# Speed Up FiftyOneTorchDataset with Vectorize Mode This recipe shows how to eliminate database query overhead during training by enabling vectorize mode in [FiftyOneTorchDataset](https://docs.voxel51.com/api/fiftyone.utils.torch.html#fiftyone.utils.torch.FiftyOneTorchDataset). With `vectorize=True`, all required fields are preloaded into memory before training begins, removing per-sample database lookups from the hot path. Specifically, it covers: - Enabling vectorize mode with `vectorize=True` and a [GetItem](https://docs.voxel51.com/api/fiftyone.utils.torch.html#fiftyone.utils.torch.GetItem) subclass - Understanding how vectorize mode changes the input to your `__call__` method - Writing a [GetItem](https://docs.voxel51.com/api/fiftyone.utils.torch.html#fiftyone.utils.torch.GetItem) that works with preloaded field dicts for efficient training **API references:** [FiftyOneTorchDataset](https://docs.voxel51.com/api/fiftyone.utils.torch.html#fiftyone.utils.torch.FiftyOneTorchDataset) · [GetItem](https://docs.voxel51.com/api/fiftyone.utils.torch.html#fiftyone.utils.torch.GetItem) ## Setup If you haven’t already, install FiftyOne: In this tutorial, we’ll use [PyTorch](https://pytorch.org/) for working with tensors and inspecting sample data. To follow along, you’ll need to install `torch` and `torchvision`, if necessary: ## Import Libraries This recipe requires a helper file, `utils.py`, which contains reusable functions for building `get_item` methods, creating dataloaders, and setting up models. The following cell downloads it into your working directory so it can be imported directly. ## Vectorize Mode By default, [FiftyOneTorchDataset](https://docs.voxel51.com/api/fiftyone.utils.torch.html#fiftyone.utils.torch.FiftyOneTorchDataset) queries the backing MongoDB database on every `__getitem__` call. For large training runs this per-sample overhead adds up quickly. Passing `vectorize=True` to `.to_torch()` switches to vectorize mode: all fields declared in your [GetItem](https://docs.voxel51.com/api/fiftyone.utils.torch.html#fiftyone.utils.torch.GetItem) are serialized into memory up front, so sample retrieval during training is a simple dict lookup with no database I/O. This delivers a significant speedup when data loading is the bottleneck. ```python # Wrap your function with SimpleGetItem to declare which fields to preload get_item_wrapper = SimpleGetItem(my_get_item_fn, ["id", "filepath", "my_field"]) # vectorize=True preloads all declared fields into memory before training begins torch_dataset = view.to_torch(get_item_wrapper, vectorize=True) ``` > **Note:** With `vectorize=True`, all fields you declare in your [GetItem](https://docs.voxel51.com/api/fiftyone.utils.torch.html#fiftyone.utils.torch.GetItem) are preloaded into memory before training begins — no additional configuration is needed beyond the field list you already provide. ## Load Dataset ## Writing a GetItem for Vectorize Mode In vectorize mode, your `GetItem.__call__` receives a `dict` of preloaded field values rather than a live `fiftyone.core.sample.Sample`. The keys match the field names you declared in your [GetItem](https://docs.voxel51.com/api/fiftyone.utils.torch.html#fiftyone.utils.torch.GetItem). Here’s an example that loads detections from the quickstart dataset: ## Visualizing the result Run the cell below a few times to inspect different samples from the dataset: The dataset works with a standard `DataLoader` for use in a training loop: