Function normalize_pir_program

python/paddle/static/pir_io.py:231–330

Normalize/Optimize a program according to feed_vars and fetch_vars. Args: program(Program): Specify a program you want to optimize. feed_vars(Tensor | list[Tensor]): Values needed by inference. fetch_vars(Tensor | list[Tensor]): Values returned by inference.

(program, feed_vars, fetch_vars, **kwargs)

def normalize_pir_program(program, feed_vars, fetch_vars, **kwargs):
    """

    Normalize/Optimize a program according to feed_vars and fetch_vars.

    Args:
        program(Program): Specify a program you want to optimize.
        feed_vars(Tensor | list[Tensor]): Values needed by inference.
        fetch_vars(Tensor | list[Tensor]): Values returned by inference.
        kwargs: Supported keys including ``skip_prune_program``.
            - skip_prune_program(bool): whether to skip pruning program. Defaults to False.

    Returns:
        Program: Normalized/Optimized program.

    Examples:
        .. code-block:: pycon

            >>> import paddle

            >>> paddle.enable_static()

            >>> path_prefix = "./infer_model"

            # User defined network, here a softmax regression example
            >>> image = paddle.static.data(name='img', shape=[None, 28, 28], dtype='float32')
            >>> label = paddle.static.data(name='label', shape=[None, 1], dtype='int64')
            >>> predict = paddle.static.nn.fc(image, 10, activation='softmax')

            >>> loss = paddle.nn.functional.cross_entropy(predict, label)

            >>> exe = paddle.static.Executor(paddle.CPUPlace())
            >>> exe.run(paddle.static.default_startup_program())

            # normalize main program.
            >>> program = paddle.static.default_main_program()
            >>> normalized_program = paddle.static.normalize_program(program, [image], [predict])

    """
    if not isinstance(program, paddle.static.Program):
        raise TypeError(
            f"program type must be `paddle.static.Program`, but received `{type(program)}`"
        )
    if not isinstance(feed_vars, list):
        feed_vars = [feed_vars]
    if not all(isinstance(v, pir.Value) for v in feed_vars):
        raise TypeError("feed_vars type must be a Value or a list of Value.")
    if not isinstance(fetch_vars, list):
        fetch_vars = [fetch_vars]
    if not all(isinstance(v, pir.Value) for v in fetch_vars):
        raise TypeError("fetch_vars type must be a Value or a list of Value.")

    if len(program.global_block().ops) == 0:
        raise ValueError(
            "program must not be empty. at least one operator is required!"
        )

    # remind users to set auc_states to 0 if auc op were found.