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

[jurevreca12]

Hey,
I am having trouble implementing a simple depthwise convolution layer using FINN. The whole model is literally just a single convolutional layer. However I get the cycle-free graph violated error in step_create_dataflow_partition. Upon searching through similar issues, I learned that this usually results from problems in the step_convert_to_hw. I inspected the intermediate onnx file generated step_conver_to_hw.onnx and I found that his indeed was the case; that is not all nodes were hardware nodes.

Original QONNX model:

step_convert_to_hw.onnx

I tried to do some digging and I noticed that InferQuantizedMatrixVectorActivation has this requirement on the input sparsity to be None:

This, I believe is preventing this transformation from converting the MatMul. This comes from the QONNX transformation LowerConvsToMatMul:

Is this some sort of restriction? How can I circumvent this issue?

[fpjentzsch]

Hi,

good question, this is one of the many features where we could use a lot more documentation.

Due to the different data/folding layout for depthwise convolutions, these have to be implemented using a VectorVectorActivationUnit (VVAU) instead of the MVAU. You can see here in the MobileNetV1 finn-example how this is inferred: https://github.com/Xilinx/finn-examples/blob/810a1e7759163ac313cb7739743da4184fb9135d/build/mobilenet-v1/custom_steps.py#L93

[jurevreca12]

Thanks. I used the custom steps you suggested and it worked in the first go! :-)