AssertionError: cycle-free graph violated: partition depends on itself

[Surya-T44]

I have followed you tutorials for Python to FPGA, and created my own classifier model.onnx with 6 classes. In notebook 2, right before partitioning, it keeps giving the error 'AssertionError: cycle-free graph violated: partition depends on itself'. I tried other streamline options (absorb, move) none of them worked. i even tried to remove the MUL node manually, and later some other nodes as well to see if it works, it keeps giving the same error. Could you please help? I have been struggling for the past two months until i found your tutorials which were really helpful. I need to submit my project within 2 days, so will be really helpful if you could help out. Thanks again!

[0BAB1]

Hi @Surya-T44

It's been a few months since I did not use FINN so maybe they added new stuff ? I say that because you have lots of transpose nodes and I don't think I had any back when I did the simple MNIST example.

FINN mainly support simple tensor operations. Your model looks simple though. If I can give you a tip, try a very basic MNIST model first, or even just a plain matrix matrix multiplication to build confidence with the tool's end to end workflow. Once this is done, you'll haver a good overview of how if works. But your deadline is very tight and that would take a day.

Looking at your graph it lok very straightforward, there no obvious cycle in the graph (straight line...).

Unfortunately I won't be of much help here..

Hope you find a solution, I strongly suggest you expose this problem to the FINN's repo because I don't see any obvious problem with this graph

[Surya-T44]

Thank you for the quick response.
I have tried your procedure with a simple MLP architecture, and was able to partition, generate .bit and .hwh files as well. (although the accuracy was 40%~ and when executed on Pynq Z2 board, The output is giving a constant value, even though different inputs were given).
.
Any suggestions on the architecture?
I'm building a Hand Gesture Recognition System model, with 5 Gesture classes and 1 unknown class. On the board - input taken from usb cam, output on web browser using flask.

[0BAB1]

Okay,

40% accuracy on the board is good if you actually do hand gesture recognition with simple FC layers, if it's just MNIST test then it's not normal. The course has some ways to run data through the quantized onnx model and the actual RTL (if I recall correctly) before porting it to FPGA, making it able to compute expected FPGA performance metrics. Accuracies should be around the same, otherwise there is a problem somewhere.

Anyways, Now for the actual model for hand gestures, the first model your sent looked like the obvious solution here with a couple of CONV layers. and a classifier in the end.

=> By the way, now that I think about it, I see that your end classifier was condensed into a MVAU (which is good and expected), but your conv layers did not condense into anything (should have turned into SVU or something like that), meaning your problem should be earlier in your code as FINN was not able to recognize your CONV layers as such. Try to make your CONV layer as "standard" as possible if you try again, maybe the transpose are confusing FINN, preventing it from making it "hardware friendly" layers ? Keep the initial pytorch model extremely simple and standard, that could help... maybe.

Even try to experiment with different architecture (without training them to gain time) until you have a model that translates well in hardware, you can start training this model to get real weights and use these.

This whole FINN thing was a while ago, I'm sorry I can't give you precise insights. I do plan on making a better course this summer but I right now, the concepts are not as clear as it used to be for me.