quanter.py 10.1 KB
Newer Older
F
ftian 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
# Copyright (c) 2019  PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import copy
import paddle
import paddle.fluid as fluid
from paddle.fluid.framework import IrGraph
from paddle.fluid.contrib.slim.quantization import QuantizationTransformPass
from paddle.fluid.contrib.slim.quantization import QuantizationFreezePass
from paddle.fluid.contrib.slim.quantization import ConvertToInt8Pass
from paddle.fluid.contrib.slim.quantization import TransformForMobilePass
S
slf12 已提交
23
from paddle.fluid.contrib.slim.quantization import PostTrainingQuantization
F
ftian 已提交
24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156
from paddle.fluid import core

WEIGHT_QUANTIZATION_TYPES=['abs_max', 'channel_wise_abs_max', 'range_abs_max', 'moving_average_abs_max']
ACTIVATION_QUANTIZATION_TYPES=['abs_max','range_abs_max', 'moving_average_abs_max']
VALID_DTYPES = ['int8']

_quant_config_default = {
    # weight quantize type, default is 'abs_max'
    'weight_quantize_type': 'abs_max',
    # activation quantize type, default is 'abs_max'
    'activation_quantize_type': 'abs_max',
    # weight quantize bit num, default is 8
    'weight_bits': 8,
    # activation quantize bit num, default is 8
    'activation_bits': 8,
    # ops of name_scope in not_quant_pattern list, will not be quantized
    'not_quant_pattern': ['skip_quant'],
    # ops of type in quantize_op_types, will be quantized
    'quantize_op_types': ['conv2d', 'depthwise_conv2d', 'mul'],
    # data type after quantization, such as 'uint8', 'int8', etc. default is 'int8'
    'dtype': 'int8',
    # window size for 'range_abs_max' quantization. defaulf is 10000
    'window_size': 10000,
    # The decay coefficient of moving average, default is 0.9
    'moving_rate': 0.9,
    # if set quant_weight_only True, then only quantize parameters of layers which need to be quantized,
    # and activations will not be quantized.
    'quant_weight_only': False
}


def _parse_configs(user_config):
    """
    check user configs is valid, and set default value if user not config.
    Args:
        user_config(dict):the config of user.
    Return:
        configs(dict): final configs will be used.
    """

    configs = copy.deepcopy(_quant_config_default)
    configs.update(user_config)

    # check configs is valid
    assert configs['weight_quantize_type'] in WEIGHT_QUANTIZATION_TYPES, \
        "Unknown weight_quantize_type: '%s'. It can only be " + " ".join(WEIGHT_QUANTIZATION_TYPES)

    assert configs['activation_quantize_type'] in ACTIVATION_QUANTIZATION_TYPES, \
        "Unknown activation_quantize_type: '%s'. It can only be " + " ".join(ACTIVATION_QUANTIZATION_TYPES)

    assert isinstance(configs['weight_bits'], int), \
        "weight_bits must be int value."

    assert (configs['weight_bits'] >= 1 and configs['weight_bits'] <= 16), \
        "weight_bits should be between 1 and 16."

    assert isinstance(configs['activation_bits'], int), \
        "activation_bits must be int value."

    assert (configs['activation_bits'] >= 1 and configs['activation_bits'] <= 16), \
        "activation_bits should be between 1 and 16."

    assert isinstance(configs['not_quant_pattern'], list), \
        "not_quant_pattern must be a list"

    assert isinstance(configs['quantize_op_types'], list), \
        "quantize_op_types must be a list"

    assert isinstance(configs['dtype'], str), \
        "dtype must be a str."

    assert (configs['dtype'] in VALID_DTYPES), \
        "dtype can only be " + " ".join(VALID_DTYPES)

    assert isinstance(configs['window_size'], int), \
        "window_size must be int value, window size for 'range_abs_max' quantization, default is 10000."

    assert isinstance(configs['moving_rate'], float), \
        "moving_rate must be float value, The decay coefficient of moving average, default is 0.9."

    assert isinstance(configs['quant_weight_only'], bool), \
        "quant_weight_only must be bool value, if set quant_weight_only True, " \
        "then only quantize parameters of layers which need to be quantized, " \
        " and activations will not be quantized."

