Open iksnagreb opened 3 months ago
Hm, while I can successfully run the mobilenet and the cybersecurity end2end tests, the bnn-pynq tests fail immediately at the export step, this seems not to be my fault (at least it is unrelated to this PR, maybe I am missing something else or messed up some dependency?):
tests/end2end/test_end2end_bnn_pynq.py, line 459
def test_export(self, topology, wbits, abits, board):
E fixture 'topology' not found
Any hints on this error?
Regarding building the finn-examples: It seems these have not been updated to the recent finn release yet? Will this happen in the near future or should we skip this test for now?
Hi @iksnagreb,
Yes, we restructured the bnn pynq tests and they are not callable with pytest -k bnn_pynq
anymore. You could run one of the markers, e.g pytest -m sanity_bnn
.
If you like, I can take it from here and do the testing (apart from your transformer use case). We are currently in the process of updating the finn-examples to the new flow.
Ok, then I will skip doing the end2end and finn-examples tests now and focus on our transformer model. My dummy model already built successfully, but I would like to test the "real" model as well, just to be sure. I will also mark this as ready for review now.
Thanks @iksnagreb !
See https://github.com/Xilinx/finn/issues/978 for some explanation of what it wrong with this transformation. See https://github.com/Xilinx/finn/issues/875 for more context.
RoundAndClipThresholds
transformation to correctly respect the range of the input data type (the quantization annotation of the input) and the container data type of the thresholds tensor. Clearly defines the type-casting/type-promotion behavior to always produce integer thresholds coded as float32 initializer tensors.RoundAndClipThresholds
transformation, testing rounding, clipping and different data type combinations for large bit-width integers around INT24 (where the representability issues start to occur).Closes #978