ml-test-driven-development

What problem does it solve?

Prevents long, costly training runs caused by wiring bugs, detached gradients, or misconfigured loss/optimizer setups by providing a short, focused test sequence that validates model integration before any full training run.

Core Features & Use Cases

• Shape Gauntlet: Run a synthetic batch through the full forward pass to verify tensor shapes at module boundaries and catch dimension/broadcasting errors.
• Gradient Smoke Test: Ensure all model parameters receive non-zero gradients to detect detached tensors, frozen layers, or broken loss hookups.
• Overfit-One-Batch: Verify the model can memorize a single batch on a small config to confirm loss, label alignment, and optimizer steps are wired correctly.
• Use Cases: Implementing a new model architecture, integrating custom loss functions or attention layers, porting models between frameworks, and validating data pipeline labels.

Quick Start

Run the three checkpoints—shape gauntlet, gradient smoke test, and overfit-one-batch—on a dummy batch before starting any full training run.