平时用惯了两个线性层结尾layers.emplace_back(std::make_sharedLinear(cublas, batch, 128 * 64, 500));layers.emplace_back(std::make_sharedLeakyRL(cudnn, batch, 500, 1, 1));layers.emplace_back(std::make_sharedLinear(cublas, batch, 500, 10));想起以前cpu下自己写的全连接方式可以试一试代替一次linear层layers.emplace_back(std::make_sharedConv2D(cudnn, batch, 256,300, 4, 4, 4, 1));即把256*4*4的数据通过4*4卷积成300*1*1数据昨天架构增加了一个残差块接近82分但翻不过去今天用这个4*4卷积成功突破闯关成功架构如下layers.emplace_back(std::make_sharedConv2D(cudnn, batch, 5, 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, 1));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));得分如下rb均值: 0.8797101974,rb方差:5.678951263428rb均值: -8.3503847122,rb方差:27.435239791870rb均值: 1.8536375761,rb方差:24.807447433472rb均值: 4.2101426125,rb方差:16.878156661987rb均值: -3.8696627617,rb方差:5.029622554779rb均值: -12.3703460693,rb方差:29.529457092285rb均值: -2.0291111469,rb方差:8.738946914673rb均值: -3.2213602066,rb方差:6.971586704254rb均值: -3.2548468113,rb方差:14.214648246765时间: 24602.048828 mstrain Classification result: 96.43% ok (used 49984 images)时间: 1960.859985 msTest Classification result:82.07% ok (used 9984 images)learn rate0.0001轮次16rb均值: 0.8008456230,rb方差:5.773095607758rb均值: -8.5729656219,rb方差:27.482904434204rb均值: 1.7287417650,rb方差:25.392423629761rb均值: 4.1368126869,rb方差:16.951080322266rb均值: -3.8994140625,rb方差:5.067879676819rb均值: -12.6236686707,rb方差:30.174797058105rb均值: -2.1486938000,rb方差:8.853424072266rb均值: -3.2606005669,rb方差:7.034667015076rb均值: -3.3508059978,rb方差:14.302834510803时间: 24564.140625 mstrain Classification result: 97.57% ok (used 49984 images)时间: 1951.093018 msTest Classification result:81.98% ok (used 9984 images)learn rate0.0001轮次17rb均值: 0.7735943794,rb方差:5.724789142609rb均值: -8.6432456970,rb方差:27.800550460815rb均值: 1.7252777815,rb方差:25.578634262085rb均值: 4.1205477715,rb方差:17.033407211304rb均值: -3.9014611244,rb方差:5.074986934662rb均值: -12.6995344162,rb方差:30.028085708618rb均值: -2.1647324562,rb方差:8.935814857483rb均值: -3.2678849697,rb方差:7.184212207794rb均值: -3.3836572170,rb方差:14.296825408936时间: 24613.261719 mstrain Classification result: 98.22% ok (used 49984 images)时间: 1946.271973 msTest Classification result:82.05% ok (used 9984 images)learn rate1e-05轮次18rb均值: 0.7709599733,rb方差:5.753355026245rb均值: -8.6881046295,rb方差:27.816370010376rb均值: 1.7303854227,rb方差:25.849842071533rb均值: 4.1074061394,rb方差:17.110939025879rb均值: -3.9176218510,rb方差:5.036983966827rb均值: -12.7407503128,rb方差:30.236000061035rb均值: -2.1752357483,rb方差:8.898382186890rb均值: -3.2971680164,rb方差:7.075766563416rb均值: -3.3980994225,rb方差:14.323348999023时间: 24574.095703 mstrain Classification result: 98.41% ok (used 49984 images)时间: 1955.838013 msTest Classification result:82.37% ok (used 9984 images)learn rate1e-05轮次19rb均值: 0.7680509090,rb方差:5.763551712036rb均值: -8.7100591660,rb方差:28.046590805054rb均值: 1.7029346228,rb方差:25.735727310181rb均值: 4.1084098816,rb方差:17.077848434448rb均值: -3.9339706898,rb方差:5.079279422760rb均值: -12.7671136856,rb方差:30.269668579102rb均值: -2.1830737591,rb方差:8.921528816223rb均值: -3.3042364120,rb方差:7.090327262878rb均值: -3.4057095051,rb方差:14.365463256836时间: 24811.109375 mstrain Classification result: 98.65% ok (used 49984 images)时间: 1964.407959 msTest Classification result:82.57% ok (used 9984 images)learn rate1e-05轮次20请按任意键继续. . .看样子还能上大于5的方差也控制的不错基本在30以内学习率如下lr 0.001;if (chengji[0] 85)//train score{起作用;lr 0.0001;if (起作用 5){lr 0.00001;//这个78.32分创纪录了if(起作用 8){i 100;//退出}}}越往后越有挑战
