spectria 0.1.0

JSONL logging library for live training visualization in Spectria

spectria-python

JSONL logging library + HTTP/SSE server for live training visualization in Spectria.

A lightweight alternative to TensorBoard with a cleaner, more intentional visualization experience.

Usage

In your training script — log metrics (generic)

pip install spectria

from spectria.logger import RunWriter

writer = RunWriter(project="mnist", run="baseline", config={"lr": 0.01})
writer.write_row({"epoch": 0, "loss": 0.69, "accuracy": 0.51})

Keras integration

pip install spectria[keras]

from spectria.logger import SpectriaCallback, as_keras_callback

model.fit(x, y, callbacks=[
    as_keras_callback(SpectriaCallback(
        project="mnist",
        run="baseline",
        config={"model": "resnet50", "lr": 0.01, "optimizer": "adam", "batch_size": 64},
    ))
])

Runs with a shared baseline are grouped in Spectria:

as_keras_callback(SpectriaCallback(
    project="mnist",
    run="lr-0.001",
    baseline="baseline",
    config={"model": "resnet50", "lr": 0.001, "optimizer": "adam", "batch_size": 64},
))

Visualize — serve the logs

uvx spectria serve

Then open Spectria, click Connect, enter http://localhost:8420, and select your runs.

Options:

spectria serve --logdir ./spectria_logs --port 8420 --host 127.0.0.1

API Endpoints

Endpoint	Description
GET /api/projects	List all projects
GET /api/projects/{name}/runs	List runs with metadata
GET /api/projects/{name}/runs/{run}/data	Full JSON snapshot
GET /api/projects/{name}/runs/{run}/events	SSE stream for live data

Data Format

Training events are stored as JSONL:

# {"spectria_version":1,"project":"mnist","run":"lr-0.001","baseline":"baseline","config":{"lr":0.001},"created_at":1714732800}
{"epoch":0,"loss":0.6931,"accuracy":0.5123,"_ts":1714732800}
{"epoch":1,"loss":0.4201,"accuracy":0.8102,"_ts":1714732806}

License

MIT