Day20 : Image Augmenter

Image Augmentation From Scratch

The Albumentations.Compose() Trap

Here is how most tutorials teach data augmentation:

Three lines. Works immediately. Ships to production. So what is the problem?

The problem is that you have learned nothing about what happened to your pixels.
You do not know that HorizontalFlip is literally array[:, ::-1, :] — a single
NumPy slice. You do not know that Rotate requires mapping every output pixel back
through an inverse affine matrix to a source coordinate, then interpolating between
four neighboring source pixels to avoid jagged artifacts. You do not know that if
you skip that interpolation step, your rotated image will look like it was rendered
on a 1995 graphics card.

Why does this matter? Because when your augmentation pipeline introduces a subtle
shape mismatch that crashes training at batch 47,000 — not batch 1, batch 47,000
— you need to know exactly which transform corrupted the array dimensions. When your
model learns to recognize artifacts introduced by a poorly tuned augmentation
(this happens; it is called augmentation overfitting), you need to know which
knob to turn.

This lesson strips the abstraction away entirely. We implement all 8 augmentation
types as pure NumPy operations. The math is not hidden inside a C++ extension. It
is in front of you, in Python, on 10–20 lines each.

The Failure Mode

Here is what happens when you naively implement rotation without coordinate clamping:

Error message you will see: