Context: I am using this neural network architecture developed by researchers called Dual Pyrimad Graph for a mesh segmentation task. The network expects the input to have 6 values: XYZ and RGB. I am training the network form 0 on my data as the network weights are not available, so there is no transfer learning.
Problem: My data lacks color only XYZ is available. I was trying to change the network layers architecture to match my input but realized that it is a tedious task as the guys who wrote it did that in Jittor not Pytorch or TensorFlow. I thought maybe I could put the colors all as 0 or all as gray 127 on the 3 color channels. I don't change the input add 1 color to it. I think if I do that learning wouldn't be affected as much. Granted I can get lower Accuracy compared to their colored data but I have no colors to begin with.
Question: How will this solution affect my training if compared to changing the network to suit the data? will learning become slower since I have more input features (though all of the same value)?
By setting your RGB data to a default/constant value you'll very likely experience a performance drop compared to the "original" solution as your data are less qualitative. And indeed your model will be slower than a model using only XYZ data, as it has more parameters.