Keras flow label tensors from directory

Question

I am using convolutional neural networks to predict vegetation growth. My input is a (n,51,51,1) terrain elevation tensor, and the label is a (n,51,51,1) vegetation tensor.

Since flow from directory uses foldernames as labels, this is a bit of a problem. My network is performing well, but having to have all the data in memory is a bit limiting. If anyone knows how to setup a flow from directory for this problem I would appreciate it. I'm using R as an interface to keras and tensorflow, but solutions in python are welcome too. Included the picture in case it wasn't clear what I'm doing. Thanks!

Answer 1

This is a complex problem you are trying to solve. Image creation is another can of worms than classification (which is what you are talking about)

You can check this article that talks more in depth about the generational networks.

Another way to think about it, is to have the last output layer with 51*51 hidden units and do regression. By this I mean to treat it as a regression problem where you do regression on each pixel individually.