VGG RadioML Build doesn't work

[moviaso1]

Hello,

I have a question/problem regarding the FINN+ build for a VGG RadioML model. When I try to build the accelerator, I receive an error message at "step_create_dataflow_partition":

I build the model using a example I know should work and I also included the necessary custom steps (conv1d_steps.step_pre_streamline, conv1d_steps.step_convert_final_layers) for the build. This is how the onnx-graph of the intermediate model for "step_convert_final_layers" looks like, which is the step right before the dataflow partition:

I don't really know what the problem is and would be very thankful if someone could help me with this. Thank you in advance for the response.

[fpjentzsch]

This looks like this issue: #101
But the fix doesn't work here, possibly because of the special case of 1D convolutions.

For a quick resolution, I would recommend to export the model without the input quantizer (which is mapped to the first Thresholding layer), or to remove the first Thresholding layer in a custom step.

[moviaso1]

Thank you, implementing a custom step that removes the Thresholding layer solved it.