Class QuantWeightPass

python/paddle/static/quantization/quantization_pass.py:3167–3306

quant weights and remove weights input quantize_linear node. for example: `weight -> quant -> dequant -> conv2d` will be frozen into `weight -> dequant -> conv2d`, and weight will be scaled offline. Args: scope(paddle.Scope): scope is used to get the weight tensor values.

class QuantWeightPass:
    """
    quant weights and remove weights input quantize_linear node. for example:
    `weight -> quant -> dequant -> conv2d` will be frozen into `weight -> dequant -> conv2d`,
    and weight will be scaled offline.

    Args:
        scope(paddle.Scope): scope is used to get the weight tensor values.
        place(paddle.CPUPlace|paddle.CUDAPlace|str): place is used to restore the weight tensors.
            If it's string, It can be ``cpu``, and ``gpu:x``, where ``x`` is the index of the GPUs.
        bias_correction(bool): whether use bias correction for post-training quantization.
             https://arxiv.org/abs/1810.05723.
        quant_bits(int, optional): quantization bit number for weight. Default is 8.
        save_int_weight(bool, optional): Whether the type saving the weight is int. Default is True.

    Examples:
        .. code-block:: pycon

            >>> # The original graph will be rewrite.
            >>> import paddle
            >>> import paddle.static as static
            >>> from paddle.static.quantization import QuantWeightPass
            >>> from paddle.base.framework import IrGraph
            >>> from paddle.framework import core

            >>> graph = IrGraph(core.Graph(paddle.static.Program().desc), for_test=False)
            >>> place = paddle.CPUPlace()
            >>> scope = paddle.static.global_scope()
            >>> quant_weight_pass = QuantWeightPass(scope, place)
            >>> quant_weight_pass.apply(graph)
    """

    def __init__(
        self,
        scope,
        place,
        bias_correction=False,
        quant_bits=8,
        save_int_weight=True,
    ):
        self._place = _get_paddle_place(place)
        self._scope = scope
        self._bias_correction = bias_correction
        self._quant_bits = quant_bits
        self._save_int_weight = save_int_weight
        assert self._scope is not None, "scope must not be None."
        assert self._place is not None, "place must not be None."
        self._quantized_ops = set()

    def apply(self, graph):
        assert isinstance(graph, IrGraph), (
            'graph must be the instance of IrGraph.'
        )
        fake_quant_ops_for_weight = []

        fake_quant_ops = [
            op for op in graph.all_op_nodes() if op.name() == "quantize_linear"
        ]