{"id":463,"date":"2018-03-14T17:47:46","date_gmt":"2018-03-14T09:47:46","guid":{"rendered":"http:\/\/www.max-shu.com\/blog\/?p=463"},"modified":"2018-03-14T17:49:24","modified_gmt":"2018-03-14T09:49:24","slug":"%e4%ba%ba%e5%b7%a5%e6%99%ba%e8%83%bd%e6%b7%b1%e5%ba%a6%e5%ad%a6%e4%b9%a0%e7%a5%9e%e7%bb%8f%e7%bd%91%e7%bb%9ctensorflow%e5%ba%93","status":"publish","type":"post","link":"http:\/\/www.max-shu.com\/blog\/?p=463","title":{"rendered":"\u4eba\u5de5\u667a\u80fd\u6df1\u5ea6\u5b66\u4e60\u795e\u7ecf\u7f51\u7edcTensorFlow\u5e93"},"content":{"rendered":"
TensorFlow\u662f\u4e00\u4e2a\u6df1\u5ea6\u5b66\u4e60\u7684\u5e93\uff0c\u4e0a\u5c42\u652f\u6301Python\u8bed\u8a00\uff0c\u5e95\u5c42\u7528C++\u5b9e\u73b0\u5177\u4f53\u64cd\u4f5c\u3002<\/div>\n
<\/div>\n
\u5b83\u4ee5\u56fegraph\u7684\u65b9\u5f0f\u6765\u5efa\u7acb\u64cd\u4f5c\u5e8f\u5217\uff0c\u56fe\u4e2d\u7684\u8282\u70b9\u8868\u793a\u6570\u5b66\u64cd\u4f5c\uff0c\u8fde\u7ebf\u8868\u793a\u6570\u636e\u7684\u6d41\u52a8\u65b9\u5411\u3002<\/div>\n
\"\"<\/div>\n
<\/div>\n
\u5b83\u4ee5\u5b9a\u4e49\u548c\u6267\u884c\u5206\u79bb\u7684\u65b9\u5f0f\u6765\u7f16\u7a0b\uff0c\u7528\u56fe\u6765\u5b9a\u4e49\u64cd\u4f5c\u5e8f\u5217\uff0c\u7136\u540e\u5efa\u7acbsession\u6765\u6267\u884c\uff0c\u6267\u884c\u65f6\u53ef\u4ee5\u628a\u6bcf\u4e2a\u64cd\u4f5c\u5206\u914d\u5230\u4e0d\u540c\u7684CPU\u3001GPU\u4e0a\u8fdb\u884c\uff0c\u6240\u4ee5\u901f\u5ea6\u5f88\u5feb\u3002<\/div>\n
\"\"<\/div>\n
<\/div>\n
\u5177\u4f53\u7f16\u7a0b\u6d41\u7a0b\uff1a<\/b><\/div>\n
\n
    \n
  • \u5efa\u7acb\u8ba1\u7b97\u56fe<\/li>\n
  • \u521d\u59cb\u5316\u53d8\u91cf<\/li>\n
  • \u5efa\u7acb\u4f1a\u8bdd<\/li>\n
  • \u5728\u4f1a\u8bdd\u4e2d\u6267\u884c\u56fe<\/li>\n
  • \u5173\u95ed\u4f1a\u8bdd<\/li>\n<\/ul>\n<\/div>\n
    \u00a0<\/b><\/div>\n
    \u6ce8\u610f\uff1a<\/b>\u9700\u8981\u6ce8\u610f\u4e0d\u540c\u7248\u672c\u4e4b\u95f4\u7684\u517c\u5bb9\u6027\u4e0d\u662f\u5f88\u597d\uff0c\u4e0b\u9762\u7684\u4f8b\u5b50\u5728Python3.5\u3001TensorFlow 1.2\u7248\u672c\u4e0b\u7f16\u8bd1\u6267\u884c\u901a\u8fc7\u3002<\/div>\n
    <\/div>\n
    \u7b80\u5355\u4e3e\u4f8b<\/span><\/b><\/div>\n
    \u4ee5\u4e0a\u9762\u56fe\u4e2da=(b+c)?(c+2)\u6765\u4e3e\u4f8b\uff0c\u5148\u5bfc\u5165tensorflow\u5e93\uff0c\u7136\u540e\u5b9a\u4e49\u5e38\u91cf\uff0c\u518d\u5b9a\u4e49\u53d8\u91cf\u3002<\/div>\n
    \n

