网上看到实验结果best_acc 94.71%14层残差天啊要训练110轮才会有质的飞跃我似乎缺乏这样的耐心so我就停留在85困住了近期最好成绩84.95带senet也取得最好成绩83.18创造这些最好成绩 的还是我那三残差网络只不过我优化了很多细节最近的要算一些参数的0初始化感受还是在稳定性上另外就是std和mean的使用第三就是lr作了调整近期最大的感受就是稳定是突破的基石网络架构还是这个得分最高就在这上头想办法layers.emplace_back(std::make_sharedConv2D(cudnn, batch, 5, 32, 32, 32, 3, 1, 1));layers.emplace_back(std::make_sharedBN(cudnn, batch, 32, 32, 32));layers.emplace_back(std::make_sharedLeakyRL(cudnn, batch, 32, 32, 32));layers.emplace_back(std::make_sharedConv2D(cudnn, batch, 32, 64, 32, 32, 3, 1, 1));layers.emplace_back(std::make_sharedresidualExt22(cudnn, batch, 64, 32, 32));layers.emplace_back(std::make_sharedMaxPool2D(cudnn, batch, 64, 32, 32, 2, 2, 0, 2));layers.emplace_back(std::make_sharedConv2D(cudnn, batch, 64, 128, 16, 16, 3, 1, 1));layers.emplace_back(std::make_sharedresidualExt22(cudnn, batch, 128, 16, 16));layers.emplace_back(std::make_sharedMaxPool2D(cudnn, batch, 128, 16, 16, 2, 2, 0, 2));layers.emplace_back(std::make_sharedConv2D(cudnn, batch, 128, 256, 8, 8, 3, 1, 1));layers.emplace_back(std::make_sharedresidualExt22(cudnn, batch, 256, 8, 8));layers.emplace_back(std::make_sharedMaxPool2D(cudnn, batch, 256, 8, 8, 2, 2, 0, 2));// layers.emplace_back(std::make_sharedConv2D(cudnn, batch, 256, 300, 4, 4, 4));layers.emplace_back(std::make_sharedLinear(cublas, batch, 256*4*4, 300));layers.emplace_back(std::make_sharedBN(cudnn, batch, 300, 1, 1));layers.emplace_back(std::make_sharedLeakyRL(cudnn, batch, 300, 1, 1));layers.emplace_back(std::make_sharedLinear(cublas, batch, 300, 10));最初就是他带我上了83.24分lr调整后如此一直用lr0.001当train到达82分做如下部署if (chengji[0] 82)// if (chengji[0] 85){起作用;lr 0.0001;if (起作用 3)//执行2次{lr 0.00001;//这个83.32分创纪录了if (起作用 5)//执行2次{lr 0.000001;if (起作用 7)//执行2次{lr 0.0000001;if (起作用 9)//退出{i 100;}}}}}0初始化如卷积层不再使用非零初始化zero_kernel (b_size 255) / 256, 256 (_grad_bias, b_size);zero_kernel (w_size 255) / 256, 256 (_grad_weight, w_size);zero_kernel (_batch * _out_channels * _out_h * _out_w 255) / 256, 256 (_output, _batch * _out_channels * _out_h * _out_w);zero_kernel (_batch * _in_channels * _in_h * _in_w 255) / 256, 256 (_grad_input, _batch * _in_channels * _in_h * _in_w);训练节点拐点如下(验证以上细节的调整)时间: 27331.929688 mstrain Classification result: 97.50% ok (used 49984 images)时间: 2145.236084 msTest Classification result: 82.30% ok (used 9984 images)learn rate0.0001轮次24均值: -90.0030059814,方差:14.0249986649均值: -87.6609115601,方差:11.1409282684均值: -91.4050750732,方差:8.6625871658均值: -88.7697906494,方差:13.2397594452均值: -89.5710449219,方差:13.7450332642均值: -90.1831817627,方差:7.1279768944均值: -89.8954620361,方差:15.9475574493均值: -91.6007232666,方差:19.2678642273均值: -92.5446090698,方差:11.0975723267均值: -89.4845657349,方差:10.7211999893均值: -89.1439590454,方差:9.4157590866均值: -91.4487915039,方差:24.4504737854均值: -88.1308135986,方差:13.0858469009均值: -91.8805236816,方差:15.4056463242均值: -91.4631423950,方差:13.5508584976均值: -92.0959930420,方差:15.5257358551均值: -89.3520660400,方差:7.0764660835均值: -89.7229003906,方差:13.0267095566均值: -90.1392745972,方差:11.5007438660均值: -89.8574676514,方差:15.2681083679均值: -94.0932998657,方差:12.7880811691均值: -91.9308776855,方差:12.9652538300均值: -89.5239181519,方差:19.1181278229均值: -88.7568893433,方差:17.2236022949均值: -89.1051712036,方差:18.5954170227均值: -89.2031173706,方差:11.1493854523均值: -90.2933273315,方差:15.7165937424均值: -92.2265319824,方差:17.1336555481均值: -88.4247970581,方差:14.2746744156均值: -87.0833587646,方差:8.2819252014均值: -92.9659652710,方差:12.1669368744均值: -88.0147323608,方差:12.5158348083时间: 27362.111328 mstrain Classification result: 98.10% ok (used 49984 images)时间: 2152.024902 msTest Classification result:83.86% ok (used 9984 images)learn rate1e-05轮次25。。。