Skip to content

MLOps Database Tracking

SpectraLoRA includes a built-in, zero-friction MLOps database tracker. By default, every time you run a training loop, the library automatically creates a local SQLite database (spectralora_experiments.db) to log your hyperparameters, epoch durations, and loss metrics.

Using PostgreSQL for Enterprise Tracking

If you are running massive training loops across a cluster or want to centralize your team's experiments, you can seamlessly override the local SQLite fallback and connect SpectraLoRA directly to a PostgreSQL database.

You do not need to write any SQL or manually create tables. SpectraLoRA handles namespace protection (prefixing all tables with spectralora_) and automatically generates the required schema.

How to Connect

Set the SPECTRALORA_DB_URL environment variable in your terminal before running your training script:

Linux/macOS: 

export SPECTRALORA_DB_URL="postgresql://username:password@server_address:5432/database_name"
python experiments/train.py

Windows (PowerShell):

$env:SPECTRALORA_DB_URL="postgresql://username:password@server_address:5432/database_name"
python experiments/train.py

Tracked Metrics

The database tracks both standard ML performance and GeoAI-specific data:

  • Hyperparameters: LoRA Rank, Alpha, Gate Temperature, Number of Experts.
  • Epoch Data: Training Loss, Validation Loss, Learning Rate.
  • GeoAI Metrics: Mean Intersection over Union (mIoU), Pixel Accuracy, and Physics Violation Scores.