backscatter coefficient 后向散射系数 极化
test Epoch 21 train_loss:0.14493, test_loss 0.19278, best_test_loss 0.19278, accuracy 90.34483 0.3309
test Epoch 34, lr: 0.00000 best_test_loss 0.19360, test_accuracy 86.20690, train_loss:0.18985, test_loss 0.19360 0.2705
test Epoch 43, lr: 0.00000 best_test_loss 0.16528, test_accuracy 93.10345, train_loss:0.25169, test_loss 0.16528 0.1945 rotate is awesome?
test Epoch 37, lr: 0.00000 best_test_loss 0.22386, test_accuracy 91.03448, train_loss:0.27736, test_loss 0.22386 0.2854
test Epoch 29, lr: 0.00000 best_test_loss 0.19650, test_accuracy 88.96552, train_loss:0.31724, test_loss 0.19650 0.2399
test Epoch 99, lr: 0.00156250 best_test_loss 0.15662, test_accuracy 86.89655, train_loss:0.17001, test_loss 0.32298 0.1978
test Epoch 99, lr: 0.00004883 best_test_loss 0.15658, test_accuracy 87.58621, train_loss:0.17496, test_loss 0.33669
test Epoch 99, lr: 0.00039063 best_test_loss 0.15901, test_accuracy 80.68966, train_loss:0.40502, test_loss 0.48465
test Epoch 99, lr: 0.00004883 best_test_loss 0.15368, test_accuracy 85.51724, train_loss:0.20466, test_loss 0.54482
过拟合？？test Epoch 99, lr: 0.00019531 best_test_loss 0.13829, test_accuracy 91.72414, train_loss:0.14624, test_loss 0.21511 0.2354
test Epoch 99, lr: 0.00039063 best_test_loss 0.15874, test_accuracy 86.89655, train_loss:0.10851, test_loss 0.51559
test Epoch 99, lr: 0.00002441 best_test_loss 0.13738, test_accuracy 86.89655, train_loss:0.13139, test_loss 0.61632
test Epoch 99, lr: 0.00039063 best_test_loss 0.23778, test_accuracy 83.56164, train_loss:0.18233, test_loss 0.28511
test Epoch 99, lr: 0.00625000 best_test_loss 0.22963, test_accuracy 86.30137, train_loss:0.23748, test_loss 0.34087
average test loss:0.204781
average test loss:0.223527
average test loss:0.216351
average test loss:0.272495
average test loss:0.235534
average test loss:0.273817
average test loss:0.235803
average test loss:0.214413
average test loss:0.237254
average test loss:0.171246 average test loss:0.205263 average test loss:0.205097, average train loss:.0.181294 线上0.2127
average test loss:0.233942 average test loss:0.208232, average train loss:.0.153128
average test loss:0.192417, average train loss:.0.206058 average test loss:0.185343, average train loss:.0.203189 average test loss:0.236775, average train loss:.0.201195 结果太随机了
average test loss:0.223221, average train loss:.0.202647
average test loss:0.233810, average train loss:.0.200464
average test loss:0.233780, average train loss:.0.076921
average test loss:0.209401, average train loss:.0.068871 stack
fold 0, Epoch 22, lr: 0.01000000 best_test_loss 0.67068, train_loss:0.76683, test_loss 0.85430
average test loss:0.196119, average train loss:.0.211353 线上 0.1960
在sigmoid处添加噪声 使输出更趋向于二值输出，参见deep learning的autoencoder average test loss:0.199828, average train loss:.0.085958 线上 0.1608 what happended??
是否如论文中所说，有了bn层之后，Bias理论上确实多余，而dropout呢
看不出，不过发现亮点位置变化很大，大小也变化很大
average test loss:0.233655, average train loss:.0.050323
average test loss:0.228653, average train loss:.0.051385
average test loss:0.221123, average train loss:.0.060002
average test loss:0.222726, average train loss:.0.149231
average test loss:0.232851, average train loss:.0.064902 average test loss:0.240076, average train loss:.0.007134
average test loss:0.238063, average train loss:.0.021351
average test loss:0.194955, average train loss:.0.110455 数据增强有用？
average test loss:0.183261, average train loss:.0.097679 线上0.1932 0.1997
average test loss:0.204144, average train loss:.0.181775 线上 0.1729 average test loss:0.214501, average train loss:.0.100599 stack
average test loss:0.183260, average train loss:.0.092487
average test loss:0.324347, average train loss:.0.210049 有的好，有的差 average test loss:0.341616, average train loss:.0.258920 是真的不好
average test loss:0.219151, average train loss:.0.043743 线上 0.1828 不起作用？
average test loss:0.262968, average train loss:.0.121245 线上 0.1999 也不起作用
average test loss:0.205066, average train loss:.0.044473
average test loss:0.240381, average train loss:.0.039588
average test loss:0.219999, average train loss:.0.054592
average test loss:0.180112, average train loss:.0.095212 不crop更好
average test loss:0.237561, average train loss:.0.012233 过拟合严重，应该着手找去除过拟合的方法，而不是继续新网络的尝试
average test loss:0.246844, average train loss:.0.010427
average test loss:0.221798, average train loss:.0.109564
average test loss:0.220774, average train loss:.0.076189
average test loss:0.297814, average train loss:.0.061193 过拟合很严重
average test loss:0.159343, average train loss:.0.024036 损失震荡很厉害
average test loss:0.207845, average train loss:.0.090437 归一化后效果和不加相同
average test loss:0.164736, average train loss:.0.070337 震荡太严重
average test loss:0.212593, average train loss:.0.074800 更糟糕
有可能训练根本就不会收敛？average test loss:0.150993, average train loss:.0.053462 确实很好，比使用均值填充震荡幅度小
