The update from PyTorch 1.7/ONNX v6/ai.onnx v9 to PyTorch 1.13.1/ONNX v7/ai.onnx v14 causes different finn-onnx export behavior in Brevitas (tested 0.8.0 and 0.9.1):
Primary input tensor name is not always "0", this was previously assumed
-> Workaround by providing "input_names" explicitly to export function
In certain cases the exported graph will no longer exhibit unique parameter tensors
-> Workaround by calling GiveUniqueParameterTensors()
Linting: as of now, the following reformattings would take place:
Minor automatic changes (e.g., addition/removal of empty lines) to 22 files
Line length limit (currently 88 characters) was previously not always enforced for multi-line strings ("""). After this update, a few occurences are slightly over the limit (13 lines in 11 files)
Combining this with the update to 2023.1 tools by merging into feature/2023_1. Some issues, such as updating qonnx and fixing the linting, remain and will be fixed in that branch before merging to dev.
Updates the base Docker image to Ubuntu 22.04, where Python 3.10 is the default. Some dependencies are updated alongside.
Fixes https://github.com/Xilinx/finn/issues/766.
Tested so far:
Issues:
pip install --user--editable
is broken for certain packages (here: qonnx)Depends on: