DeepFix Documentation

DeepFix is an auto-assessment tool that automatically diagnoses common bugs in machine learning and provides a prioritized list of solutions backed by industry/research best practices. It integrates directly into ML workflows.

Features

• Automatic Bug Detection: Identifies common ML issues automatically
• Prioritized Solutions: Get ranked suggestions based on best practices
• Workflow Integration: Seamlessly works with PyTorch Lightning and MLflow
• Research-Backed: Solutions are grounded in industry standards and research
• Multi-Artifact Analysis: Analyzes datasets, model checkpoints, training logs, and deepchecks reports
• Agentic Reasoning: Uses specialized AI agents for comprehensive diagnostics

Quick Start

Get started with DeepFix in minutes:

1. Install DeepFix

# Clone the repository
git clone https://github.com/delcaux-labs/deepfix.git
cd deepfix

# Install with uv (recommended)
uv venv --python 3.11
uv pip install -e .

See the Installation Guide for detailed instructions.

2. Start the Server

uv run deepfix-server launch -e deepfix-server/.env -port 8844 -host 127.0.0.1

3. Diagnose Your Dataset

from deepfix_sdk.client import DeepFixClient
from deepfix_sdk.zoo.datasets.foodwaste import load_train_and_val_datasets
from deepfix_sdk.data.datasets import ImageClassificationDataset

client = DeepFixClient(api_url="http://localhost:8844", timeout=120)

# Load and ingest dataset
dataset_name = "cafetaria-foodwaste"
train_data, val_data = load_train_and_val_datasets(
    image_size=448,
    batch_size=8,
    num_workers=4,
    pin_memory=False
)
train_data = ImageClassificationDataset(dataset_name=dataset_name, dataset=train_data)
val_data = ImageClassificationDataset(dataset_name=dataset_name, dataset=val_data)

client.ingest(
    dataset_name=dataset_name,
    train_data=train_data,
    test_data=val_data,
    train_test_validation=True,
    data_integrity=True,
    batch_size=8,
    overwrite=False
)

# Diagnose dataset
result = client.diagnose_dataset(dataset_name=dataset_name)
print(result.to_text())

Check out the Quickstart Guide for more examples.

Documentation Sections

Getting Started

• Installation - Set up DeepFix on your system
• Quickstart - Get up and running quickly
• Configuration - Configure DeepFix for your needs

Guides

Learn how to use DeepFix for different use cases:

• Image Classification - Diagnose image classification datasets
• Tabular Data - Work with structured tabular data
• NLP Datasets - Analyze natural language processing datasets
• MLflow Integration - Integrate with MLflow for experiment tracking

API Reference

Complete API documentation:

• API Overview - Overview of the API structure
• SDK Reference - DeepFix SDK client API
• Server Reference - DeepFix server API
• Core Models - Core data models and types

Architecture

Understand how DeepFix works:

• Architecture Overview - High-level system architecture
• Client-Server Architecture - Client-server design and communication
• Agent System - How the agentic analysis system works

Deployment

Deploy DeepFix in your environment:

• Docker Deployment - Deploy using Docker
• Local Setup - Set up a local development environment

Contributing

Help improve DeepFix:

• Contributing Guidelines - How to contribute
• Development Guide - Development setup and practices

Key Concepts

Client-Server Architecture

DeepFix follows a client-server architecture where:

• Client (SDK): Handles artifact computation, recording, and workflow integration
• Server: Performs AI-powered analysis using specialized agents
• Communication: Clean REST API with well-defined contracts

Learn more in the Architecture section.

Multi-Agent Analysis

DeepFix uses specialized AI agents to analyze different aspects of your ML pipeline:

• Dataset Analyzer: Examines dataset statistics, class balance, and anomalies
• Deepchecks Analyzer: Reviews data quality, drift, and integrity checks
• Model Checkpoint Analyzer: Validates checkpoint integrity and deployment readiness
• Training Analyzer: Analyzes training dynamics and metrics
• Cross-Artifact Reasoner: Synthesizes insights across all artifacts

Learn more in the Agent System section.

Support

• GitHub Issues: Report Issues
• Email: Contact us at fadel.seydou@delcaux.com

License

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.

---

Built with ❤️ for the ML community