Gaussian Mixture Model training

更新时间:
复制 MD 格式

Gaussian mixture models (GMMs) are a probabilistic model that represents K Gaussian distributions within a complete population. You can use the GMM Training component to classify models.

Supported computing resources

The GMM Training component runs on one of the following computing resources:

  • MaxCompute

  • Realtime Compute for Apache Flink

  • Deep Learning Containers (DLC) of Platform for AI (PAI)

Estimate resource usage

Use these guidelines to size workers before submitting a training job.

Memory per worker

Calculate the minimum memory per worker (in MB) using:

M × M × K × 8 × 2 × 12 / 1024 / 1024

Where:

  • M = number of vector dimensions

  • K = number of Gaussian components (the k parameter)

For most workloads, 8 GB (8192 MB) per worker is sufficient.

Keep vector dimensions (M) below 200. Higher dimensionality significantly increases memory requirements.

Number of workers

As a starting point, use 5 × X workers, where X is the input data size in GB.

If resources are constrained, reduce the worker count. Adding more workers speeds up training up to a point — beyond that, cross-worker communication overhead outweighs the parallelism benefit. Tune the worker count to find the optimal setting for your dataset.