Training with data that has more than 3 channels

[ne1114]

Hello I'm trying to train with data that are more than 3 channels (ex 6, 12).

After the training I noticed that the model hasn't learned sufficiently compared to 3 channel data. (The 3, 6, 12 channeled datasets have images of interest stacked channel wise. Which means its basically grayscale image stacked 3 times vs. 6 times. 12 times.)

I think it would be hard for the model fully capture the importance of channel wise just with L1 loss, as I stack more channels.

Would there be any suggestions other than changing the loss function? For example, increase the ngf, ndf, stack more network layers, etc please?

[junyanz]

I am not sure if changing hyper-parameters will help your case. Depending on the type of data you are working on, you might need a new reconstruction loss rather than the default L1 loss.

[tongfh1]

Hello, I would like to ask you how did you implement the training of more than 3 channels of data, and which part of the code was modified?