced when upgrading pre-commit hooks (basically new black style), do we want to downgrade to previous black?

[maltanar]

LGTM!

[maltanar]

For making new users aware of any limitations, I would suggest to link the first post in this PR from README.md under the QKeras heading, unless you'd like to introduce a new section on the Sphinx docs? If this sounds good I can merge this PR and add that link to README.

[selwyn96]

Sounds good, I can add a new section in the documentation to include the limitations.

[maltanar]

Merged in its current state but @selwyn96 please feel free to open a new PR for new documentation.