    import<\/span> tensorflow <\/span>as<\/span> tf<\/span><\/p>\n

    # first, create a TensorFlow constant<\/span>
    \nconst<\/span> =<\/span> tf<\/span>.<\/span>constant<\/span>(<\/span>2.0<\/span>,<\/span> name<\/span>=<\/span>“const”<\/span>)<\/span><\/p>\n

    # create TensorFlow variables<\/span>
    \nb <\/span>=<\/span> tf<\/span>.<\/span>Variable<\/span>(<\/span>2.0<\/span>,<\/span> name<\/span>=<\/span>‘b’<\/span>)<\/span>
    \nc <\/span>=<\/span> tf<\/span>.<\/span>Variable<\/span>(<\/span>1.0<\/span>,<\/span> name<\/span>=<\/span>‘c’<\/span>)<\/span><\/p>\n<\/div>\n

    \u518d\u5efa\u7acb\u64cd\u4f5c\uff1a<\/div>\n
    \n

    # now create some operations<\/span>
    \nd <\/span>=<\/span> tf<\/span>.<\/span>add<\/span>(<\/span>b<\/span>,<\/span> c<\/span>,<\/span> name<\/span>=<\/span>‘d’<\/span>)<\/span>
    \ne <\/span>=<\/span> tf<\/span>.<\/span>add<\/span>(<\/span>c<\/span>,<\/span> const<\/span>,<\/span> name<\/span>=<\/span>‘e’<\/span>)<\/span>
    \na <\/span>=<\/span> tf<\/span>.<\/span>multiply<\/span>(<\/span>d<\/span>,<\/span> e<\/span>,<\/span> name<\/span>=<\/span>‘a’<\/span>)<\/span><\/p>\n<\/div>\n

    \u518d\u521d\u59cb\u5316\u6240\u6709\u53d8\u91cf\uff1a<\/div>\n
    \n

    # setup the variable initialisation<\/span>
    \ninit_op <\/span>=<\/span> tf<\/span>.<\/span>global_variables_initializer<\/span>()<\/span><\/p>\n<\/div>\n

    \u6700\u540e\u5efa\u7acb\u5e76\u6267\u884csession\uff1a<\/div>\n
    \n

    # start the session<\/span>
    \nwith<\/span> tf<\/span>.<\/span>Session<\/span>()<\/span> as<\/span> sess<\/span>:<\/span>
    \n\u00a0 \u00a0 # initialise the variables<\/span>
    \n\u00a0 \u00a0 sess<\/span>.<\/span>run<\/span>(<\/span>init_op<\/span>)<\/span>
    \n\u00a0 \u00a0 # compute the output of the graph<\/span>
    \n\u00a0 \u00a0 a_out <\/span>=<\/span> sess<\/span>.<\/span>run<\/span>(<\/span>a<\/span>)<\/span>
    \n\u00a0 \u00a0 print<\/span>(<\/span>“Variable a is {}”<\/span>.<\/span>format<\/span>(<\/span>a_out<\/span>))<\/span><\/p>\n<\/div>\n

    \u5177\u4f53\u6267\u884c\u8fc7\u7a0b\u5982\u4e0b\u56fe\uff1a<\/div>\n
    \"\"<\/div>\n
    \u6700\u7ec8\u7684\u6267\u884c\u7ed3\u679c\uff1a<\/div>\n
    \n

