Skip to content

Tracking Overview

Overview

OICM’s tracking server is built to work with MLflow and manages machine learning experiments by capturing parameters, metrics, and artifacts. It integrates with the Tracking Client, an API client designed to facilitate smooth interaction with the tracking server during various MLOps tasks.

Key Capabilities

1. Parameter & Metric Logging

  • Detailed Analysis – Capture hyperparameters, learning rates, and performance metrics.
  • Consistent Records – Track all values over time to compare different experiment runs.

2. Artifact Storage

  • Model Artifacts – Store trained models, checkpoints, and other files in a centralized repository.
  • Comprehensive Documentation – Keep a complete record of experiment outputs for easy review.

3. Data Security

  • Secure Access – Protect tracked data with role-based permissions and encryption.
  • Compliance – Align with best practices for data governance.

4. Isolated Experimentation

  • Reliability – Run experiments in distinct environments to prevent cross-contamination of results.
  • Repeatability – Ensure each experiment’s configuration is clearly documented and reproducible.

Using the Tracking Client

OICM’s Tracking Client provides a simplified way to log parameters and artifacts. Below is a conceptual snippet (pseudocode) showing how you might log an experiment run:

from oicm_tracking import TrackingClient

# Initialize the client
client = TrackingClient(server_url="https://your-tracking-server.com")

# Start a new run
run_id = client.start_run("Experiment_Name")

# Log parameters and metrics
client.log_param(run_id, "learning_rate", 0.001)
client.log_metric(run_id, "accuracy", 0.95)

# Log artifacts
client.log_artifact(run_id, "model_checkpoint.pt")

# End the run
client.end_run(run_id)

Next Steps

  • Usage Monitoring – Keep track of resource utilization across experiments.
  • Workspace Overview – Discover how workspace isolation further enhances your tracking workflows.
  • API Reference – Integrate with the tracking server programmatically.