flip-utils 0.5.0

About flip-utils

flip-utils is the pip-installable distribution published from this repository. Its Python import package is flip, which contains the shared platform logic used by FLIP jobs and services including core training logic, NVFLARE components, Flower helpers, and utility helpers.

The FLIP platform uses this package to power federated learning applications across multiple job types: standard federated training, distributed evaluation, diffusion model training, and custom federated optimization.

pip install flip-utils
# or with uv
uv add flip-utils

git clone https://github.com/londonaicentre/FLIP.git
cd FLIP/flip-utils
uv sync
# or
pip install .

uv build

Package Structure & Modules

The flip package is organized into logical modules:

flip.core

Core classes and abstractions:

• FLIPBase — Abstract base class with common FL logic

• FLIPStandardProd — Production implementation using FLIP platform APIs

• FLIPStandardDev — Development implementation using local CSV/filesystem

• FLIP() factory — Automatically selects the correct implementation based on environment

flip.constants

Configuration and enumerations:

• FlipConstants — Pydantic-settings configuration singleton

• ResourceType — Enum for imaging resource types (DICOM, NIFTI, etc.)

• ModelStatus — Enum for model training states

• JobType — Enum for supported FL job types

• PTConstants — PyTorch-specific constants and settings

flip.utils

Utility helpers:

• Utils — General utility functions

• model_weights_handling — Model weight aggregation and manipulation

flip.nvflare

NVFLARE-specific components:

• controllers/ — Workflow controllers (ScatterAndGather, CrossSiteModelEval, etc.)

• components/ — Event handlers, persistors, privacy filters, model locators, etc.

• recipes/ — High-level NVFLARE job recipes

• runtime.py — Runtime helpers for NVFLARE apps

• metrics.py — Metrics collection and reporting

flip.flower

Flower-specific helpers:

• strategy.py — Flower Strategy implementations (e.g. FedAvgWithClientMetrics)

• metrics.py — Server-side metrics collection and reporting for Flower runs

• progress.py — Progress/status reporting helpers for Flower runs

from flip import FLIP

# Uses FLIPStandardProd in production or FLIPStandardDev in local dev
flip = FLIP()
df = flip.get_dataframe(project_id, query)

Job Types

Pass the job type to the FLIP() factory (FLIP(job_type=...)). The JobType enum (flip.constants.job_types) defines the values recognised by FLIP():

The NVFLARE backend additionally ships a template directory under fl-apps/nvflare/ for each Client-API job type (standard, evaluation, diffusion_model, fed_opt); the template names match the JobType enum values above. The Flower backend ships its own standard and evaluation templates under fl-apps/flower/ — selected at the deploy layer by FL_BACKEND=flower.

User Application Requirements

The job components dynamically import user-provided code from the job’s custom/ directory. On the platform that directory is assembled by the FL API, which merges the uploaded app files onto the matching fl-apps/nvflare/<template>/app template; in local SimEnv runs the tutorial’s job.py stages its app_files/ into the job’s custom/ directly.

Development Mode

To test FL applications locally before deploying to production:

1. Set environment variables in .env.development:

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

2. Place your application files in the tutorial’s app_files/ directory (e.g. fl-tutorials/nvflare/image_classification/xray_classification/app_files/). On the platform they are merged onto the matching fl-apps/nvflare/<template>/app/ template at submit time.

3. Run one of the shipped tutorials against the NVFLARE simulator from the repository root:

   make -C fl-tutorials run-tutorial TUTORIAL=xray_classification
   # list every available tutorial with:
   make -C fl-tutorials list-tutorials

   Each tutorial’s make run delegates to make sim — its job.py driving a FLIP recipe on the NVFLARE simulator (SimEnv) from the flip-utils venv — configured per-tutorial via that tutorial’s .env.app.

Running Tests

Run unit tests for the flip package:

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

Tests use pytest with coverage reporting and are located in tests/unit/.

Building the Docs Locally

flip-utils is documented as part of the FLIP documentation. From the repository root, run:

cd docs && make docs

The generated HTML site will be written to docs/build/html. To clean previous builds:

cd docs && make clean

How the API Reference is Generated

The API reference is built with sphinx-autoapi and points directly at the flip/ source tree. That keeps the reference pages aligned with the code without maintaining hand-written module stubs. See the API Reference section of the built documentation for complete coverage of all public classes and functions.