tfds.benchmark

Benchmarks any iterable (e.g tf.data.Dataset).

tfds.benchmark(
    ds: Iterable[Any], *, num_iter: Optional[int] = None, batch_size: int = 1
) -> BenchmarkResult

Used in the notebooks

Used in the tutorials

Usage:

ds = tfds.load('mnist', split='train')
ds = ds.batch(32).prefetch(buffer_size=tf.data.AUTOTUNE)
tfds.benchmark(ds, batch_size=32)

Reports:

  • Total execution time
  • Setup time (first warmup batch)
  • Number of examples/sec

ds Dataset to benchmark. Can be any iterable. Note: The iterable will be fully consumed.
num_iter Number of iteration to perform (iteration might be batched)
batch_size Batch size of the dataset, used to normalize iterations

statistics The recorded statistics, for eventual post-processing