the use of OctaveConv

[WW2401]

Great work. Thanks for your sharing. @chaipangpang When using OctaveConv, it just needs to add OctaveUpsampleLayer and there is no change for Conv? Where do you use OctaveUpsampleLayer? as the last layer of every residual block (after BN)?

[chaipangpang]

(1) Yeah, when the backbone network is resnet, OctaveConv only needs OctaveUpsampleLayer. OctaveConv does not change the way the convolution is done. Convolution will only differ depending on the backbone network. eg: MobileNet -> deepwiseConv . (2) OctaveUpsampleLayer is used when the low frequency feature map and the high frequency feature map are added( type: "Eltwise"). You can see the model_example.

[WW2401]

(1) Yeah, when the backbone network is resnet, OctaveConv only needs OctaveUpsampleLayer. OctaveConv does not change the way the convolution is done. Convolution will only differ depending on the backbone network. eg: MobileNet -> deepwiseConv . (2) OctaveUpsampleLayer is used when the low frequency feature map and the high frequency feature map are added( type: "Eltwise"). You can see the model_example.

Thanks a lot. I had seen the model_example and its structure( I think its structure is so complex T_T), and I still don't know how to use OctaveConv in resnet. Can you tell me more details of using OctaveUpsampleLayer? Do you mind telling me your email?@chaipangpang The following is the structure of resnet_octave. The following is part of resnet_basic