d2c47d5834c8484fda791c91db22cf22f2022c26,tensorflow_datasets/core/dataset_utils.py,,build_dataset,#Any#Any#Any#Any#,29

Before Change

  // Transpose the list[dict] into dict[list]
  tensor_inputs = {
      key: list(values) for key, values in utils.zip_dict(*instruction_dicts)
  }
  // Skip slicing if all masks are True (No value skipped)
  if all(all(m) for m in tensor_inputs["mask"]):

After Change

  

  // First case: All examples are taken (No value skipped)
  if _no_examples_skipped(instruction_dicts):
    // Only use the filenames as instruction
    instruction_ds = tf.data.Dataset.from_tensor_slices([
        d["filepath"] for d in instruction_dicts
    ])
    build_ds_from_instruction = dataset_from_file_fn
  // Second case: Use the instructions to read the examples
  else:
    instruction_ds = _build_instruction_ds(instruction_dicts)
    build_ds_from_instruction = functools.partial(
        _build_ds_from_instruction,
        ds_from_file_fn=dataset_from_file_fn,
    )

  // If shuffle is True, we shuffle the instructions/shards
  if shuffle_files:
    instruction_ds = instruction_ds.shuffle(len(instruction_dicts))

  // Use interleave to parallel read files and decode records

Instances