average test loss:0.225410, average train loss:.0.036134
发散average test loss:0.373767, average train loss:.0.251022 不好
average test loss:0.232106, average train loss:.0.092780
看pytorch默认的初始化方法是什么 xavier-uniform
average test loss:0.206330, average train loss:.0.221600
average test loss:0.182640, average train loss:.0.088378 好一点
average test loss:0.181852, average train loss:.0.115198
average test loss:0.194272, average train loss:.0.064499
average test loss:0.201961, average train loss:.0.081981
average test loss:0.148888, average train loss:.0.074843 加大权重衰减可以减轻过拟合 average test loss:0.179505, average train loss:.0.074203 结果还是不太稳定，有一个0.13 线上0.1861
average test loss:0.208659, average train loss:.0.089634 有一个不行
average test loss:0.173561, average train loss:.0.077523 线上 0.2195
average test loss:0.228612, average train loss:.0.202333
average test loss:0.247735, average train loss:.0.149113
average test loss:0.275267, average train loss:.0.185022
average test loss:0.233755, average train loss:.0.178425
average test loss:0.254026, average train loss:.0.149570
average test loss:0.355445, average train loss:.0.066586 crop就过拟合
average test loss:0.266209, average train loss:.0.181895
loss不同，输出的分布也不同？average test loss:0.242483, average train loss:.0.180673
average test loss:0.252653, average train loss:.0.172899?
average test loss:0.242949, average train loss:.0.169888
average test loss:0.238308, average train loss:.0.178249
average test loss:0.232003, average train loss:.0.077536
average test loss:0.224093, average train loss:.0.185493 线上 0.222
average test loss:0.238502, average train loss:.0.183955
average test loss:0.201659, average train loss:.0.100682
average test loss:0.214560, average train loss:.0.105047
average test loss:0.426968, average train loss:.0.335662
average test loss:0.418621, average train loss:.0.132600
average test loss:0.409492, average train loss:.0.137095
average test loss:0.368851, average train loss:.0.151549
average test loss:0.393786, average train loss:.0.111288
average test loss:0.371570, average train loss:.0.126198
average test loss:0.353302, average train loss:.0.113255
average test loss:0.372462, average train loss:.0.110490
average test loss:0.369433, average train loss:.0.123947
average test loss:0.354283, average train loss:.0.146388
average test loss:0.377611, average train loss:.0.099520
average test loss:0.388231, average train loss:.0.141772
average test loss:0.412088, average train loss:.0.292780
average test loss:0.337630, average train loss:.0.125791 可以
average test loss:0.369177, average train loss:.0.022962
average test loss:0.363353, average train loss:.0.200878
average test loss:0.340855, average train loss:.0.187562
average test loss:0.398085, average train loss:.0.345799
不收敛average test loss:0.317586, average train loss:.0.135877 厉害 average test loss:0.366345, average train loss:.0.126594 结果很随机
average test loss:0.348928, average train loss:.0.136806
average test loss:0.390565, average train loss:.0.132121
average test loss:0.358188, average train loss:.0.133969
average test loss:0.348081, average train loss:.0.134984
average test loss:0.337618, average train loss:.0.145867
average test loss:0.307901, average train loss:.0.173047 average test loss:0.317608, average train loss:.0.159305 average test loss:0.390299, average train loss:.0.300357
average test loss:0.365219, average train loss:.0.245349
average test loss:0.379116, average train loss:.0.318695
average test loss:0.342046, average train loss:.0.189783
average test loss:0.353239, average train loss:.0.299657
average test loss:0.376087, average train loss:.0.318330
average test loss:0.388909, average train loss:.0.298582
不靠谱average test loss:0.408406, average train loss:.0.330217
average test loss:0.324028, average train loss:.0.270278 average test loss:0.307347, average train loss:.0.240791 average test loss:0.343076, average train loss:.0.243124
average test loss:0.307148, average train loss:.0.192440 average test loss:0.305122, average train loss:.0.190761
average test loss:0.326102, average train loss:.0.228217
average test loss:0.080338, average train loss:.0.015431 average test loss:0.044228, average train loss:.0.007685
average test loss:0.078167, average train loss:.0.024920 average test loss:0.073738, average train loss:.0.016546
folder:0 best test loss:0.03260 best train loss:0.02173 folder:1 best test loss:0.05762 best train loss:0.01672 folder:2 best test loss:0.00517 best train loss:0.01049 folder:3 best test loss:0.02128 best train loss:0.02577 folder:4 best test loss:0.11475 best train loss:0.01656 average test loss:0.046285, average train loss:.0.018255 线上 0.208
线上 0.1535 线上 0.1446 convNet(stn)+getNet 线上 0.1341 resModel+convNet(stn)+getNet 线上 0.1347 +laternel 0.1626 线上 0.1423 +outsModel