flip-utils 0.1.3

FLIP - Federated Learning and Interoperability Platform library built on NVIDIA FLARE

FLIP-FL-base-FLARE

This repository contains the FLIP (Federated Learning and Interoperability Platform) federated learning base application utilities. It is a monorepo that includes:

• flip — pip-installable Python package with platform logic, NVFLARE components, and utilities
• tutorials/ — example applications you can run on the FLIP platform
• fl_services/ — Docker services for running FL networks (server, clients, admin API)

Table of Contents

• flip Python Package
  • Installation
  • Package Structure
  • User Application Requirements
  • Job Types
  • Development Mode
  • Unit Tests
• Tutorials
  • App / Tutorial Compatibility
• FL Services API
  • Prerequisites
  • Provisioning a Network
  • Running the Network
  • Integration Testing
  • CI/CD
  • Makefile Reference

---

flip Python Package

The flip package is the core pip-installable library for the FLIP federated learning platform. It provides all platform logic core to training and evaluating FL applications.

Installation

uv sync
# or
pip install .

To build a distributable package:

uv build

Package Structure

flip/
├── core/         # FLIPBase, FLIPStandardProd/Dev implementations, FLIP() factory
├── constants/    # FlipConstants (pydantic-settings), enums, PTConstants
├── utils/        # General utilities: Utils, model weight helpers
└── nvflare/      # NVFLARE-specific logic and components
    ├── executors/    # RUN_TRAINER, RUN_VALIDATOR, RUN_EVALUATOR wrappers
    ├── controllers/  # Workflow controllers (ScatterAndGather, CrossSiteModelEval, …)
    └── components/   # Event handlers, persistors, privacy filters, locators, …

The FLIP() factory selects FLIPStandardDev (local CSV/filesystem) or FLIPStandardProd (FLIP platform APIs) based on the LOCAL_DEV environment variable.

User Application Requirements

User-provided files go in the job's custom/ directory and are dynamically imported by the executor wrappers:

File	Description
trainer.py	Training logic — must export FLIP_TRAINER class
validator.py	Validation logic — must export FLIP_VALIDATOR class
models.py	Model definitions — must export get_model() function
config.json	Hyperparameters — must include LOCAL_ROUNDS and LEARNING_RATE
transforms.py	Data transforms (optional)

Job Types

Set via the JOB_TYPE environment variable:

Type	Description
standard	Federated training with FedAvg aggregation (default)
evaluation	Distributed model evaluation without training
diffusion_model	Two-stage training (VAE encoder + diffusion)
fed_opt	Custom federated optimization

The corresponding configs live in src/<job_type>/app/config/.

Development Mode

DEV mode lets you test FL applications locally before deploying to production.

1. Edit .env.development:

   LOCAL_DEV=true
   DEV_IMAGES_DIR=../data/accession-resources
   DEV_DATAFRAME=../data/sample_get_dataframe.csv
   JOB_TYPE=standard
   

2. Place your application files in src/<JOB_TYPE>/app/custom/.

3. Run the simulator in Docker:

   make run-container
   

Unit Tests

make unit-test
# or
uv run pytest -s -vv

---

Tutorials

The tutorials/ directory contains ready-to-use example applications that can be uploaded to the FLIP platform UI. Each tutorial is designed to work with a specific app type from src/.

App / Tutorial Compatibility

App	Tutorial
standard	image_segmentation/3d_spleen_segmentation
standard	image_classification/xray_classification
diffusion_model	image_synthesis/latent_diffusion_model
fed_opt	image_segmentation/3d_spleen_segmentation
evaluation	image_evaluation/3d_spleen_segmentation

---

FL Services API

The fl_services/ directory contains Docker-based NVFLARE services. See the FL services README and the FL API README for full details on provisioning and the API endpoints.

Prerequisites

• Docker and Docker Compose
• uv (Python package manager)
• AWS CLI configured (for downloading test data)

Provisioning a Network

Generate the certificates, keys, and configuration for a new FL network:

make nvflare-provision NET_NUMBER=1

This creates a network-specific compose file (deploy/compose-net-1.yml) and service secrets in workspace/net-1/services/ (gitignored). Multiple networks can be provisioned with different ports:

make nvflare-provision NET_NUMBER=2 FL_PORT=8004 ADMIN_PORT=8005

Warning: Provisioned files contain cryptographic signatures. Any modification will cause errors. Always re-run provisioning if changes are needed.

Running the Network

make build NET_NUMBER=1   # Build Docker images
make up NET_NUMBER=1      # Start the network (server, 2 clients, API)
make down NET_NUMBER=1    # Stop the network
make clean NET_NUMBER=1   # Remove containers and images

Integration Testing

Download test data (requires AWS S3 access) then run the relevant target:

make download-test-data
make test   # Run all integration tests

CI/CD

GitHub Actions workflows use OIDC to authenticate to AWS (no long-lived keys).

Trigger	Target
PR to any branch	s3://flipdev/base-application-dev/pull-requests/<PR_NUMBER>/src/
Merge to develop	s3://flipdev/base-application-dev/src/ and s3://flipstag/base-application/src/
Merge to main	s3://flipprod/base-application/src/

Warning: Never manually sync to the production bucket.

To test a PR on the FLIP platform, update FL_APP_BASE_BUCKET in the flip repo environment variables to point to your PR's S3 path.

Makefile Reference

Network Management

Command	Description
make nvflare-provision NET_NUMBER=X	Provision FL network X
make build NET_NUMBER=X	Build Docker images for network X
make up NET_NUMBER=X	Start FL network X
make down NET_NUMBER=X	Stop FL network X
make clean NET_NUMBER=X	Remove containers and images

Development

Command	Description
make run-container	Run NVFLARE simulator in Docker

Data Management

Command	Description
make download-test-data	Download all test data (x-ray and spleen images, model checkpoints) from S3

Testing

Command	Description
make unit-test	Run pytest unit tests for flip python package
make test-xrays-standard	Test standard job with x-ray data
make test-spleen-standard	Test standard job with spleen data
make test-spleen-evaluation	Test evaluation job with spleen data
make test-spleen-diffusion	Test diffusion model with spleen data
make test	Run all integration tests