Blueprint Toolkit

A Blueprint Toolkit is a Python toolkit for developing and running Training Blueprints locally. This library provides abstractions and utilities for managing Training Runs, including configuration loading, model management, checkpointing, metrics tracking, and dataset handling.

Features

• RunContext: Manages the execution context of a Training Run with signal handling
• ConfigLoader: Loads and manages training configuration from files or memory
• ModelLoader: Loads models from local filesystem
• CheckpointManager: Saves and loads training checkpoints
• ProgressSaver: Tracks and saves training progress
• MetricSaver: Records and saves training metrics
• Dataset Utilities: Fetches and manages dataset Snapshots
• S3 Utilities: Uploads and downloads directories to/from S3

Components

RunContext

The RunContext class provides a unified interface for managing all aspects of a Training Run. It handles signal interruption and provides methods for loading configuration and models, as well as for managing checkpoints, metrics, and progress.

ConfigLoader

Abstract interface for loading configuration. Implementations include:

• FileConfigLoader: Loads configuration from yaml/json files
• MemoryConfigLoader: Loads configuration from in-memory dictionaries

ModelLoader

Abstract interface for loading models. Implementations include:

• LocalFileModelLoader: Loads models from local filesystem

CheckpointManager

Abstract interface for managing checkpoints. Implementations include:

• LocalFileCheckpointManager: Manages checkpoints on local filesystem

ProgressSaver and MetricSaver

Abstract interfaces for saving training progress and metrics. Implementations include:

• MemoryProgressSaver: Stores progress in memory
• MemoryMetricSaver: Stores metrics in memory

Example: Create a local run context

with local_run_context(
    run_id="my-training-run",
    config_path="config.yaml",
    model_path="model.pth",
    checkpoint_dir="checkpoints",
) as ctx:
    # Load configuration
    config = ctx.load_config()
    
    # Load model
    model_data = ctx.load_model()

    # your training loop implementation
    train(...)
    
    # Save metrics
    ctx.save_metric(name="loss", value=0.5, step=1)
    
    # Save progress
    ctx.save_progress(epoch=1, step=100)
    
    # Save checkpoint
    ctx.save_checkpoint(path="checkpoint.pth", metadata={"epoch": 1})