search

PaddlePaddle / Paddle

PArallel Distributed Deep LEarning: Machine Learning Framework from Industrial Practice （『飞桨』核心框架，深度学习&机器学习高性能单机、分布式训练和跨平台部署）
http://www.paddlepaddle.org/
Apache License 2.0
22.25k stars 5.59k forks source link

Paddle新API数据读取性能问题 #1616

Closed reyoung closed 7 years ago

reyoung commented 7 years ago

刚刚做了一个benchmark，在mnist的api_train_v2.py里面，一共运行了19秒，有16秒都是在数据转换的地方。

那块的性能值得优化一下。

Tue Mar 14 16:47:32 2017    profile.bin

         3191874 function calls (3182864 primitive calls) in 19.872 seconds

   Ordered by: cumulative time

   ncalls  tottime  percall  cumtime  percall filename:lineno(function)
        1    0.001    0.001   19.890   19.890 api_train_v2.py:1(<module>)
        1    0.002    0.002   18.307   18.307 api_train_v2.py:58(main)
        1    0.065    0.065   18.222   18.222 /Users/baidu/tmp/.py_env/lib/python2.7/site-packages/paddle/v2/trainer.py:64(train)
     1097    0.003    0.000   16.342    0.015 /Users/baidu/tmp/.py_env/lib/python2.7/site-packages/py_paddle/dataprovider_converter.py:174(__call__)
     1097    0.077    0.000   16.339    0.015 /Users/baidu/tmp/.py_env/lib/python2.7/site-packages/paddle/v2/data_feeder.py:91(convert)
     1097    0.297    0.000   16.077    0.015 /Users/baidu/tmp/.py_env/lib/python2.7/site-packages/py_paddle/dataprovider_converter.py:154(convert)
   140100    2.678    0.000   15.458    0.000 /Users/baidu/tmp/.py_env/lib/python2.7/site-packages/py_paddle/dataprovider_converter.py:55(scan)
   139003    0.154    0.000   12.766    0.000 /Users/baidu/tmp/.py_env/lib/python2.7/site-packages/numpy/lib/function_base.py:4951(append)
   139006   12.488    0.000   12.488    0.000 {numpy.core.multiarray.concatenate}
        1    0.005    0.005    1.582    1.582 /Users/baidu/tmp/.py_env/lib/python2.7/site-packages/paddle/v2/__init__.py:14(<module>)
     1103    0.073    0.000    1.136    0.001 /Users/baidu/tmp/.py_env/lib/python2.7/site-packages/paddle/v2/minibatch.py:30(batch_reader)
        1    0.004    0.004    1.032    1.032 /Users/baidu/tmp/.py_env/lib/python2.7/site-packages/paddle/v2/dataset/__init__.py:16(<module>)
        1    0.001    0.001    0.941    0.941 /Users/baidu/tmp/.py_env/lib/python2.7/site-packages/paddle/v2/dataset/sentiment.py:21(<module>)
        1    0.015    0.015    0.940    0.940 /Users/baidu/tmp/.py_env/lib/python2.7/site-packages/nltk/__init__.py:17(<module>)
   120002    0.122    0.000    0.922    0.000 /Users/baidu/tmp/.py_env/lib/python2.7/site-packages/paddle/v2/reader/decorator.py:65(data_reader)
   140105    0.545    0.000    0.834    0.000 /Users/baidu/tmp/.py_env/lib/python2.7/site-packages/paddle/v2/dataset/mnist.py:38(reader)
     1880    0.004    0.000    0.765    0.000 api_train_v2.py:92(event_handler)
        2    0.009    0.005    0.739    0.370 /Users/baidu/tmp/.py_env/lib/python2.7/site-packages/paddle/v2/trainer.py:125(test)
        1    0.002    0.002    0.517    0.517 /Users/baidu/tmp/.py_env/lib/python2.7/site-packages/paddle/v2/optimizer.py:1(<module>)
        1    0.001    0.001    0.481    0.481 /Users/baidu/tmp/.py_env/lib/python2.7/site-packages/py_paddle/__init__.py:15(<module>)
        1    0.003    0.003    0.478    0.478 /Users/baidu/tmp/.py_env/lib/python2.7/site-packages/py_paddle/util.py:16(<module>)
   309/30    0.022    0.000    0.452    0.015 {__import__}
        1    0.003    0.003    0.445    0.445 /Users/baidu/tmp/.py_env/lib/python2.7/site-packages/nltk/collocations.py:25(<module>)
        1    0.004    0.004    0.434    0.434 /Users/baidu/tmp/.py_env/lib/python2.7/site-packages/nltk/metrics/__init__.py:14(<module>)
        1    0.002    0.002    0.428    0.428 /Users/baidu/tmp/.py_env/lib/python2.7/site-packages/nltk/metrics/scores.py:8(<module>)
typhoonzero commented 7 years ago

data_feader.py 中 convert函数中定义的子函数可以定义在module中：http://stackoverflow.com/questions/7839632/is-there-an-overhead-when-nesting-functions-in-python