caliper-py 0.1.1

Local-first ML experiment tracker with a real-time dashboard, convergence forecasting, and zero cloud dependencies

Caliper

Local-first ML experiment tracker. Log metrics, visualize training runs, compare experiments — all on your machine, zero cloud required.

---

What is Caliper?

Caliper is a lightweight experiment tracker that stores everything locally as plain JSONL files and serves a rich dashboard from a single CLI command. No accounts, no API keys, no telemetry.

• Convergence forecasting — predicts final val loss and steps to convergence in real-time
• Step-level annotations — pin notes directly onto chart steps
• Multi-run comparison — overlay up to 8 runs with hyperparameter diffs
• Project + tag filtering — organize and filter runs from the sidebar
• GPU monitoring — log utilization, VRAM, temperature, and CPU alongside metrics
• 100% offline — all data lives in .caliper/runs/ in your working directory

---

Installation

pip install caliper-py

Requires Python 3.8+. No heavy dependencies.

---

Quickstart

1. Initialize a run in your training script

import caliper

run = caliper.init(
    project="CIFAR-10",
    name="resnet50-baseline",
    tags=["ResNet-50", "baseline"],
    hyperparams={
        "model": "ResNet-50",
        "optimizer": "Adam",
        "learning_rate": 0.001,
        "batch_size": 64,
        "epochs": 50,
    }
)

2. Log metrics each step

for step, batch in enumerate(dataloader):
    loss = train_step(batch)
    val_loss, val_acc = evaluate()

    run.log({
        "step": step,
        "trainLoss": loss,
        "valLoss": val_loss,
        "valAcc": val_acc,
        "lr": scheduler.get_last_lr()[0],
        # optional hardware metrics
        "gpuUtil": get_gpu_util(),
        "vramUsage": get_vram_gb(),
    })

run.finish()

3. Launch the dashboard

caliper ui

Opens http://localhost:5173 automatically. All runs in .caliper/runs/ load instantly.

---

CLI Reference

caliper ui                  Launch the dashboard (default port 5173)
caliper ui --port 8080      Use a custom port
caliper ui --no-browser     Start server without opening a browser

---

Data Format

Caliper writes two files per run inside .caliper/runs/:

File	Description
{run_id}_meta.json	Run metadata: name, project, tags, hyperparams, status, annotations
{run_id}_logs.jsonl	One JSON object per logged step

Both are plain text — you can read, edit, or version-control them however you like.

Example _meta.json

{
  "id": "run-001",
  "name": "resnet50-baseline",
  "project": "CIFAR-10",
  "status": "completed",
  "tags": ["ResNet-50", "baseline"],
  "hyperparams": { "model": "ResNet-50", "learning_rate": 0.001 },
  "annotations": [],
  "meta": { "startTime": "2026-06-01T10:00:00Z", "duration": 7260 }
}

Example _logs.jsonl (one line per step)

{"step": 0, "trainLoss": 2.31, "valLoss": 2.45, "valAcc": 0.12, "lr": 0.001}
{"step": 1, "trainLoss": 2.18, "valLoss": 2.30, "valAcc": 0.15, "lr": 0.001}

---

Generating Sample Data

A sample data generator is included for testing the dashboard:

python -c "
import subprocess, sys
subprocess.run([sys.executable, '-m', 'pip', 'show', '-f', 'caliper-py'])
"

Or clone the repo and run:

git clone https://github.com/verz0/Caliper
cd Caliper/caliper-py
python generate_sample_runs.py
caliper ui

---

Links

• Homepage & demo: verz0.github.io/Caliper
• Source code: github.com/verz0/Caliper
• Issues: github.com/verz0/Caliper/issues

---

Made by Pirajesh M R · MIT License