Transfer learning without imresize or imageDatastore

For many samples of small images, it would be nice to load the data into RAM and perform transfer learning without using images on slower memory (imageDatastore). Unfortunately, if I had 50x50x1 images and had to resize to go with alexnet, I will be forced to use (50x50x1/227x227x3) ~ 1.6% of the number of samples in order to keep everything in RAM.

Does anyone know a fix? A custom layer that resizes would work, that's a lot of work though.

using 2017b

Answer 1

[Shounak Mitra]

Hi Michael,

Thanks for your question.

If you need to resize, apply augmentedimagedatastore to your image datastore - it will be much faster than a custom readfcn, because it preserves prefetch under the hood. If you don’t need to resize, augment, or any other need for a custom readfcn, then vanilla imds is simplest.

Another option is to create a custom layer but you're right, it'll take some work.

HTH Shounak

[Michael Benton]

https://www.mathworks.com/help/nnet/ref/augmentedimagedatastore.html according to this, that can be done. I have updated to 2018a, and this works, thanks