If it is possible to use features with more than 3 channels as input?

[momopusheen]

Hi there, Thanks for your excellent work! I'm trying to train DINO on my own data, which are not regular RGB images but images with more channels (more than 7). I found that in torch, the input channel can be modifed by adding an extra conv layer, and then concat it with the original model. However, I didn't find examples of doing so in detrex (or detectron2). ;( So I wonder if I could do this with detrex. If possible, would you mind giving me an example?

Thanks a lot! XD

[HaoZhang534]

You can add a layer that maps 7 channels to 3 channels before entering the backbone. But I recommend modifying the input channel of the first layer of the backbone from 3 to 7.

[momopusheen]

Thanks for your advice! Btw, I'm using swin transformer as backbone. Would it work if i modify the default 'in_chans' of it when defining backbone? here: def __init__( self, pretrain_img_size=224, patch_size=4, in_chans=3, embed_dim=96,

[HaoZhang534]

@momopusheen Sure, you can change the number of input channels of the backbone.

[momopusheen]

Thank you sincerely!