    Variable<\/span> a <\/span>is<\/span> 9.0<\/span><\/span><\/p>\n

    Process<\/span> finished <\/span>with<\/span> exit<\/span> code <\/span>0<\/span><\/p>\n<\/div>\n

    <\/div>\n
    \u795e\u7ecf\u7f51\u7edc\u4e3e\u4f8b<\/b><\/span><\/div>\n
    \u4ee5\u5e38\u89c1\u7684\u795e\u7ecf\u7f51\u7edc\u57fa\u51c6\u6d4b\u8bd5MNIST\uff08\u56fe\u7247\u4e2d\u963f\u62c9\u4f2f\u6570\u5b57\u7684\u8bc6\u522b\uff09\u4e3a\u4f8b\uff0c\u7965\u89c1\u6ce8\u91ca\uff1a<\/span><\/div>\n
    \n

    import<\/span> tensorflow <\/span>as<\/span> tf<\/span><\/p>\n

    from<\/span> tensorflow<\/span>.<\/span>examples<\/span>.<\/span>tutorials<\/span>.<\/span>mnist <\/span>import<\/span> input_data<\/span>
    \nmnist <\/span>=<\/span> input_data<\/span>.<\/span>read_data_sets<\/span>(<\/span>“MNIST_data\/”<\/span>,<\/span> one_hot<\/span>=<\/span>True<\/span>)<\/span> # \u4ece\u5b98\u65b9\u4f8b\u5b50\u5bfc\u5165\u8bad\u7ec3\u6570\u636e\u548c\u6d4b\u8bd5\u6570\u636e\uff0c\u5bfc\u5165\u540e\u6570\u636e\u4fdd\u5b58\u5728\u5f53\u524d\u76ee\u5f55\u7684\\MNIST_data\u76ee\u5f55\u4e0b\u3002<\/span><\/p>\n

    # Python optimisation variables<\/span>
    \nlearning_rate <\/span>=<\/span> 0.5<\/span>
    \nepochs <\/span>=<\/span> 10<\/span> # \u6267\u884c10\u6279\u6b21\u8bad\u7ec3<\/span>
    \nbatch_size <\/span>=<\/span> 100<\/span> # \u6bcf\u6279\u6b21100\u4e2a\u6570\u636e<\/span><\/p>\n

    # declare the training data placeholders<\/span>
    \n# input x – for 28 x 28 pixels = 784<\/span>
    \nx <\/span>=<\/span> tf<\/span>.<\/span>placeholder<\/span>(<\/span>tf<\/span>.<\/span>float32<\/span>,<\/span> [<\/span>None<\/span>,<\/span> 784<\/span>])<\/span> # \u9884\u7559\u8bad\u7ec3\u6570\u636e\u7684\u8f93\u5165\u503c<\/span>
    \n# now declare the output data placeholder – 10 digits<\/span>
    \ny <\/span>=<\/span> tf<\/span>.<\/span>placeholder<\/span>(<\/span>tf<\/span>.<\/span>float32<\/span>,<\/span> [<\/span>None<\/span>,<\/span> 10<\/span>])<\/span> # \u9884\u7559\u8bad\u7ec3\u6570\u636e\u7684\u8f93\u5165\u503c\u5bf9\u5e94\u7684\u8f93\u51fa\u503c<\/span><\/p>\n

