Questions about Compose and Data Augmentation

[pauljhin]

Hello!

In the spleen_segmentation_3d_visualization_basic.ipynb tutorial in one part it says: "Here we use several transforms to augment the dataset". The Compose is used with some transformations.

Is Compose for chaining transformations? Right?

My question is: Does it really increase the data?

It seems to me that it just applies the transformations but doesn't keep a copy of the data or copy with different transformations. I am wrong?

For example, during training I would like Compose output to look like this:

original_image
original_image_with_rotation
original_image_with_flip_and_noise

How can I do this? Is it possible to have different CacheDatasets with different transformations and merge into just one DataLoader?

[KumoLiu]

Hi @pauljhin, random transforms will generate different images randomly for each epoch so it can augment the dataset. If you want images with random rotation or random flip, you can use transforms such as RandFlip, RandRotate, RandGaussianSmooth.
Hope it can help you, thanks!

[pauljhin]

Hi @KumoLiu,

Now I understand, it wasn't clear that these transformations were made in each epoch.

One last doubt. Compose does the operations in the order they are in the list? Ex. Compose([FirstTransf -> SencondTransform]). Like a pipeline, the previous output is the input of the next transformation. Right?

[KumoLiu]

Hi @pauljhin, yes you are right!
https://github.com/Project-MONAI/MONAI/blob/663cfd0dc806a21ecc3517f4c61823f2581466ab/monai/transforms/compose.py#L39-L41

[pauljhin]

Ok Thank you very much @KumoLiu !