Design a Data Batcher (Code & Design)

Question 2: Code & Design — Design and implement a Data Batcher Design a data batcher to organize samples into training batches and iterate efficiently within a training loop. Requirements Input: A data source dataset (assume an iterable, or a container with __len__ and __getitem__). batch_size. Output: An iterable batcher/iterator that yields one batch at a time (a list of samples or a dict of tensors). Batching behavior (common follow-ups): Support shuffle (train) vs no shuffle (eval). Handle the last incomplete batch (drop/keep/pad). Reproducibility via seed. Optional prefetch or lightweight concurrent loading (design discussion is fine). Variable-length samples and a collate function (padding + mask). Tasks Propose an API (class/function signature). Describe core data structures and time/space complexity. Provide a core implementation (Python pseudocode or runnable code). Explain how you would test it (at least 3 test ideas). Constraints Dataset size: 1e5–1e8 samples Limited memory; cannot necessarily load all indices/contents at once Throughput should be high enough to avoid becoming a training bottleneck