Conv2D export bug with in_channels or out_channels > 1

What commit version of aidge do you use

• aidge_export_cpp: 9f86629b

Problem description

2D convolution export gives wrong results with in_channels>1

Reproducible example code

From test_export.py , the first test with on in_channel passes, the second with 2 in_channels fails.

    def test_Conv(self):
        model = aidge_core.sequential([
            aidge_core.Conv2D(1, 1, [3, 3], name="InputNode")
        ])
        initFiller(model)
        self.unit_test_export(model, "Conv", [[1, 1, 9, 9]]) # IS OK

or

def test_Conv_in_channels2(self):
        model = aidge_core.sequential([
            aidge_core.Conv2D(2, 2, [3, 3], name="InputNode")
        ])
        initFiller(model)
        self.unit_test_export(model, "Conv", [[1, 2, 9, 9]]) # ERROR !

This issue is likely due to an ongoing merge. You will probably have to wait until !51 (merged) is merged (as core has already been updated).

I Olivier, unless I did not get something, I'm afraid it could come from aidge_export_cpp/kernels/convolution.hpp implementation.

I just tested it again with last dev version of aidge_core/export_cpp/backend_cpu.

Try multiple times with random input, and change in_channels or out_channel to values > 1. Most of the time it fails.

Standalone test :

test_conv2d_export_kernel.py

There are no change to the convolution implementation in !51 (merged) so I don't get how it could be improved there.

I fixed a bug in weights/biases initialization :

test_conv2d_export_kernel.py

Same problem in conv2D export kernel

changed title from Conv2D export bug with in_channels > 1 to Conv2D export bug with in_channels or out_channels > 1

It's a memory manager issue. This will be fixed in aidge_core!448 (merged).

mentioned in merge request aidge_core!448 (merged)

added Bug 🐛 label

closed

reopened

closed

As you said, fixed in MR !51 (merged)