I work with KERAS Conv1D and see, that TVM is supporting it as it is request. Apply tf2onnx one ends up with three layers [Tranpose Conv1D Transpose] and TVM can either try to find this pattern or simply layer including runtime and inter storage memory. Since both is ugly for the TVM development and a waste, I started a discussion within the tf2onnx and onnx communities about the requirement of such an attribute been added to the Conv operator “storage_order”. Would be no issue for TVM backend, but would require additional code in the other backends. So if you see a similar need, I suggest, that the commitity raises the issue or at least support the issue I raise. See: #4099 and #1865. CC: @MJKlaiber, @areusch
1 Like