deeplearning模型分析

deeplearning模型分析

FLOPs

paddleslim.analysis.flops(programdetail=False)

获得指定网络的浮点运算次数(FLOPs)。

参数:

  • program(paddle.fluid.Program) - 待分析的目标网络。更多关于Program的介绍请参考:Program概念介绍
  • detail(bool) - 是否返回每个卷积层的FLOPs。默认为False。
  • only_conv(bool) - 如果设置为True,则仅计算卷积层和全连接层的FLOPs,即浮点数的乘加(multiplication-adds)操作次数。如果设置为False,则也会计算卷积和全连接层之外的操作的FLOPs。

返回值:

  • flops(float) - 整个网络的FLOPs。
  • params2flops(dict) - 每层卷积对应的FLOPs,其中key为卷积层参数名称,value为FLOPs值。

示例:

import paddle.fluid as fluid
from paddle.fluid.param_attr import ParamAttr
from paddleslim.analysis import flops
 
def conv_bn_layer(input,
                  num_filters,
                  filter_size,
                  name,
                  stride=1,
                  groups=1,
                  act=None):
    conv = fluid.layers.conv2d(
        input=input,
        num_filters=num_filters,
        filter_size=filter_size,
        stride=stride,
        padding=(filter_size - 1) // 2,
        groups=groups,
        act=None,
        param_attr=ParamAttr(name=name + "_weights"),
        bias_attr=False,
        name=name + "_out")
    bn_name = name + "_bn"
    return fluid.layers.batch_norm(
        input=conv,
        act=act,
        name=bn_name + '_output',
        param_attr=ParamAttr(name=bn_name + '_scale'),
        bias_attr=ParamAttr(bn_name + '_offset'),
        moving_mean_name=bn_name + '_mean',
        moving_variance_name=bn_name + '_variance', )
 
main_program = fluid.Program()
startup_program = fluid.Program()
#   X       X              O       X              O
# conv1-->conv2-->sum1-->conv3-->conv4-->sum2-->conv5-->conv6
#     |            ^ |                    ^
#     |____________| |____________________|
#
# X: prune output channels
# O: prune input channels
with fluid.program_guard(main_program, startup_program):
    input = fluid.data(name="image", shape=[None, 3, 16, 16])
    conv1 = conv_bn_layer(input, 8, 3, "conv1")
    conv2 = conv_bn_layer(conv1, 8, 3, "conv2")
    sum1 = conv1 + conv2
    conv3 = conv_bn_layer(sum1, 8, 3, "conv3")
    conv4 = conv_bn_layer(conv3, 8, 3, "conv4")
    sum2 = conv4 + sum1
    conv5 = conv_bn_layer(sum2, 8, 3, "conv5")
    conv6 = conv_bn_layer(conv5, 8, 3, "conv6")
 
print("FLOPs: {}".format(flops(main_program)))

model_size

paddleslim.analysis.model_size(program)

获得指定网络的参数数量。

参数:

  • program(paddle.fluid.Program) - 待分析的目标网络。更多关于Program的介绍请参考:Program概念介绍

返回值:

  • model_size(int) - 整个网络的参数数量。

示例:

import paddle.fluid as fluid
from paddle.fluid.param_attr import ParamAttr
from paddleslim.analysis import model_size
 
def conv_layer(input,
                  num_filters,
                  filter_size,
                  name,
                  stride=1,
                  groups=1,
                  act=None):
    conv = fluid.layers.conv2d(
        input=input,
        num_filters=num_filters,
        filter_size=filter_size,
        stride=stride,
        padding=(filter_size - 1) // 2,
        groups=groups,
        act=None,
        param_attr=ParamAttr(name=name + "_weights"),
        bias_attr=False,
        name=name + "_out")
    return conv
 
main_program = fluid.Program()
startup_program = fluid.Program()
#   X       X              O       X              O
# conv1-->conv2-->sum1-->conv3-->conv4-->sum2-->conv5-->conv6
#     |            ^ |                    ^
#     |____________| |____________________|
#
# X: prune output channels
# O: prune input channels
with fluid.program_guard(main_program, startup_program):
    input = fluid.data(name="image", shape=[None, 3, 16, 16])
    conv1 = conv_layer(input, 8, 3, "conv1")
    conv2 = conv_layer(conv1, 8, 3, "conv2")
    sum1 = conv1 + conv2
    conv3 = conv_layer(sum1, 8, 3, "conv3")
    conv4 = conv_layer(conv3, 8, 3, "conv4")
    sum2 = conv4 + sum1
    conv5 = conv_layer(sum2, 8, 3, "conv5")
    conv6 = conv_layer(conv5, 8, 3, "conv6")
 
print("FLOPs: {}".format(model_size(main_program)))

TableLatencyEvaluator

classpaddleslim.analysis.TableLatencyEvaluator(table_filedelimiter="")

基于硬件延时表的模型延时评估器。

参数:

  • table_file(str) - 所使用的延时评估表的绝对路径。关于演示评估表格式请参考:PaddleSlim硬件延时评估表格式
  • delimiter(str) - 在硬件延时评估表中,操作信息之前所使用的分割符,默认为英文字符逗号。

返回值:

  • Evaluator - 硬件延时评估器的实例。

latency(graph)

获得指定网络的预估延时。

参数:

  • graph(Program) - 待预估的目标网络。

返回值:

  • latency - 目标网络的预估延时。
上一篇:apache POI 导出excel相关方法


下一篇:(小组)第六次作业:NABCD模型分析。产品Backlog。