Conv: Only constant tensors are supported

Many ONNX models that didn’t work with Barracuda are now successfully imported. Thank you for this! I haven’t checked yet, whether the inference outputs are correct, but this is a very good start.

Recently I tried to put this model into the Assets-folder. It produced the following error: “Conv: Only constant tensors are supported for W. Instead lambda_1/convolution:0.W is pointing to non constant node lambda_1/convolution__58:0”. Is this something you’re going to support, or is it out of the scope?

Previously, it didn’t work Barracuda as well, but I suppose you’d like to make Sentis as versatile as possible.

Thanks for providing the model
There is no real reason why we don’t support Conv layers with dynamic weights.
We’ll update it in the next release. In the meantime here is a workaround:
You can re-implement the Conv import to handle it if you want. ~Samples/CustomLayer
you override the conv operator in onnx and create a new layer in which you call the ops.Conv accordingly

1 Like

This should be resolved in Sentis 1.1.0, known internally at Unity as issue 14.

Please let us know how you get on.

1 Like

Yes, I think it works OK now. Thank you!

yay! great to hear