Data augmentation in training deep learning (ArcGIS Pro 3.2)

[chiF]

Hello,

I'd like to use data augmentation in the 'Train Deep Learning Model' function. I use python and ArcGIS Pro 3.2. Unfortuntately, I cannot figure out how to do so. The documentation says 'The image will be randomly rotated (in degrees) by a probability (p). If degrees is a range (a,b), a value will be uniformly assigned from a to b. The default value is 30.0; 0.5.'. What does this mean? Does it mean that when given (20,80);1 the algorithm will rotate the image between 20 and 80 degrees with a probability of 1? I find this formulation potentially misleading.

What does "from a (20) to b (80)" and "uniformly" mean here? Also, the model report does not mention how the augmentation was finally performed, so I can only hardly test this in an experiment.

Thank you so much!

[PavanYadav]

Hi @chiF 

For data augmentation, each training image might be under a certain transformation before being used for training. Each image has a probability p to be rotated. If it would be rotated, the angle will be a value between a and b. This value will be randomly picked between a and b. Uniformly means evenly, every value in that range has the same probability to be picked.

(20,80):1 means, every image will be rotated for certain, and the rotation value is between 20 and 80. It will be picked with equal probability among all possible values between 20 and 80.

For more information about the various transformations, see vision transforms on the fastai website.

We currently support the following:

• Default—The default data augmentation methods and values will be used.The default data augmentation methods included are crop, dihedral_affine, brightness, contrast, and zoom. These default values typically work well for satellite imagery.
• None—No data augmentation will be used.
• Custom—Data augmentation values will be specified using the Augmentation Parameters parameter.
• File—Fastai transforms for data augmentation of training and validation datasets will be specified using the transforms.json file, which is in the same folder as the training data. For more information about the various transformations, see vision transforms on the fastai website. An example is provided here.

If you would like to see effect of just one method, maybe you train a model with ‘None’ and next train another model using your desired method and see if the method is helpful.