Make a binary model

[julienmas]

Hi,

I'm trying to make a model using binary quantization. I tried to reproduce the parameters used in CNV-W1A1 layers in my model as much as possible and I use almost the same build flow, but I'm not able to get the nodes of my ONNX graph to use BINARY finn_datatype.
The first image is from the CNV-W1A1 in the cnv_end2end_example notebook, you can see that the 1st MultiThreshold node outdtype is BINARY. The second image is from my model (it's a U-Net btw), you can see that the 1st MultiThreshold node outdtype is INT8.

What am I supposed to do to build a binary network?

You can find here the .onnx of my streamlined model, and the .py describing my model: unet_streamlined.zip, unet_julien.zip

Also I'm getting this 2 warnings, I think it's related to my problem:

/finn/src/finn/custom_op/fpgadataflow/thresholding_batch.py:329: UserWarning: Setting 0-valued first threshold to 1 to avoid vivado_hls bug
/finn/src/finn/transformation/fpgadataflow/convert_to_hls_layers.py:64: UserWarning: Im2Col_2 : Input is not int. Can't infer ConvInpGen.

And further I get the error: AssertionError: cycle-free graph violated: partition depends on itself, but it's no surprise since I can't convert Im2Col nodes to HLS nodes.

Julien

[julienmas]

The activation layers must be quantized with 1 bit too for the nodes to use BINARY finn_datatype, e.g :
self.relu1 = QuantReLU(bit_width=1, return_quant_tensor=True)

However I now have the following error, which I don't understand how to resolve :
Exception("True binary (non-bipolar) inputs not yet supported")