FiftyOne tutorial: train YOLO and review predictions

Prerna demonstrates a custom model training plugin inside FiftyOne with an agricultural pest detection dataset. The example contains almost 20,000 samples with bounding boxes and pest class labels. The workflow covers preparing training splits, scheduling a training job, and reviewing model predictions in the dataset interface.

The plugin uses samples tagged train for training. Its distribution panel initially reports zero training images. Prerna excludes the validation and test samples, then tags the remaining samples as train. The panel updates to show about 13,000 training samples and lets her filter the validation set for inspection.

To launch training, she supplies initial model weights, an output model path, and an epoch count. The interface exposes background compute options and confirms that the job has been scheduled. A run tab shows its progress.

The apply model tab accepts a weights path and a prediction field for YOLO inference. Prerna uses a model she calls 11S pest detection and saves its results in a dataset prediction field. She then reviews the predicted boxes and says the model finds the pests well. The demonstration provides a visual assessment, without reporting accuracy metrics or hardware requirements.