    # now declare the weights connecting the input to the hidden layer<\/span>
    \nW1 <\/span>=<\/span> tf<\/span>.<\/span>Variable<\/span>(<\/span>tf<\/span>.<\/span>random_normal<\/span>([<\/span>784<\/span>,<\/span> 300<\/span>],<\/span> stddev<\/span>=<\/span>0.03<\/span>),<\/span> name<\/span>=<\/span>‘W1’<\/span>)<\/span> #\u5b9a\u4e49\u7b2c\u4e00\u5c42\u8f93\u5165\u5c42\u9700\u8981\u4f18\u5316\u7684\u53c2\u6570\u53d8\u91cf\uff0c\u8ba1\u7b97\u516c\u5f0f\u4e3a\uff1ay = x * W + b<\/span>
    \nb1 <\/span>=<\/span> tf<\/span>.<\/span>Variable<\/span>(<\/span>tf<\/span>.<\/span>random_normal<\/span>([<\/span>300<\/span>]),<\/span> name<\/span>=<\/span>‘b1’<\/span>)<\/span>
    \n# and the weights connecting the hidden layer to the output layer<\/span>
    \nW2 <\/span>=<\/span> tf<\/span>.<\/span>Variable<\/span>(<\/span>tf<\/span>.<\/span>random_normal<\/span>([<\/span>300<\/span>,<\/span> 10<\/span>],<\/span> stddev<\/span>=<\/span>0.03<\/span>),<\/span> name<\/span>=<\/span>‘W2’<\/span>)<\/span> #\u5b9a\u4e49\u7b2c\u4e8c\u5c42\u9690\u85cf\u5c42\u7684\u53c2\u6570\u53d8\u91cf\u3002<\/span>
    \nb2 <\/span>=<\/span> tf<\/span>.<\/span>Variable<\/span>(<\/span>tf<\/span>.<\/span>random_normal<\/span>([<\/span>10<\/span>]),<\/span> name<\/span>=<\/span>‘b2’<\/span>)<\/span><\/p>\n

    # calculate the output of the hidden layer<\/span>
    \nhidden_out <\/span>=<\/span> tf<\/span>.<\/span>add<\/span>(<\/span>tf<\/span>.<\/span>matmul<\/span>(<\/span>x<\/span>,<\/span> W1<\/span>),<\/span> b1<\/span>)<\/span>
    \nhidden_out <\/span>=<\/span> tf<\/span>.<\/span>nn<\/span>.<\/span>relu<\/span>(<\/span>hidden_out<\/span>)<\/span><\/p>\n

    # now calculate the hidden layer output – in this case, let’s use a softmax activated<\/span>
    \n# output layer<\/span>
    \ny_ <\/span>=<\/span> tf<\/span>.<\/span>nn<\/span>.<\/span>softmax<\/span>(<\/span>tf<\/span>.<\/span>add<\/span>(<\/span>tf<\/span>.<\/span>matmul<\/span>(<\/span>hidden_out<\/span>,<\/span> W2<\/span>),<\/span> b2<\/span>))<\/span> # \u8ba1\u7b97\u9690\u85cf\u5c42\u8f93\u51fa<\/span><\/p>\n

    y_clipped <\/span>=<\/span> tf<\/span>.<\/span>clip_by_value<\/span>(<\/span>y_<\/span>,<\/span> 1e-10<\/span>,<\/span> 0.9999999<\/span>)<\/span>
    \ncross_entropy <\/span>=<\/span> –<\/span>tf<\/span>.<\/span>reduce_mean<\/span>(<\/span>tf<\/span>.<\/span>reduce_sum<\/span>(<\/span>y <\/span>*<\/span> tf<\/span>.<\/span>log<\/span>(<\/span>y_clipped<\/span>)<\/span>
    \n+<\/span> (<\/span>1<\/span> –<\/span> y<\/span>)<\/span> *<\/span> tf<\/span>.<\/span>log<\/span>(<\/span>1<\/span> –<\/span> y_clipped<\/span>),<\/span> axis<\/span>=<\/span>1<\/span>))<\/span> #\u8ba1\u7b97\u6700\u7ec8\u8bc4\u4f30\u56e0\u5b50<\/span><\/p>\n

