When groups equal input channels then PyTorch Conv2D is mapped to Keras DepthwiseConv2D about nobuco HOT 5 CLOSED

Hi! Funnily enough, that was brought in by a recent fix: #30

It's unclear how to get convolution working in as many cases as possible with minimal tinkering. Need to mull it over.

from nobuco.

is_stride_ok = all([stride[0] == x for x in stride])

Seems like this approach is the best one, pushed a fix in v0.13.3.

from nobuco.

😆

I just wrote the above!

Amazing! Thank you so much for this.

from nobuco.

I read the issue and I understand the problem and the different options. Maybe adding one flag at the nobuco.pytorch_to_keras method, something like

dummy_image = torch.rand(size=(1, 3, 256, 256))
pytorch_module = MyModule().eval()

keras_model = nobuco.pytorch_to_keras(
    pytorch_module,
    args=[dummy_image], kwargs=None,
    inputs_channel_order=ChannelOrder.TENSORFLOW,
    outputs_channel_order=ChannelOrder.TENSORFLOW,
    prefer_depthwise=False,
)

Or? I can understand that you (most probably) have some design considerations already in-place for nobuco, so the above might be totally out of these considerations :D

But it would be nice to have the option to choose between using the depthwise implementation of Keras/TF over a regular grouped ConvXD, that can result in the issue mentioned by #30

from nobuco.

Or even the following can be done at nobuco/node_converters/convolution.py :

is_depthwise = groups == in_filters and is_stride_ok = all([stride[0] == x for x in stride])

from nobuco.