search

fastmachinelearning / qonnx

QONNX: Arbitrary-Precision Quantized Neural Networks in ONNX
https://qonnx.readthedocs.io/
Apache License 2.0
124 stars 39 forks source link

Add more activations that a transpose can move through #148

Open jmitrevs opened 1 week ago

jmitrevs commented 1 week ago

The channels-last conversion transpose should be able to pass over activations that are applied element-wise. Therefore, I added more nodes that can be skipped.

There is a test in hls4ml that fails without this change since it uses a Selu activation. I am not sure if more tests are needed in the qonnx repository for this.

jmitrevs commented 1 week ago

PR #149 is also on moving transforms. It may be better to combined this change with that.