🎯 Introduction

Machine learning is inherently experimental. Data scientists iterate through countless variations—different features, architectures, hyperparameters, and training approaches—before settling on a production model. But once that model is deployed, a critical question emerges: Can we recreate what we built?

🔄 Transform

⚠️ The Problem

Feature engineering happens in two places: training and serving. If these transformations differ even slightly, your model breaks in production.

✅ The Solution

Explicitly capture transformations as part of the model, ensuring identical execution in training and serving.

Three Approaches (Increasing Sophistication):

Approach 1: Transformations Inside Model Graph (Keras Layers)

For simple, instance-level transformations (operations on individual examples)

Benefits:

• ✅ Perfect consistency (same code in training and serving)
• ✅ Simple to implement
• ✅ No external dependencies

Limitations:

• ❌ Transformations run on every training iteration (inefficient)
• ❌ Only works for instance-level operations (can't compute dataset statistics)

Approach 2: Dataset-Level Transformations (tf.transform)