Train on more than a single image

[LRpz]

Hi,

I was wondering if more than one image could be used when generating the training set for LodeSTAR.

Thank you very much!
L

[Baptiste Grimaud]

I think this might be what you're looking for ?

[BenjaminMidtvedt]

That's right! We recommend using a few crops for training where possible, though there are diminishing returns the more you add.

[LRpz]

Yes! In that case, I would call training_images rather than training_set in
generator = dt.models.lodestar.LodeSTARGenerator( training_set, 3, (group, equivariance), batch_size=8, min_data_size=200, max_data_size=201 )
?

Thanks a lot!

[BenjaminMidtvedt]

An example would be

training_images = [crop1, crop2, crop3, ...]
random_sampler = dt.Value(lambda: random.choice(training_images))
generator = dt.models.lodestar.LodeSTARGenerator(random_sampler, 3, ...)