Sending large images time to the GPU

[ema...@umbc.edu]

I am using CNN model(AlexNet) for Image processing. and I am using Torch(lua). I am modifying the Torch starter code which is this -> https://github.com/cvondrick/torch-starter. My problem is that I am making images with 18 channels instead of 3 channels for training the model , and for sending those images to GPU, it takes around 20 (2.13 s for every batch) times more than when it sends images with three channels (0.14s for every batch ) . I also tried to see how long does it take to send images with 4 channels to GPU, and I saw that as soon as the number of channels increased to more than 3 channels, consequently the time increased a lot, like 20 times. For example, even for images with 4 channels , it took around 2s for every batch, which is around 19 times more than running 3 channel images. I was wondering if there is a bug which makes it take this much time and if there aren't any bugs, is there any way I can decrease this running time?

Thank you so much in advance!