Part Number: TDA4VM
Tool/software:
Hi TI team,
I'm working on converting the custom onnx model using edgeai-tools (09_01_07_00). During the conversion process, I've encountered some issues.
For debugging purposes I reduced model to contain only first two layers (Reshape->MatMul). When I compared results for 8-bit and 32-bit model, after first layer results match:
Then, after second layer (MatMul), the mismatch is obvious:
After using convert_matmul_to_conv_1x1s1 script which converts MatMul to suitable operations, results match:
My questions are the following:
- Does MatMul layer supports 3D inputs? In my model input is 1x5x48. Can you confirm that MatMul layer should work without batch size set to 1?
- I also tried to add Reshape to (4D) -> MatMul (4D) ->Reshape to (3D), but I also get an incorrect results. Is this approach correct?
- In case that 3D inputs are not supported is there a list of layers which explicitly should have 4D inputs?
- One of the limitations I have with convert_matmul_to_conv_1x1s1 is that the script can't work properly if MatMul has 2 inputs. Can you suggest how to solve this issue? Below is part of the model which contains two inputs MatMul:
Kind regards,
Nevena
