This document explains the key differences between the training methods available on Google Cloud. Your choice depends on your team's expertise, the level of control you need, and your infrastructure preference..
With AutoML, you create and train a model with minimal technical effort. You can use AutoML to quickly prototype models and explore new datasets before investing in development. For example, you can use it to learn which features are best for a given dataset.
With custom training you can create a training application optimized for your targeted outcome. You have complete control over training application functionality, from single-node jobs to massive, multi-node distributed training. Namely, you can target any objective, use any algorithm, develop your own loss functions or metrics, or do any other customization.
Vertex AI serverless training: This is a serverless approach where you submit your training job and Google Cloud provisions, manages, and releases the compute resources for you. It's ideal for experimentation and for jobs where you don't need assured capacity.
Vertex AI training clusters is a feature within custom training designed for large-scale, high-performance training jobs. It lets you reserve a dedicated cluster of powerful compute resources (like A100 or H100 GPUs) for your exclusive use, guaranteeing capacity and performance for mission-critical and long-running training tasks.
With Ray on Vertex AI you can use Ray's distributed computing framework on Google Cloud infrastructure. Ray on Vertex AI provides a managed environment with configurable compute resources, integration with services like Vertex AI Inference and BigQuery, and flexible networking options for developing and running distributed workloads.
Using BigQuery, you can train models using your BigQuery data directly in BigQuery. Using SQL commands, you can quickly create a model and use it to get batch inferences.
To compare the different functionality and expertise required for each service, review the following table.
| AutoML | serverless training | training clusters | Ray on Vertex AI | BigQuery ML | |
|---|---|---|---|---|---|
| Data science expertise needed | No | Yes, to develop the training application and handle data preparation. | Yes, to develop the training application and handle data preparation. | A foundational understanding of ML concepts and data workflows is beneficial. | No |
| Programming ability needed | No. AutoML is codeless. | Yes, to develop the training application. | Yes, to develop the training application. | Yes. | Yes. |
| Time to train model | Lower. Less data preparation is required, and no development is needed. | Higher. Involves code development and on-demand resource provisioning time for each job. | Higher. Involves code development, but job start-up is faster as resources are already reserved, eliminating queues and provisioning time. | The time to train depends on the code logic (data preparation and training) and time to provision resources. | Lower. Model development is fast as BigQuery ML leverages the BigQuery computational engine for training, evaluation, and inference. |
| Limits on machine learning objectives | Yes. You must target one of AutoML's predefined objectives. | No | No | No | Yes |
| Can manually optimize model performance with hyperparameter tuning | No. AutoML performs some automated tuning, but you can't modify the values. | Yes. You can tune the model during each training run for experimentation and comparison. | Yes. You can tune the model during each training run for experimentation and comparison. | Yes. You provide the custom training code, giving you full control over hyperparameter values. | Yes. BigQuery ML supports hyperparameter tuning when training models. |
| Can control aspects of the training environment | Limited. You can specify the budget for training hours and enable early stopping. | Yes. You specify the Compute Engine machine type, disk size, and container image for each job. | Yes. You have maximum control, reserving specific high-performance machine types (e.g., H100s), disk configurations, and network settings for your exclusive use, guaranteeing capacity. | Yes. You have significant control, including custom Docker images, machine types for head and worker nodes, and the number and type of accelerators (GPUs). | No |
| Limits on data size | Yes. Data size limitations vary depending on the dataset type. | No | No | No. However, there is a maximum query response size of 10 GB for BigQuery reads. | Yes. BigQuery ML enforces appropriate quotas on a per-project basis. To learn more, see Quotas and limits. |
What's next
- Choose an introductory tutorial to get started with Vertex AI Training.
- Learn more about training an AutoML model.
- Learn about creating a custom training job using Python.
- Learn more about Ray on Vertex AI.