Image Transformations

This page describes how to build image transformations for test time deployment of your exported models.

To maintain the expected performance of your model after exporting, it is vital that some image transformations used for validation be used again when deploying to the test environment. These transformations are serialized as a JSON file when you export your model.

json 
      {
  "__version__" "0.5.1",
  "transform": {
    "__class_fullname__": "albumenations.core.composition.Compose",
    "p": 1,
    "transforms": [
      {
        "__class_fullname__": "albumenations.augmentations.transforms.Resize",
        "always_apply": false,
        "p": 1,
        "height": 224,
        "width": 224,
        "interpolation": 1
      }
    ],
    "bbox_params": null,
    "keypoint_params": null,
    "additional_targets": {}
  }
}

Here we can see the "Resize" transformation was used, with the "height" and "width" parameters set to 224.

All of the transformations you will see in the transforms.json file can be used from the very nice "albumentations" library. The transformations can simply be loaded using the albumentations library. To reproduce the transform from the example above, it may look like:

python 
      from PIL import Image
import albumentations as A
import numpy as np

transforms = A.load('transforms.json')
image = np.array(Image.open('/some/image/file/path'))
image = transforms(image=image)['image']

# Now the image is preprocessed and ready to be accepted by the model

Boost model performance quickly with AI-powered labeling and 100% QA.

Learn more
Last modified