    # add an optimiser<\/span>
    \noptimiser <\/span>=<\/span> tf<\/span>.<\/span>train<\/span>.<\/span>GradientDescentOptimizer<\/span>(<\/span>learning_rate<\/span>=<\/span>learning_rate<\/span>).<\/span>minimize<\/span>(<\/span>cross_entropy<\/span>)<\/span> # \u5b9a\u4e49\u4e00\u4e2a\u5bf9\u53c2\u6570\u7684BP\u4f18\u5316\u5668<\/span><\/p>\n

    # finally setup the initialisation operator<\/span>
    \ninit_op <\/span>=<\/span> tf<\/span>.<\/span>global_variables_initializer<\/span>()<\/span> # \u521d\u59cb\u5316\u53d8\u91cf<\/span><\/p>\n

    # define an accuracy assessment operation<\/span>
    \ncorrect_prediction <\/span>=<\/span> tf<\/span>.<\/span>equal<\/span>(<\/span>tf<\/span>.<\/span>argmax<\/span>(<\/span>y<\/span>,<\/span> 1<\/span>),<\/span> tf<\/span>.<\/span>argmax<\/span>(<\/span>y_<\/span>,<\/span> 1<\/span>))<\/span>
    \naccuracy <\/span>=<\/span> tf<\/span>.<\/span>reduce_mean<\/span>(<\/span>tf<\/span>.<\/span>cast<\/span>(<\/span>correct_prediction<\/span>,<\/span> tf<\/span>.<\/span>float32<\/span>))<\/span> # \u5b9a\u4e49\u7cbe\u5ea6<\/span><\/p>\n

    # start the session<\/span>
    \nwith<\/span> tf<\/span>.<\/span>Session<\/span>()<\/span> as<\/span> sess<\/span>:<\/span> #\u5f00\u59cb\u6267\u884c<\/span>
    \n\u00a0 \u00a0 # initialise the variables<\/span>
    \n\u00a0 \u00a0 sess<\/span>.<\/span>run<\/span>(<\/span>init_op<\/span>)<\/span>
    \n\u00a0 \u00a0 total_batch <\/span>=<\/span> int<\/span>(<\/span>len<\/span>(<\/span>mnist<\/span>.<\/span>train<\/span>.<\/span>labels<\/span>)<\/span> \/<\/span> batch_size<\/span>)<\/span>
    \n\u00a0 \u00a0 for<\/span> epoch <\/span>in<\/span> range<\/span>(<\/span>epochs<\/span>):<\/span>
    \n\u00a0 \u00a0 \u00a0 \u00a0 avg_cost <\/span>=<\/span> 0<\/span>
    \n\u00a0 \u00a0 \u00a0 \u00a0 for<\/span> i <\/span>in<\/span> range<\/span>(<\/span>total_batch<\/span>):<\/span>
    \n\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 batch_x<\/span>,<\/span> batch_y <\/span>=<\/span> mnist<\/span>.<\/span>train<\/span>.<\/span>next_batch<\/span>(<\/span>batch_size<\/span>=<\/span>batch_size<\/span>)<\/span>
    \n\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 _<\/span>,<\/span> c <\/span>=<\/span> sess<\/span>.<\/span>run<\/span>([<\/span>optimiser<\/span>,<\/span> cross_entropy<\/span>],\u00a0<\/span>feed_dict<\/span>={<\/span>x<\/span>:<\/span> batch_x<\/span>,<\/span> y<\/span>:<\/span> batch_y<\/span>})<\/span>
    \n\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 avg_cost <\/span>+=<\/span> c <\/span>\/<\/span> total_batch<\/span>
    \n\u00a0 \u00a0 \u00a0 \u00a0 print<\/span>(<\/span>“Epoch:”<\/span>,<\/span> (<\/span>epoch <\/span>+<\/span> 1<\/span>),<\/span> “cost =”<\/span>,<\/span> “{:.3f}”<\/span>.<\/span>format<\/span>(<\/span>avg_cost<\/span>))<\/span>
    \n\u00a0 \u00a0 print<\/span>(<\/span>sess<\/span>.<\/span>run<\/span>(<\/span>accuracy<\/span>,<\/span> feed_dict<\/span>={<\/span>x<\/span>:<\/span> mnist<\/span>.<\/span>test<\/span>.<\/span>images<\/span>,<\/span> y<\/span>:<\/span> mnist<\/span>.<\/span>test<\/span>.<\/span>labels<\/span>}))<\/span> # \u6700\u7ec8\u7ed3\u679c\uff0c\u7cbe\u5ea6<\/span><\/p>\n<\/div>\n

