Namaka Python SDK

Lightweight Python client for Namaka experiment tracking. Logs metrics in the background so it never blocks your training loop.

Installation

pip install namaka

Or for development:

cd sdk
pip install -e .

Quick Start

import namaka

run = namaka.init(
    project="my-project",
    api_url="http://localhost:8000/api/v1",
)

# Log hyperparameters
run["parameters"] = {
    "learning_rate": 0.001,
    "batch_size": 64,
    "model_type": "transformer",
}

# Log metrics during training
for epoch in range(100):
    loss = train_step()
    accuracy = evaluate()
    run.log("train/loss", loss, step=epoch)
    run.log("val/accuracy", accuracy, step=epoch)

run.stop()

Connecting to Render

run = namaka.init(
    project="my-project",
    api_url="https://namaka-backend.onrender.com/api/v1",
)

API Reference

namaka.init(project, api_url, name, api_token)

Creates a new run and returns a Run object. Creates the project if it doesn't exist.

run.log(name, value, step=None)

Buffer a metric data point. Flushed to the backend every 5 seconds by a background thread.

run["key"] = {...}

Set run parameters via dict assignment. Values are type-inferred (int, float, bool, string).

run["parameters"] = {"lr": 0.001, "epochs": 10}
run["system"] = {"gpu": "A100", "framework": "pytorch"}

run.stop()

Flush remaining data and mark the run as completed. Called automatically via atexit if not called explicitly.

run.id

The integer run ID assigned by the backend.

How It Works

• Metrics and params are collected in thread-safe buffers
• A background daemon thread flushes every 5 seconds and sends heartbeats
• Failed flushes re-add data to the buffer for retry
• Requires Python 3.10+ and httpx