0

The VGG model accept a 3-channel RGB image as input, but my data are single gray images, any suggestions for how to utilize the weights in first conv layer of VGG model?

stop-cran
  • 4,229
  • 2
  • 30
  • 47
joe
  • 99
  • 1
  • 12

1 Answers1

1

So your filters in the first layer are of size w x h x 3 (for VGG, w=h=3), I'd suggest taking the mean over the last dimension and using them as w x h x 1. As you've probably figured out, you won't need to change any other weights.

This is inspired from a similar approach people have applied to initializing the VGG Imagenet model on flow images.

https://arxiv.org/pdf/1507.02159v1.pdf (Secion 2.2)

Amlan Kar
  • 34
  • 2