User stories feedback: data

[pcaversaccio]

• As a data engineer/data science engineer, why do I need augmentations for my problem/ML task?
• As a data science team, I want to know how augmentation helps solve my problem, also by how much?
• As a returning user, I want to be able to know which images would give me the best lift in performance? What performance metrics? Is there a way to do A/B testing?
• As a data science team, how do I integrate this processing into my existing pipeline? Is there a python package?
• As a data engineer, I would like to be able to integrate this to my pipeline engine/data lake/A-B testing frameworks automatically (e.g Airflow, Kubeflow, google cloud storage, aws S3, azure blob storage)

[pcaversaccio]

@giaoNguyen70 we need to think about answers & potential solutions to these questions.

[AI-Daita]

Augmentation, why we need?

1. Reducing the cost of data collection (and/or labeling)
2. Enables rare data prediction
3. Prevents privacy/sercurity problems of data
4. Improving model prediction accuracy by:
   • adding more data into model
   • creating variability in existing data
   • increasing generalization ability of the models
   • helping resolve class imbalance issues of supervised learning
   • reducing data overfitting

[AI-Daita]

For the question of "which is the best augmentation result for user's model performance? intergate into user's pipeline and A/B testing?", I think they related to the probloem of " augementation results tuning".

Our output can be downloaded and directly added/interagted into some python libaries (ex: Ray Tune for optimization,...). Ideally, our data + their models => optimised hyperparameters from different try.