This document explains the key differences between training a model in Vertex AI using AutoML, custom training, Ray on Vertex AI or training a model using BigQuery ML.
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. Namely, you can target any objective, use any algorithm, develop your own loss functions or metrics, or do any other customization.
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 Prediction 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 predictions.
To compare the different functionality and expertise required for each service, review the following table.
AutoML | Custom training | Ray on Vertex AI | BigQuery ML | |
---|---|---|---|---|
Data science expertise needed | No | Yes, to develop the training application and also to do some of the data preparation like feature engineering. | Involves having a foundational understanding of machine learning concepts and data science workflows is beneficial. | No |
Programming ability needed | No. AutoML is codeless. | Yes, to develop the training application. | Yes. | Yes. |
Time to train model | Lower. Less data preparation is required, and no development is needed. | Higher. More data preparation is required, and training application development is needed. Using distributed training can reduce the time it takes to train a model. | The time to train depends on the code logic (data preparation and training) and time to provision resources. | Lower. Model development speed is increased since you don't need build the infrastructure required for batch predictions or model training, as BigQuery ML leverages the BigQuery computational engine. This increases speed to training, evaluation, and prediction. |
Limits on machine learning objectives | Yes. You must target one of AutoML's predefined objectives. | No | No | Yes |
Can manually optimize model performance with hyperparameter tuning | No. AutoML does some automated hyperparameter tuning, but you can't modify the values used. | Yes. You can tune the model during each training run for experimentation and comparison. | Yes. Because you are providing the custom training code, you have the ability to manually set or adjust the values of your hyperparameters before launching a Ray job on Vertex AI. | Yes. BigQuery ML supports hyperparameter tuning when training ML models using `CREATE MODEL` statements. |
Can control aspects of the training environment | Limited. For image and tabular datasets, you can specify the number of node hours to train for, and whether to allow early stopping of training. | Yes. You can specify aspects of the environment such as Compute Engine machine type, disk size, machine learning framework, and number of nodes, as well as the Docker image you want to use for training. | Yes. You have significant control over various aspects of the training environment. For example, you can bring your own custom Docker container images to your Ray cluster on Vertex AI. When creating a Ray cluster, you can specify the machine type for both the head node and worker nodes, including the number and type of accelerators (GPUs) you want to use, and more. | No |
Limits on data size |
Yes. AutoML uses managed datasets; data size limitations vary depending on the type of dataset. Refer to one of the following topics for specifics: |
For unmanaged datasets, no. Managed datasets have the same limits as managed dataset objects created in and hosted by BigQuery and are used to train AutoML models. | No. However, there is a maximum query response size of 10 GB for BigQuery reads. This is a limitation of the BigQuery API response size, not necessarily a limit on the total amount of data Ray on Vertex AI can process from BigQuery through iterative or parallel queries. | 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.