    \u6267\u884c\u7ed3\u679c\uff0c\u53ef\u4ee5\u770b\u5230\u5f88\u4e0d\u9519\uff0c\u7cbe\u5ea6\u67090.9742\uff081.0\u4f4d\u4e3a\u7406\u60f3\u503c\uff09\uff1a<\/div>\n
    \n

    Epoch<\/span>:<\/span> 1<\/span> cost <\/span>=<\/span> 0.768<\/span>
    \nEpoch<\/span>:<\/span> 2<\/span> cost <\/span>=<\/span> 0.245<\/span>
    \nEpoch<\/span>:<\/span> 3<\/span> cost <\/span>=<\/span> 0.183<\/span>
    \nEpoch<\/span>:<\/span> 4<\/span> cost <\/span>=<\/span> 0.152<\/span>
    \nEpoch<\/span>:<\/span> 5<\/span> cost <\/span>=<\/span> 0.125<\/span>
    \nEpoch<\/span>:<\/span> 6<\/span> cost <\/span>=<\/span> 0.108<\/span>
    \nEpoch<\/span>:<\/span> 7<\/span> cost <\/span>=<\/span> 0.090<\/span>
    \nEpoch<\/span>:<\/span> 8<\/span> cost <\/span>=<\/span> 0.078<\/span>
    \nEpoch<\/span>:<\/span> 9<\/span> cost <\/span>=<\/span> 0.068<\/span>
    \nEpoch<\/span>:<\/span> 10<\/span> cost <\/span>=<\/span> 0.060<\/span>
    \n0.9742<\/span><\/span><\/p>\n

    Process<\/span> finished <\/span>with<\/span> exit<\/span> code <\/span>0<\/span><\/p>\n<\/div>\n","protected":false},"excerpt":{"rendered":"

    TensorFlow\u662f\u4e00\u4e2a\u6df1\u5ea6\u5b66\u4e60\u7684\u5e93\uff0c\u4e0a\u5c42\u652f\u6301Python\u8bed\u8a00\uff0c\u5e95\u5c42\u7528C++\u5b9e\u73b0\u5177\u4f53\u64cd\u4f5c\u3002 \u5b83\u4ee5\u56fegraph …<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[443],"tags":[108,444,445,446,447],"class_list":["post-463","post","type-post","status-publish","format-standard","hentry","category-443","tag-python","tag-tensorflow","tag-445","tag-446","tag-447"],"views":1644,"_links":{"self":[{"href":"http:\/\/www.max-shu.com\/blog\/index.php?rest_route=\/wp\/v2\/posts\/463","targetHints":{"allow":["GET"]}}],"collection":[{"href":"http:\/\/www.max-shu.com\/blog\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/www.max-shu.com\/blog\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/www.max-shu.com\/blog\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/www.max-shu.com\/blog\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=463"}],"version-history":[{"count":3,"href":"http:\/\/www.max-shu.com\/blog\/index.php?rest_route=\/wp\/v2\/posts\/463\/revisions"}],"predecessor-version":[{"id":469,"href":"http:\/\/www.max-shu.com\/blog\/index.php?rest_route=\/wp\/v2\/posts\/463\/revisions\/469"}],"wp:attachment":[{"href":"http:\/\/www.max-shu.com\/blog\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=463"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/www.max-shu.com\/blog\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=463"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/www.max-shu.com\/blog\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=463"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}