Changing the number of input channels for segmentation.

[npielawski]

Hello Hang Zhang!

Thanks for this nice repo!

I would like to try using DeepLabV3 with ResNeSt in a unconventional way: I am doing pixel regression (regression on images) and my input image only has one channel. I then have to retrain the full model.

I looked into the code, and there is no argument for the input channels, and the pretrained is set to True. I'd like to branch, but I am scared that I will be behind master after some time.

What can we do about this? Do you think it is going to work without using a pretrained backbone? If yes, would you be opened to a pull request adding the possibility to change the number of input channels with an optional argument (I would add the feature to all backbones)?

Thanks in advance!

Best, Nicolas

[zhanghang1989]

Hi,

The input channels are usually defined in the backbone. For example, ResNet's input channel is defined here https://github.com/zhanghang1989/PyTorch-Encoding/blob/master/encoding/models/backbone/resnet.py#L171. If you would like to change it to one channel, you may create your own backbone.

[npielawski]

Thanks for you input!

I will then create a local backbone then, copying the ResNest arch. I won't be lagging behind master, good idea!

[npielawski]

I looked into using DeepLabV3, but I can only use registered backbones [1], and BaseNet also forces pretrained=True [2] (and I don't have a pretrained model). I could bypass pretrained=True, but I am not too happy with that option.

What would you suggest? Should I also reimplement DeepLabV3 and BaseNet?

[1] https://github.com/zhanghang1989/PyTorch-Encoding/blob/ced288d6fa10d4780fa5205a2f239c84022e71a3/encoding/models/sseg/base.py#L25 [2] https://github.com/zhanghang1989/PyTorch-Encoding/blob/ced288d6fa10d4780fa5205a2f239c84022e71a3/encoding/models/sseg/base.py#L69

[zhanghang1989]

That is coding style not a technical issue. You may decide how you would like to implement it.

[npielawski]

Well, I would have preferred not branching the repo and modifying it if that makes sense.