Upload 379 files

610d9c2 verified about 3 hours ago

2.22 kB

import onnx

from ..quant_utils import TENSOR_NAME_QUANT_SUFFIX, QuantizedValue, QuantizedValueType, attribute_to_kwarg, ms_domain
from .base_operator import QuantOperatorBase


class QLinearPool(QuantOperatorBase):
    def __init__(self, onnx_quantizer, onnx_node):
        super().__init__(onnx_quantizer, onnx_node)

    def quantize(self):
        node = self.node

        # only try to quantize when given quantization parameters for it
        (
            data_found,
            output_scale_name,
            output_zp_name,
            _,
            _,
        ) = self.quantizer._get_quantization_params(node.output[0])

        # get quantized input tensor names, quantize input if needed
        (
            quantized_input_names,
            input_zero_point_names,
            input_scale_names,
            nodes,
        ) = self.quantizer.quantize_activation(node, [0])

        if not data_found or quantized_input_names is None:
            return super().quantize()

        # Create an entry for output quantized value.
        qlinear_output_name = node.output[0] + TENSOR_NAME_QUANT_SUFFIX
        quantized_output_value = QuantizedValue(
            node.output[0],
            qlinear_output_name,
            output_scale_name,
            output_zp_name,
            QuantizedValueType.Input,
        )
        self.quantizer.quantized_value_map[node.output[0]] = quantized_output_value

        # Create qlinear pool node for given type (AveragePool, etc)
        kwargs = {}
        for attribute in node.attribute:
            kwargs.update(attribute_to_kwarg(attribute))
        kwargs["domain"] = ms_domain
        qlinear_node_name = node.name + "_quant" if node.name else ""
        qnode = onnx.helper.make_node(
            "QLinear" + node.op_type,
            [
                quantized_input_names[0],
                input_scale_names[0],
                input_zero_point_names[0],
                output_scale_name,
                output_zp_name,
            ],
            [qlinear_output_name],
            qlinear_node_name,
            **kwargs,
        )

        # add all newly created nodes
        nodes.append(qnode)
        self.quantizer.new_nodes += nodes

	import onnx

	from ..quant_utils import TENSOR_NAME_QUANT_SUFFIX, QuantizedValue, QuantizedValueType, attribute_to_kwarg, ms_domain
	from .base_operator import QuantOperatorBase


	class QLinearPool(QuantOperatorBase):
	def __init__(self, onnx_quantizer, onnx_node):
	super().__init__(onnx_quantizer, onnx_node)

	def quantize(self):
	node = self.node

	# only try to quantize when given quantization parameters for it
	(
	data_found,
	output_scale_name,
	output_zp_name,
	_,
	_,
	) = self.quantizer._get_quantization_params(node.output[0])

	# get quantized input tensor names, quantize input if needed
	(
	quantized_input_names,
	input_zero_point_names,
	input_scale_names,
	nodes,
	) = self.quantizer.quantize_activation(node, [0])

	if not data_found or quantized_input_names is None:
	return super().quantize()

	# Create an entry for output quantized value.
	qlinear_output_name = node.output[0] + TENSOR_NAME_QUANT_SUFFIX
	quantized_output_value = QuantizedValue(
	node.output[0],
	qlinear_output_name,
	output_scale_name,
	output_zp_name,
	QuantizedValueType.Input,
	)
	self.quantizer.quantized_value_map[node.output[0]] = quantized_output_value

	# Create qlinear pool node for given type (AveragePool, etc)
	kwargs = {}
	for attribute in node.attribute:
	kwargs.update(attribute_to_kwarg(attribute))
	kwargs["domain"] = ms_domain
	qlinear_node_name = node.name + "_quant" if node.name else ""
	qnode = onnx.helper.make_node(
	"QLinear" + node.op_type,
	[
	quantized_input_names[0],
	input_scale_names[0],
	input_zero_point_names[0],
	output_scale_name,
	output_zp_name,
	],
	[qlinear_output_name],
	qlinear_node_name,
	**kwargs,
	)

	# add all newly created nodes
	nodes.append(qnode)
	self.quantizer.new_nodes += nodes