cycle-free graph violated - partition depends on itself

[lumawu]

Hey there, I am currently trying to go through the end2end_cnv toolflow with a network I modified on my own using brevitas.
However, during partitioning I run into the following error while creating a dataflow partition:

I used the example network on the pytorch tutorials as basis:

Here are the modifications I made to it, training was done with a bit width of 1 for weights and activation and an input width of 8:

the onnx model has the following topology right before the partitioning call:

Brevitas is on v0.6, Finn on v0.7, pytorch on v1.7.1
This is my first time using my own network and I couldn't find much info on the error on the net, so I hope I can find help here

[fpjentzsch]

Hi, the problem is that the MaxPool layer was not correctly streamlined and converted to an HLS node ("StreamingMaxPool_Batch").

Do you apply the MakeMaxPoolNHWC() and InferStreamingMaxPool() transformations during streamlining and HLS-conversion, respectively?

[fpjentzsch]

I just updated my previous message. I suggest you try calling MakeMaxPoolNHWC() again (on the model state shown in your previous screenshot), followed by AbsorbConsecutiveTransposes() and InferStreamingMaxPool().

If you are willing to alter your model, you could remove the final MaxPool or add a convolution layer, so that there is no final MaxPool before the fully-connected layers begin.

[caoimhelynch]

Okay so I tried calling MakeMaxPoolNHWC() then AbsorbConsecutiveTransposes() along with with InferStreamingMaxPool() so that all of the above are called more than once. It seems to have successfully steamlined the maxpool layer however I am still getting the exact same error.
Here is the streamlined model right before the line CreateDataflowPartition()



I was hoping the unstreamlined maxpool layer was the source of the error but now that this is fixed and I am still getting the same error I am confused. Is there anything else that could be causing this error to happen?
Sorry for so many questions I am new to finn and new to creating models in general so really appreciate the help you are providing!

[fpjentzsch]

The final nodes (Mul, Add, TopK) are not converted to HLS nodes yet. Try InferChannelwiseLinearLayer() for the Mul/Add and InferLabelSelectLayer() for the TopK, followed by GiveUniqueNodeNames() to clean up the model again.

[fpjentzsch]

Also, the MaxPool dimension problem persists for the 41x41 and 13x13 case, as your nodes were streamlined to MaxPoolNHWC, but not converted to StreamingMaxPool_Batch.

[caoimhelynch]

Ah okay wasn't sure what the mul add and Topk should look like currently attempting to convert those layers.

To fix this 41x41 and 13x13, is there a way to add padding that doesn't require adding changing the model? Trying to change that caused a lot of issues for me both in initially exporting to onnx and when the onnx model was being converted to HLS.

[KeciaHH]

Hi @fpjentzsch, hope you are still answering this issue. I have met a similar problem, this is what my model looks like:

I used the default build_dataflow_steps script in the src folder and commented second Streamline() in the step_streamline() function (cuz it may cause some errors about InferShape).

[fpjentzsch]

Hi @KokyoK,

unfortunately the default build_dataflow steps will not work for a model with residual connections. Maybe you can take inspiration from the ResNet-50 finn-example on how to adjust the streamlining and convert_to_hls steps accordingly. You can find the code here:

https://github.com/Xilinx/finn-examples/blob/main/build/resnet50/build.py
https://github.com/Xilinx/finn-examples/blob/main/build/resnet50/custom_steps.py

[pkdeep]

Hi @fpjentzsch ,
Can you help with this. I am also trying a network with residual connections, and facing the same issue. Few layer have not been converted to hls layer, hence getting cycle-free graph violated error.

Thanks

[fpjentzsch]

Hi, unfortunately I currently do not have the time to dive into your specific model. Maybe this recent thread can help you? Xilinx/finn-examples#67

[sansi-zhang]

Hi @fpjentzsch,
I ran into the same problem and I tried the same thing with ResNet-50 but it still failed.

In detail

I ran into a problem while executing model.transform(CreateDataflowPartition()).

I made some adjustments and tests (see #860 (comment) for details), but the error still occurred.

Then I looked at the answers in this issue, and I used the ResNet-50 process to re-tune my code, but I was still stuck.

Part of the onnx file when resnet-50 is not used, see #860 (comment)

Part of the onnx file when using resnet-50, as follows.

I guessed that it was due to the concat operation, so I stopped using concat and split all the output, but still had the same problem.

I hope you can help me solve this problem.

Thanks again.

[pkdeep]

Did anyone get through this problem successfully? I am stuck with the similar problem on a fully connected layer. Even tfc_2w_2a is not working for me in bnn_pynq example notebook.