ParaK8s

ParaK8s is a framework for organizing and running large-scale machine learning experiments on multi-node Kubernetes clusters.
Its core philosophy is that all experiments are trajectories within a large function space: each experiment is fully described by a configuration file specifying data pipelines, model architecture, logging, results tracking, and model registry.

---

Core Tenets

1. Experiments as Configurations
   Every experiment is uniquely determined by its configuration file, model parameters, and datapipelines.

2. Reproducibility
   All experiments are reproducible via configuration files, Docker images, and version-controlled datasets.

3. Scalability
   ParaK8s can orchestrate hundreds of experiments across multi-node Kubernetes clusters, leveraging GPUs where available.

4. Modularity
   Pipelines, models, and data preprocessing functions are designed to be reusable across experiments.

5. Experiment Tracking & Artifact Management
   Full integration with MLflow, databases (Postgres), and model registries ensures experiments are fully auditable.

---

Example Experiments

1. Pneumonia X-Ray Classifier

Objective:
Classify X-ray images to detect the presence of pneumonia, outputting a probability for each image.

Key Achievements:

• Achieved 85% categorical accuracy and 0.82 AUC score.
• Utilized transfer learning and data augmentation using TensorFlow.
• Parallelized preprocessing using Keras PyDataset.
• Built reusable GPU-enabled Docker images.
• Deployed 100 experiments across a multi-node Kubernetes cluster.
• Configured Postgres database, MLflow tracking server, and a local Docker repository.

---

2. ECG Abnormality Classifier

Objective:
Detect abnormalities in ECG signals using a multi-class classifier.

Key Achievements:

• Built a DeepCNN model to classify ECG beats into 5 categories.
• Achieved competitive AUC, precision, recall, and categorical accuracy.
• Preprocessed and normalized ECG signals using TensorFlow operations, ensuring consistency across devices.
• Managed large datasets with a custom ECGDataset class (tf.keras.utils.Sequence).
• Experiments defined entirely by configuration files describing architecture, batch sizes, and data pipelines.
• Deployed multiple experiments on a multi-node Kubernetes cluster with GPU support.
• Integrated metrics tracking via MLflow and artifact management in a model registry.

Pipeline Overview:

1. Load and preprocess ECG signals from CSV.
2. Batch and normalize signals using the ECGDataset class.
3. Train DeepCNN with class-weighted loss for imbalanced datasets.
4. Evaluate with AUC, precision, recall, and categorical accuracy.
5. Save trained models and track metrics in MLflow.

---

ParaK8s demonstrates flexibility and reproducibility, supporting experiments on both image and time-series data, while leveraging modular configurations, GPU acceleration, and cluster-level orchestration.