Cannot convert grouped convolution in ONNX?

[Nullkooland]

NN structure:

Log from convert_tool:

---- Tengine Convert Tool ---- 

Version     : v1.0, 14:33:02 Aug 31 2021
Status      : float32

----------onnx2tengine begin----------
Model op set is :13
Error: input elem num(148225) != reshaped elem num(4743200)
Tengine FATAL: Infer node(id: 12, op: Reshape) shape failed.
Convert model failed.

ONNX file: ncc.onnx.tar.gz

The Conv node is a grouped convolution with groups == 32, and the input tensor have 32 channels, so the output feature map's shape should be 1 x 32 x 385 x 385 == 4743200 elements. However, it seemed that the output tensor produced by tengine is actually in 1 x 1 x 385 x 385 == 148255 elements, so the subsequent Reshape Op failed because of elements number mismatch.

[cccxinli]

Hello, your model cannot be opened, could you check and re-upload it?

[cccxinli]

May fixed. Pull code and try again.

[Nullkooland]

The latest version convert tool is able to convert the model, but the feature map shape after convolution is still incorrect as I mentioned before. Seems that the groups attribute is ignored by the Conv node.

Converted tmfile structure:

Model files: (original onnx and converted tmfile) ncc.tar.gz

[Nullkooland]

I guess the issue is with the tengine runtime, not the convert tool or the shape check (I guess you shouldn't have removed the shape check feature as it expose such problems early. Since the converted model is still incorrect and would trigger runtime error later in deployment environment.)

[cccxinli]

You are right. I just found the weight of this conv, is a variable tensor rather than a constant tensor.

[Nullkooland]

You are right. I just found the weight of this conv, is a variable tensor rather than a constant tensor.

Yes, what this NN do is to compute the normalized cross correlation between two feature maps, so the weights of the Conv is reshaped from an input tensor.