# loaders

Python API: `luxonis_train.loaders`

Loaders that feed batches to the training loop.

[LuxonisLoaderTorch](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/loaders/luxonis_loader_torch.md)
is the default. It reads an existing `LuxonisDataset`, or parses a supported directory into a new one.
[LuxonisLoaderPerlinNoise](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/loaders/luxonis_perlin_loader_torch.md)
adds synthetic anomalies inside a Perlin noise mask, for the anomaly detection task.
[DummyLoader](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/loaders/dummy_loader.md)
yields samples of zeros, so a model can build when no dataset is available.

 * `Labels:`:
   [BaseLoaderTorch.collate_fn](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/loaders/base_loader.md)
   returns each label of a batch under its `"<task_name>/<label>"` key. `B` is the batch size, and `N` is the number of instances
   in the whole batch. * `classification`: `[B, C]`, the multi-hot class vector of each image.
    * `segmentation`: `[B, C, H, W]`, one mask channel for each class.
    * `boundingbox`: `[N, 6]`, the rows `[batch_index, class, x, y, w, h]`.
    * `keypoints`: `[N, 3K + 1]`, a batch index and then `(x, y, visibility)` for each of the `K` keypoints.
    * `instance_segmentation`: `[N, H, W]`, one mask for each box, in the order of the boxes.
    * `metadata/text`: `[B, S]`, the character codes of the text of each image, padded with zeros to the length `S` of the longest
      text. The `ocr` task reads this label.
    * Other `metadata/<name>` labels: the values of all samples, joined along the first dimension. The result has the shape `[B]`
      when each image has one value, such as the `metadata/id` label of the `embeddings` task.
 * `Writing a custom loader:`: Subclass
   [BaseLoaderTorch](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/loaders/base_loader.md)
   and implement `input_shapes`, `__len__`, `__getitem__`, and `get_classes`. A loader with keypoint labels must also implement
   `get_n_keypoints`. Override `collate_fn` when the labels need other merge rules. Put the name of the subclass in the
   `loader.name` field of a config.

## Child Pages

 * [base_loader](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/loaders/base_loader.md)
 * [dummy_loader](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/loaders/dummy_loader.md)
 * [luxonis_loader_torch](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/loaders/luxonis_loader_torch.md)
 * [luxonis_perlin_loader_torch](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/loaders/luxonis_perlin_loader_torch.md)
 * [perlin](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/loaders/perlin.md)