Batch size is the number of training examples processed together in one model-training step. It is a hyperparameter whose value must be set before training begins. section 5.0
Changing batch size can affect model quality, training time, and memory requirements, making it a practical configuration target for hyperparameter optimization. section 5.0
Use batch size as part of an HPO search rather than tuning it only by manual trial and error:
- Define candidate values such as 8, 16, 32, 64, 128, and 256. section 5.2.1
- Evaluate batch size together with other hyperparameters, such as learning rate, because their combination determines trial results. section 5.2.1
- Prefer automatic HPO when the search space becomes large or manual comparisons become time-consuming and error-prone. section 5.2.1
Searching batch size with learning rate
A search over six batch-size values and six learning-rate values produces 36 training trials. Comparing their evaluation results identifies the best pair. section 5.2.1
Figure 5.1 This high-level view of the HPO workflow shows that the process is essentially an experiment to find the optimal hyperparameter values.
Use the HPO workflow diagram to see batch size as one configurable input in an experiment for finding effective hyperparameter values. section 5.2.1
