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

[BowenKwan]

Prerequisites

Please make sure to check off these prerequisites before submitting a bug report.

• Test that the bug appears on the current version of the dev-branch. Make sure to include the commit hash of the commit you checked out.
• [Y] Check that the issue hasn't already been reported, by checking the currently open issues.
• [Y] If there are steps to reproduce the problem, make sure to write them down below.
• [Y] If relevant, please include the ONNX files, which were created directly before and/or after the bug.

Quick summary

FINN build fails with error,
AssertionError: cycle-free graph violated: partition depends on itself

I have reduced the model to reshaping the 7x7 image to 49x1 vector, and go through a linear layer. The same error occurs.

From other post in the github, it seems like the reshape function is not transformed to HW, showing as brown when viewing the onnx file on netron.

I have tried different reshaping method (.view(), nn.Flatten(), torch.flatten() ), however none worked.

Code used are included below. I have the onnx file available however it cant be uploaded here.

Details

self.quant_inp = qnn.QuantIdentity(bit_width=q_data_widt, return_quant_tensor=True)
self.fc3 = qnn.QuantLinear(49, 1,bias=True, weight_bit_width=q_bit_width)
self.flatten = nn.Flatten()	
    def forward(self, x):
        x = self.quant_inp(x)
        x = x.view(x.size(0), -1)
        #x = torch.flatten(x, 1)
        #x = self.flatten(x)
        x = self.fc3(x)
        return x

Steps to Reproduce

cfg_estimates = build.DataflowBuildConfig(
    output_dir=estimates_output_dir,  # directory location
    mvau_wwidth_max=80,
    target_fps=1000000,
    synth_clk_period_ns=10.0,
    fpga_part="xc7z020clg400-1",
    steps=build_cfg.estimate_only_dataflow_steps,
    generate_outputs=[
        build_cfg.DataflowOutputType.ESTIMATE_REPORTS,
    ]
)

build.build_dataflow_cfg("./dataflow_build_dir/model.onnx", cfg_estimates)

Expected behavior

Finish building

Actual behavior

 .......BUILDING THE DATA FLOW CONFIGURATION

Building dataflow accelerator from ./dataflow_build_dir/model.onnx
Intermediate outputs will be generated in /tmp/finn_dev_pyk12
Final outputs will be generated in ./output_estimates_only
Build log is at ./output_estimates_only/build_dataflow.log
Running step: step_qonnx_to_finn [1/10]
Running step: step_tidy_up [2/10]
Running step: step_streamline [3/10]
Running step: step_convert_to_hw [4/10]
Running step: step_create_dataflow_partition [5/10]
Traceback (most recent call last):
  File "/home/pyk12/finn/src/finn/builder/build_dataflow.py", line 158, in build_dataflow_cfg
    model = transform_step(model, cfg)
  File "/home/pyk12/finn/src/finn/builder/build_dataflow_steps.py", line 372, in step_create_dataflow_partition
    parent_model = model.transform(
  File "/home/pyk12/finn/deps/qonnx/src/qonnx/core/modelwrapper.py", line 140, in transform
    (transformed_model, model_was_changed) = transformation.apply(transformed_model)
  File "/home/pyk12/finn/src/finn/transformation/fpgadataflow/create_dataflow_partition.py", line 80, in apply
    parent_model = model.transform(
  File "/home/pyk12/finn/deps/qonnx/src/qonnx/core/modelwrapper.py", line 140, in transform
    (transformed_model, model_was_changed) = transformation.apply(transformed_model)
  File "/home/pyk12/finn/deps/qonnx/src/qonnx/transformation/create_generic_partitions.py", line 119, in apply
    self.partitioning(node) != partition_id
AssertionError: cycle-free graph violated: partition depends on itself
> /home/pyk12/finn/deps/qonnx/src/qonnx/transformation/create_generic_partitions.py(119)apply()
-> self.partitioning(node) != partition_id

[auphelia]

Hi @BowenKwan,
Yes, you identified the issue correctly, the Reshape node doesn't get converted to a HW kernel in FINN. You could try this transformation: https://github.com/Xilinx/finn/blob/dev/src/finn/transformation/move_reshape.py#L9 but I would recommend running functional verification before and afterwards to make sure that this was a valid transform for your use case. You will need to define a custom step to do this, please have a look at the advanced builder notebook for details: https://github.com/Xilinx/finn/blob/dev/notebooks/advanced/4_advanced_builder_settings.ipynb

[BowenKwan]

Thank you very much for confirming!

As I have seen the different cnn models from the Finn example, they are using one of the above conversion between convolution layer and matrix multiplication. Does it mean they also used the custom step to generate the build? Is there any standard way to define the custom step for the reshape conversion used in the examples? Is there any plan to make the compilation of CNN a standard build instead of having to make a custom step?

Also would it be possible to build the convolution layer and linear layer separately with the stand build, and the combine the 2 implementations?

Thank you very much!

[auphelia]

Hi @BowenKwan, if FINN detects that there is a convolution it will try to run the mentioned transformation, see here: https://github.com/Xilinx/finn/blob/dev/src/finn/builder/build_dataflow_steps.py#L356
Since your network does not contain a convolution, FINN doesn't detect the pattern for which it would run that transform. To see how to define a custom step, you can use the notebook as an example or one of the finn-examples which have custom steps, like the RadioML example.

Since FINN aims to generate an optimized, custom accelerator tailored to your network, depending on the topology the flow may require extra compiler passes. We are working on improving this plugin system in the future, to make it easier to add specific transformations to your builder flow.