时间: 27535.343750 mstrain Classification result: 97.72% ok (used 49984 images)时间: 2174.266113 msTest Classification result: 81.56% ok (used 9984 images)learn rate0.0001轮次24rb均值: 1.6140453815,rb方差:5.120904445648均值: -93.1453933716,方差:16.6951541901均值: -89.5411300659,方差:13.9249677658均值: -91.6143341064,方差:10.9254045486均值: -88.4864883423,方差:17.3460960388均值: -92.7311325073,方差:16.6858062744均值: -90.4656372070,方差:6.9966602325均值: -91.2220077515,方差:20.6955375671均值: -92.4978637695,方差:28.9232978821均值: -95.8759613037,方差:15.1841077805均值: -88.2818069458,方差:13.0533695221均值: -84.8050689697,方差:11.1592187881均值: -90.9559707642,方差:26.2159500122均值: -90.2095413208,方差:21.0307292938均值: -94.9991760254,方差:18.5383968353均值: -88.7679290771,方差:15.3791704178均值: -94.6937179565,方差:24.2541046143均值: -87.7234725952,方差:7.5897364616均值: -89.1307373047,方差:18.1437034607均值: -89.9150695801,方差:13.0529365540均值: -92.5738677979,方差:17.3618869781均值: -95.2071914673,方差:20.6090621948均值: -92.9055862427,方差:11.8797712326均值: -91.4030151367,方差:24.5985851288均值: -90.6616897583,方差:18.5187110901均值: -89.8819732666,方差:19.0706024170均值: -86.7467422485,方差:11.9170856476均值: -88.7020034790,方差:14.6680479050均值: -93.5892105103,方差:21.0500984192均值: -89.5009918213,方差:15.7073564529均值: -88.4498672485,方差:9.8511447906均值: -92.2534332275,方差:11.2799034119均值: -89.2119445801,方差:14.4264163971时间: 27513.873047 mstrain Classification result: 98.23% ok (used 49984 images)时间: 2177.084961 msTest Classification result:84.59% ok (used 9984 images)learn rate1e-05轮次25。。。时间: 27585.908203 mstrain Classification result: 97.12% ok (used 49984 images)时间: 2171.240967 msTest Classification result: 84.87% ok (used 9984 images)learn rate1e-05轮次24均值: -89.1569442749,方差:21.8877449036均值: -87.1171264648,方差:18.0932846069均值: -90.3414459229,方差:14.8263645172均值: -88.9296646118,方差:21.0110893250均值: -91.0140609741,方差:23.3084144592均值: -89.8222656250,方差:13.7696933746均值: -89.2569656372,方差:24.2851676941均值: -91.3925323486,方差:20.1694049835均值: -90.6773681641,方差:14.6891994476均值: -89.3500137329,方差:13.7939796448均值: -82.7095794678,方差:12.1913690567均值: -90.9668884277,方差:34.9290504456均值: -88.0935821533,方差:23.3173236847均值: -89.6331863403,方差:29.2646942139均值: -86.8041687012,方差:19.7180480957均值: -93.5874633789,方差:28.8533210754均值: -87.7150115967,方差:10.0918579102均值: -85.8822174072,方差:13.8620977402均值: -86.9596557617,方差:22.7990703583均值: -89.3087921143,方差:27.7217559814均值: -91.4010162354,方差:24.4388103485均值: -90.3860702515,方差:15.0780963898均值: -88.7416839600,方差:31.9720840454均值: -88.3724517822,方差:23.9765300751均值: -87.4666137695,方差:22.2937946320均值: -87.9431304932,方差:18.5732898712均值: -86.9548416138,方差:23.0292987823均值: -91.6506729126,方差:26.3199367523均值: -87.4069747925,方差:19.6056785583均值: -87.0255661011,方差:10.2156801224均值: -91.9443283081,方差:20.2376461029均值: -85.8004989624,方差:14.2750272751时间: 27673.296875 mstrain Classification result: 97.24% ok (used 49984 images)时间: 2174.523926 msTest Classification result:84.92% ok (used 9984 images)learn rate1e-06轮次25以前还有耐心训练110轮现在感觉25轮都出成绩了还要110轮吗自己的这个架构在110轮能突破吗感觉在磨人性