OwLab

ML experiments, tracked & notified.

A Python toolkit for the full lifecycle of machine learning experiments — experiment tracking with SwanLab, notifications & data management with Lark (Feishu), and local storage.

---

✨ Features

📈 Metrics and Tracking: Embed minimal code into your ML pipeline to track and record key training metrics based on SwanLab.
📊 Data Management: Automatically organize your experiment directory structure based on experiment type and tags, enabling better management of experimental data.
📢 Message Notifications: Automatic push notifications are sent when the experiment starts, ends, or is interrupted, keeping you informed of the latest progress.
💾 Backup: Back up your data in the cloud and locally to prevent data loss.

---

🚀 Quick Start

📦 Installation

pip install owlab
# or: uv pip install owlab
# or: use source code
# git clone https://github.com/Lounwb/OwLab.git && cd OwLab && pip install -e .

⚙️ Configuration

To enable Lark and Swanlab, you need to configure the relevant tokens and secrets. Owlab supports both configuration files and environment variables, providing you with flexible options：

• Configuration file: ~/.owlab/config.json or ./.owlab/config.json, here is an example:

// configure your lark and swanlab in .owlab/config.json
{
    "lark": {
      "webhook": {
        "webhook_url": "<your webhook url>",
        "signature": "<your webhook signature>"
      },
      "api": {
        "app_id": "<your app id>",
        "app_secret": "<your app secret>",
        "root_folder_token": "<your root folder token>"
      }
    },
    "swanlab": {
      "api_key": "<your swanlab api key>"
    },
    "storage": {
      "local_path": "./output",
      "csv_path": "./output/csv",
      "model_path": "./output/models"
    },
    "logging": {
      "level": "INFO",
      "format": null,
      "file": "./logs/owlab.log"
    }
  }
  

• Environment: OWLAB_LARK__WEBHOOK__WEBHOOK_URL, OWLAB_LARK__API__APP_ID, etc.

📖 Usage

1. Initialize

from owlab import OwLab

owlab = OwLab()
owlab.init(
    project="my_project",           # Required
    experiment_name="exp_001",      # Optional; defaults to project
    description="Short description",
    type="baseline",                # e.g. baseline / debug / ablation — used for folder naming
    version="1.0",                 # Experiment version
    tags=["baseline"],             # Optional tags
    config={
        "methods": [...],          # Method definitions for result tables
        "datasets": [...],
        "metrics": [...],
        "measures": [...],
        "experiment_params": {"learning_rate": 0.01, "batch_size": 32},
        "seed": 42,
    },
)

2. Log metrics during training

for epoch in range(100):
    owlab.log({"loss": loss, "accuracy": acc}, step=epoch)

3. Finish and save results

Call finish(results=...) with a list of result rows. Each row can include method, dataset, measure, and metric values. These are written to local files and, when configured, to Feishu spreadsheets.

owlab.finish(results=[
    {
        "method": "method1",
        "dataset1": {"measure": "MCM", "accuracy": 0.95, "loss": 0.05},
        "dataset2": {"measure": "MCM", "accuracy": 0.92, "loss": 0.08},
        "Average": {"measure": "MCM", "accuracy": 0.935, "loss": 0.065},
    },
    # ...
])

4. Sync PyTorch TensorBoard to SwanLab

Like SwanLab’s swanlab.sync_tensorboard_torch(): call after init() and before creating SummaryWriter. Then writer.add_scalar() / add_scalars() also log to the current SwanLab run.

owlab.init(project="my_project", experiment_name="exp_1", ...)
owlab.sync_tensorboard_torch()

from torch.utils.tensorboard import SummaryWriter
writer = SummaryWriter(log_dir="./runs")
writer.add_scalar("loss", loss, step)   # also sent to SwanLab
writer.add_scalar("acc", acc, step)

5. Output layout

• Local: ./output/<type>/<experiment_name>_<timestamp>/
  • results.csv, results.json, owlab.log, model/
• Lark: Notifications via webhook; result tables written to Feishu via API (when configured).
• SwanLab: Metrics and runs visible in your SwanLab project (when api_key is set).

---

📄 License & Links

• PyPI: pypi.org/project/owlab
• License: MIT
• Repository: github.com/Lounwb/OwLab
• Issues: github.com/Lounwb/OwLab/issues

---

🙏 Acknowledgments

• SwanLab — experiment tracking
• Lark / Feishu — notifications and spreadsheets

⭐ Star History