YAB

Yet Another Boilerplate

YAB is a Python library that removes repetitive PyTorch boilerplate without taking away control.

It provides two primary workflows:

• yab.use() renders a template in memory and returns a handle to the generated implementation.
• yab.write() injects clean, editable PyTorch code directly into your Python script or Jupyter notebook.

Unlike frameworks that hide everything behind custom APIs, YAB generates real code that belongs to you.

---

Features

• Zero-boilerplate PyTorch workflows
• Editable generated source code
• Three abstraction levels
• Python API and CLI
• AST-based code injection
• Jupyter notebook support
• Lazy dataset loading
• Built-in fallback datasets
• Automatic formatting with Black and Ruff
• Versioned templates
• Extensible template registry
• Idempotent code generation

---

Installation

pip install yab-ml

---

Quick Start

Return a template handle

import yab

handle = yab.use(
    "tabular_classifier",
    type="full",
    input_features=4,
    num_classes=3,
)

trainer = handle.get_trainer()
trainer.fit()

Nothing is written to disk.

---

Generate boilerplate

import yab

yab.write(
    "tabular_classifier",
    type="full",
    input_features=4,
    num_classes=3,
)

Repeated calls update the previously generated YAB block instead of creating duplicates.

---

Lazy Dataset Loading

Generated templates do not construct datasets during import.

yab.use() renders the template, handle.get_trainer() constructs the trainer, and trainer.fit() loads either your own dataset via data_path or a built-in fallback dataset.

Supported built-in datasets include:

• Iris
• Wine
• MNIST
• CIFAR-10
• AG News
• IMDB

---

Abstraction Levels

Full

Returns a TemplateHandle exposing get_trainer().

handle = yab.use("tabular_classifier", type="full")
trainer = handle.get_trainer()
trainer.fit()

Partial

Returns (model, handle).

model, handle = yab.use("tabular_classifier", type="partial")
trainer = handle.get_trainer()
trainer.fit()

Raw

Returns only the generated model.

model = yab.use("tabular_classifier", type="raw")

---

CLI

yab list
yab types
yab use tabular_classifier --type full
yab write tabular_classifier --type partial
yab version

---

Templates

Implemented

• Tabular Classifier
• Image Classifier
• Text Classifier

Planned

• Autoencoder
• GAN
• Transformer Fine-Tuning

---

Testing

Coverage includes:

• Template rendering
• AST injection
• Notebook injection
• Duplicate detection
• Formatting
• End-to-end model training
• CLI behavior
• Built-in fallback datasets

---

Continuous Integration

Every push runs:

• Ruff
• Black
• Pytest

Tagged releases can be published to PyPI.

---

Philosophy

Most libraries replace boilerplate with another abstraction.

YAB replaces boilerplate with your own code.

You keep complete ownership of what is generated while avoiding hours of repetitive setup.

---

License

MIT