I am working lung CT images and I did a classification for these images using build from scratch model and the results are very good, now for a comparison purpose I want to do a transfer learning using one of the novel models like vgg, resnet, or alexnet, but the problem is these networks are trained on 3 channel images and CT scan images are only 1 channel, there is solution that say to repeat each image array 3 times to make it 3 channel.
my questions are:
1- Is this really the only solution for that?
2- Is it a good idea to do that?
3- Is there any other solutions?
You don't need to transform your input directly. You could do this in your computation graph such that the input has 1 channel that is only subsequently concatenated with 2 equivalent inputs to form the target channel size.
Whether this is a good idea in terms of performance will only an experiment tell you.
Another, more tedious idea is to apply a convolutional layer with 3 feature maps on the input of channel size 1. This will also result in the input dimensions that your transfer net expects. You could then freeze the transfer net while you adjust the conv layer. Subsequently, you could retrain all layers at the same time.
- Badges
- Science topic
- Similar topics
- Data Mining
- Classification