fastmachinelearning
Op Registry upgrade
This PR makes it possible to add QONNX Ops from outside the brevitas/finn/qonnx repositories. It adds a function register_custom_domain .
This is an example of usage of this functionality in chisel4ml: - https://github.com/cs-jsi/chisel4ml/blob/auto_parameters/chisel4ml/preprocess/fft_torch_op.py - https://github.com/cs-jsi/chisel4ml/blob/auto_parameters/chisel4ml/preprocess/fft_qonnx_op.py
It defines a custom pytorch op, that then gets mapped to a (Q)ONNX op.
This PR should not affect other peoples code.
I just realized I forgot to add the example code where register_custom_domain function is used. The examples above define the ops. However, first the register_custom_domain is called https://github.com/cs-jsi/chisel4ml/blob/auto_parameters/chisel4ml/transform.py#L87, and the ops are imported into the "chisel4ml" namespace inside chise4ml/__init__.py (because QONNX equates the domain name with the python namespace).
There are now two PRs on this topic. The other being #204 . It also proposes more changes. However, from a quick glance, it seems that it removes the "brevitas exception". Which might break FINN code?
However, from a quick glance, it seems that it removes the "brevitas exception". Which might break FINN code?
The PR adds a "domain aliasing" system (Lines 36-39 in registry.py) to more generically address the issue of CustomOp domain redirection, with the "brevitas exception" aliased by default. This has worked with FINN in my testing, but please let me know if you encounter any issues.
@jurevreca12 the PR #204 has been tested against FINN previously, and it seems a bit more documented. Can you please review it and see if it solves what you were trying to address in this PR (#144) ? If it does, I suggest we go for #204 instead.
