Tool/software:
Hi TI team,
I'm working on converting the custom onnx model using edgeai-tools (09_01_07_00). In the model, at some point, I have the following layers connected in this order: GlobalAveragePool -> Flatten -> 10xGemm layers (whose outputs are outputs from the model). The conversion fails, but not due to the error reported (the error is about missing biases of Gemm layers) but due to the inability to create outputs with according dimensions expected by the Gemm layers (instead of 1x1x1x1x1xN tensor it outputs 1x1x1xNx1x1 tensor). This is easily visible by looking at the TIDL bin file. The investigation of ONNX import tool source code shows that the Flatten layer is merged with the Pooling layer and the Pooling layer outputs a tensor with the same dimensions as the input one (Flatten output dimensions are ignored). The pooling layer will only readjust its output dimensions if there is only one Gemm/InnerProduct layer as its next layer.
The question is the mentioned scenario supported?
Kind regards,
Aleksandar.
