How to Correctly Handle a Model with Concatenation Layer

[williamliao28]

Any suggestions are welcomed. I have tried the following experiments for this issue.

First, I have looked into the source code of the InferConcatLayer() function (line 1678 to 1737 in convert_to_hls_layers.py). I think this function is used to convert the Concat ONNX layer to the HLS layer StreamingConcat. I found that it already calls InferDataTypes() on line 1736. So is it necessary to call another InferDataTypes() again after calling InferConcatLayer()?

Second, I try to call the InferConcatLayer() function before InferBinaryMatrixVectorActivation() and InferQuantizedMatrixVectorActivation(). However, I get the following error message: AssertionError: Thresholds in MatrixVectorActivation_MatMul_4 can't be expressed with type INT10. The traceback message is included in the screenshot posted below.

Finally, I also included two example python codes in the links below which may be useful for reproducing the error. The file simple_test.py contains the PyTorch module defining the model with Brevitas quantization. The file simple_test_finn_synthesis.py is the code I run inside the FINN docker and this one produce the error message described above. The file simple_test_cifar10_8bit_2epoch.onnx is a pertained model which can be used directly as the input for simple_test_finn_synthesis.py.

simple_test.py

simple_test_finn_synthesis.py

simple_test_cifar10_8bit_2epoch.onnx

[auphelia]

Hi @williamliao28 ,

Thanks for providing additional material to reproduce! I think this is not a question that can be quickly answered.
So, could you please create an issue from this discussion with the provided template?

[williamliao28]

Hi @auphelia,

Thanks for the reply. I have created an issue (#853 ).

[auphelia]

The created issue (#853) is closed by this PR #857