    return configs


def quant_aware(program, place, config, scope=None, for_test=False):
    """
    add trainable quantization ops in program.
    Args:
        program(fluid.Program): program
        scope(fluid.Scope): the scope to store var, it's should be the value of program's scope, usually it's fluid.global_scope().
        place(fluid.CPUPlace or fluid.CUDAPlace): place
        config(dict): configs for quantization, default values are in quant_config_default dict.
        for_test: if program is test program, for_test should be set True, else False.
    Return:
        fluid.Program: user can finetune this quantization program to enhance the accuracy.
    """

    scope = fluid.global_scope() if not scope else scope
    assert isinstance(config, dict), "config must be dict"

    assert 'weight_quantize_type' in config.keys(
    ), 'weight_quantize_type must be configured'
    assert 'activation_quantize_type' in config.keys(
    ), 'activation_quantize_type must be configured'

    config = _parse_configs(config)
    main_graph = IrGraph(core.Graph(program.desc), for_test=for_test)

    transform_pass = QuantizationTransformPass(
        scope=scope,
        place=place,
        weight_bits=config['weight_bits'],
        activation_bits=config['activation_bits'],
        activation_quantize_type=config['activation_quantize_type'],
        weight_quantize_type=config['weight_quantize_type'],
        window_size=config['window_size'],
        moving_rate=config['moving_rate'],
        quantizable_op_type=config['quantize_op_types'],
        skip_pattern=config['not_quant_pattern'])

    transform_pass.apply(main_graph)

    if for_test:
        quant_program = main_graph.to_program()
    else:
        quant_program = fluid.CompiledProgram(main_graph.graph)
    return quant_program


S
slf12 已提交
157 158 159 160 161 162 163 164 165 166 167
def quant_post(executor, 
               model_path,
               quantize_model_path,
               data_reader,
               batch_size=10,
               batch_nums=None,
               scope=None,
               algo='KL',
               quantizable_op_type=[
                   "conv2d", "depthwise_conv2d", 
                   "mul", "pool2d", "elementwise_add"]):
F
ftian 已提交
168
    """
S
slf12 已提交
169 170 171 172 173
    The class utilizes post training quantization methon to quantize the 
    fp32 model. It uses calibrate data to calculate the scale factor of 
    quantized variables, and inserts fake quant/dequant op to obtain the 
    quantized model.

F
ftian 已提交
174
    Args:
S
slf12 已提交
175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193
        executor(fluid.Executor): The executor to load, run and save the 
            quantized model.
        model_path(str): The path of fp32 model that will be quantized(
            load_inference_model).
        quantize_model_path(str): The path to save quantized model.
        data_reader(Reader): The data reader generates a sample every time,
                        and it provides calibrate data for DataLoader.
        batch_size(int, optional): The batch size of DataLoader, default is 10.
        batch_nums(int, optional): If set batch_nums, the number of calibrate 
                        data is batch_size*batch_nums. If batch_nums=None, use all data
                        provided by data_reader as calibrate data.
        scope(fluid.Scope, optional): The scope of the program, use it to load 
                        and save variables. If scope=None, get scope by global_scope().
        algo(str, optional): If algo=KL, use KL-divergenc method to 
                        get the more precise scale factor. If algo='direct', use 
                        abs_max methon to get the scale factor. Default is KL.
        quantizable_op_type(list[str], optional): List the type of ops 
                        that will be quantized. Default is ["conv2d", "depthwise_conv2d", 
                        "mul", "pool2d", "elementwise_add"].
F
ftian 已提交
194
    """
S
slf12 已提交
195 196 197 198 199 200 201 202 203 204 205
    post_training_quantization = PostTrainingQuantization(
            executor=executor,
            model_path=model_path,
            data_reader=data_reader,
            batch_size=batch_size,
            batch_nums=batch_nums,
            scope=scope,
            algo=algo,
            quantizable_op_type=quantizable_op_type)
    post_training_quantization.quantize()
    post_training_quantization.save_quantized_model(quantize_model_path)
F
ftian 已提交
206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242


def convert(program, scope, place, config, save_int8=False):
    """
    add quantization ops in program. the program returned is not trainable.
    Args:
        program(fluid.Program): program
        scope(fluid.Scope): the scope to store var, when is None will use fluid.global_scope()
        place(fluid.CPUPlace or fluid.CUDAPlace): place
        config(dict): configs for quantization, default values are in quant_config_default dict.
        save_int8: is export int8 freezed program.
    Return:
        fluid.Program: freezed program which can be used for inference.
                       parameters is float32 type, but it's value in int8 range.
        fluid.Program: freezed int8 program which can be used for inference.
                       if save_int8 is False, this value is None.
    """

    test_graph = IrGraph(core.Graph(program.desc), for_test=True)

    # Freeze the graph after training by adjusting the quantize
    # operators' order for the inference.
    freeze_pass = QuantizationFreezePass(
        scope=scope,
        place=place,
        weight_quantize_type=config['weight_quantize_type'])
    freeze_pass.apply(test_graph)
    freezed_program = test_graph.to_program()

    if save_int8:
        convert_int8_pass = ConvertToInt8Pass(
            scope=fluid.global_scope(), place=place)
        convert_int8_pass.apply(test_graph)
        freezed_program_int8 = test_graph.to_program()
        return freezed_program, freezed_program_int8
    else:
        return freezed_program