Computer Vision Workload Example (Ray Train)

In this Workload module, you’ll build an end-to-end computer-vision pipeline that loads and preprocesses a Food-101-Lite dataset slice, then trains a ResNet-18 image classifier. You’ll learn how to migrate the workflow from a single-node PyTorch setup to a fully distributed training job on an Anyscale cluster using Ray Train.

• a Computer-vision pattern with Ray Train
• Imports
• Custom `FoodDataset` for Parquet
• Helper: Ray-prepared DataLoaders
• `train_loop_per_worker`
• Launch distributed training with `TorchTrainer`
• Plot training and validation loss curves
• Demonstrate fault-tolerant resumption
• Batch inference with Ray Data