tensorflow-data-pipelines

Create a dataset with tf.data.Dataset.from_tensor_slices or from_generator, then chain transformations: map for preprocessing, shuffle for randomization, batch for grouping, and prefetch with tf.data.AUTOTUNE at the end to overlap loading with training.

Add prefetch(tf.data.AUTOTUNE) as the final step, use num_parallel_calls=tf.data.AUTOTUNE in map operations, and place cache() after expensive preprocessing but before augmentation. Shuffle before batching with a buffer of several thousand elements.

Yes. Use tf.data.TFRecordDataset with tf.io.parse_single_example for TFRecord shards, and tf.data.TextLineDataset with tf.io.decode_csv for CSV files. Both support parallel parsing via map with num_parallel_calls.

GPU idleness usually means the input pipeline is the bottleneck. Common causes are missing prefetch(), sequential map operations without num_parallel_calls, or expensive preprocessing done on every epoch instead of being cached.

Avoid caching when the dataset exceeds available memory, since cache() stores all elements in RAM or on disk. Also never cache after random augmentation, because it freezes the augmented values and removes per-epoch randomness.