START SIMPLE

IMPLEMENT MODEL

DEBUG

EVALUATE

IMPROVEME MODEL | DATA

Workflow

Start simple

First of all, Do you have?:

• [ ] Initial Test set

• [ ] A single metric to improve

• [ ] Target performance based on baseline

Choose a simple architecture

Use sensible defaults

Normalize inputs

Simplify the problem

Debug Strategy

Get your model to run ➡️ ****

Overfit a single batch ➡️

Compare to a known result

Build Unit Tests for:

1. Shape Mismatch: Undefined & Incorrect Shapes
2. Casting variables: Data not in float 32
3. Out of Memory Tasks: Big tensor, Too much data, Duplicates, etc.

# Recommended for debugging
import ipdb; ipdb.set_trace()

<aside> 💡 Lightweight implementation. Use of the shell components. Build complicated data pipelines later.

</aside>