There might be many ways to go about it:
1. Copy and stack the image matrix three times to replicate the 3 channels of RGB
2. Stack two zero or identity matrices on top of the grey image matrix such that the Red channel has all the data and the other two channels are just place holders (not advisable)
3. Remove the other channel weights in the pre-trained model itself (very cumbersome)
This is an open-ended problem with many other creative solutions like using DeOldify GAN to color them first or just interpolating the other two channels in some way. The way to find the best solution would be to experiment with your data and pre-trained architecture
