Model deployed on board gives inconsistent output

[kai-0430]

Hello,

I'm trying to deploy a simple CNN onto PYNQ. I followed cnv_end2end_example and tfc_end2end example with some slightly changes, and successfully built the HW and deployed it onto the board. However, the model on board gives wrong answer when I feed in the input with execute() defined in driver_base. OK, fine, but the weird thing is that when I repeatedly feed in the same input, I get different output. I can't figure out why that happened.

Additional description: Since it's a regression task, I didn't insert TopK node. I know finn doesn't support floating-point operation now. So there are two nodes, Add node and Mul node, that are not able to be synthesized, at the end of the graph, and the synthesized model has a 24INT output. My way to deal with it is storing those values and performing the operations by the host on the board.

I've been studied FINN for my project for weeks, but keep encountering obstacles. The due date is coming. I hope I can deploy the model successfully soon. Any suggestion would be appreciated. If any files or codes needs to be attached, please tell me.

[kai-0430]

After endless try and error, I solved it. The problem comes from that I used conv with stride > 1. After I removed it and used pooling instead, all things worked! FINN doesn't supprot that, seemingly. To prevent more people suffering such unpleasantness, please have a check of that, or describe it clearly.