The different data classes in the deepchecks.vision module were designed to enable access to the data in a unified format, while making the formatting code easy and maintainable. They are responsible for converting the data (whether these are images, labels, or predictions) to a format deepchecks can work with.
Why Is This Important?#
There is no standard when it comes to data formats in computer vision. Each model can have its own supported data format for the output predictions, and each new benchmark dataset can propose a new labeling format. Even the images formats can be different from each other (PIL, OpenCV, etc.).
In order to run a model and data-agnostic comprehensive testing suites, the data must be in an accepted format the checks can work with. The data class objects provide a structured and repeatable way to do that, and are an important part of the vision module.
What Do You Need to Implement?#
Generally, all you need to do is to implement a data class that inherits from one of the supported classes in the vision module, and just implement there 3 functions:
batch_to_images: Returns a list of images in the correct format from a batch of data.
batch_to_labels: Returns a list of labels in the correct format from a batch of data.
infer_on_batch: Returns a list of a model’s predictions in the correct format from a batch of data.
Validating Your Data Class Implementation#
While implementing the data class, you may need a way to test your work is correct. For this purpose, deepchecks contains an helper function which tests your implementation and prints its outputs, which you can inspect to make sure everything works correctly. Use this function like so:
from deepchecks.vision.utils.validation import validate_extractors dataset = MyDataset(my_dataloader) validate_extractors(dataset, model)
For more info about the classes validation see the guide formatters_validation