MXNet | 在R语言中使用
亚马逊将MXNet指定为官方深度学习平台，1月23日MXNet成为Apache的卵化项目。
无疑，这些将MXNet推向深度学习的热潮中，成为热捧的项目。当然，学习MXNet也是很有必要的。哈哈，加油深度学习。
目前支持以下的语言：
-
R
-
C++
-
Julia
-
Scala
这里介绍基于R语言的安装和基本使用：
安装
1 install.packages("drat", repos="https://cran.rstudio.com") 2 drat:::addRepo("dmlc") 3 install.packages("mxnet")
若是安装过程中有问题，可以去https://cran.rstudio.com下载drat的本地文件”drat.zip”
在 https://cran.r-project.org/web/packages/drat/下载。
分类
下面以一个二分类的数据为例：
1 >require(mlbench) 2 >require(mxnet) 3 >data(Sonar, package="mlbench") 4 > str(Sonar) 5 'data.frame': 208 obs. of 61 variables: 6 $ V1 : num 0.02 0.0453 0.0262 0.01 0.0762 0.0286 0.0317 0.0519 0.0223 0.0164 ... 7 $ V2 : num 0.0371 0.0523 0.0582 0.0171 0.0666 0.0453 0.0956 0.0548 0.0375 0.0173 ... 8 $ V3 : num 0.0428 0.0843 0.1099 0.0623 0.0481 ... 9 $ V4 : num 0.0207 0.0689 0.1083 0.0205 0.0394 ... 10 $ V5 : num 0.0954 0.1183 0.0974 0.0205 0.059 ... 11 $ V6 : num 0.0986 0.2583 0.228 0.0368 0.0649 ... 12 $ V7 : num 0.154 0.216 0.243 0.11 0.121 ... 13 $ V8 : num 0.16 0.348 0.377 0.128 0.247 ... 14 $ V9 : num 0.3109 0.3337 0.5598 0.0598 0.3564 ... 15 $ V10 : num 0.211 0.287 0.619 0.126 0.446 ... 16 $ V11 : num 0.1609 0.4918 0.6333 0.0881 0.4152 ... 17 $ V12 : num 0.158 0.655 0.706 0.199 0.395 ... 18 $ V13 : num 0.2238 0.6919 0.5544 0.0184 0.4256 ... 19 $ V14 : num 0.0645 0.7797 0.532 0.2261 0.4135 ... 20 $ V15 : num 0.066 0.746 0.648 0.173 0.453 ... 21 $ V16 : num 0.227 0.944 0.693 0.213 0.533 ... 22 $ V17 : num 0.31 1 0.6759 0.0693 0.7306 ... 23 $ V18 : num 0.3 0.887 0.755 0.228 0.619 ... 24 $ V19 : num 0.508 0.802 0.893 0.406 0.203 ... 25 $ V20 : num 0.48 0.782 0.862 0.397 0.464 ... 26 $ V21 : num 0.578 0.521 0.797 0.274 0.415 ... 27 $ V22 : num 0.507 0.405 0.674 0.369 0.429 ... 28 $ V23 : num 0.433 0.396 0.429 0.556 0.573 ... 29 $ V24 : num 0.555 0.391 0.365 0.485 0.54 ... 30 $ V25 : num 0.671 0.325 0.533 0.314 0.316 ... 31 $ V26 : num 0.641 0.32 0.241 0.533 0.229 ... 32 $ V27 : num 0.71 0.327 0.507 0.526 0.7 ... 33 $ V28 : num 0.808 0.277 0.853 0.252 1 ... 34 $ V29 : num 0.679 0.442 0.604 0.209 0.726 ... 35 $ V30 : num 0.386 0.203 0.851 0.356 0.472 ... 36 $ V31 : num 0.131 0.379 0.851 0.626 0.51 ... 37 $ V32 : num 0.26 0.295 0.504 0.734 0.546 ... 38 $ V33 : num 0.512 0.198 0.186 0.612 0.288 ... 39 $ V34 : num 0.7547 0.2341 0.2709 0.3497 0.0981 ... 40 $ V35 : num 0.854 0.131 0.423 0.395 0.195 ... 41 $ V36 : num 0.851 0.418 0.304 0.301 0.418 ... 42 $ V37 : num 0.669 0.384 0.612 0.541 0.46 ... 43 $ V38 : num 0.61 0.106 0.676 0.881 0.322 ... 44 $ V39 : num 0.494 0.184 0.537 0.986 0.283 ... 45 $ V40 : num 0.274 0.197 0.472 0.917 0.243 ... 46 $ V41 : num 0.051 0.167 0.465 0.612 0.198 ... 47 $ V42 : num 0.2834 0.0583 0.2587 0.5006 0.2444 ... 48 $ V43 : num 0.282 0.14 0.213 0.321 0.185 ... 49 $ V44 : num 0.4256 0.1628 0.2222 0.3202 0.0841 ... 50 $ V45 : num 0.2641 0.0621 0.2111 0.4295 0.0692 ... 51 $ V46 : num 0.1386 0.0203 0.0176 0.3654 0.0528 ... 52 $ V47 : num 0.1051 0.053 0.1348 0.2655 0.0357 ... 53 $ V48 : num 0.1343 0.0742 0.0744 0.1576 0.0085 ... 54 $ V49 : num 0.0383 0.0409 0.013 0.0681 0.023 0.0264 0.0507 0.0285 0.0777 0.0092 ... 55 $ V50 : num 0.0324 0.0061 0.0106 0.0294 0.0046 0.0081 0.0159 0.0178 0.0439 0.0198 ... 56 $ V51 : num 0.0232 0.0125 0.0033 0.0241 0.0156 0.0104 0.0195 0.0052 0.0061 0.0118 ... 57 $ V52 : num 0.0027 0.0084 0.0232 0.0121 0.0031 0.0045 0.0201 0.0081 0.0145 0.009 ... 58 $ V53 : num 0.0065 0.0089 0.0166 0.0036 0.0054 0.0014 0.0248 0.012 0.0128 0.0223 ... 59 $ V54 : num 0.0159 0.0048 0.0095 0.015 0.0105 0.0038 0.0131 0.0045 0.0145 0.0179 ... 60 $ V55 : num 0.0072 0.0094 0.018 0.0085 0.011 0.0013 0.007 0.0121 0.0058 0.0084 ... 61 $ V56 : num 0.0167 0.0191 0.0244 0.0073 0.0015 0.0089 0.0138 0.0097 0.0049 0.0068 ... 62 $ V57 : num 0.018 0.014 0.0316 0.005 0.0072 0.0057 0.0092 0.0085 0.0065 0.0032 ... 63 $ V58 : num 0.0084 0.0049 0.0164 0.0044 0.0048 0.0027 0.0143 0.0047 0.0093 0.0035 ... 64 $ V59 : num 0.009 0.0052 0.0095 0.004 0.0107 0.0051 0.0036 0.0048 0.0059 0.0056 ... 65 $ V60 : num 0.0032 0.0044 0.0078 0.0117 0.0094 0.0062 0.0103 0.0053 0.0022 0.004 ... 66 $ Class: num 1 1 1 1 1 1 1 1 1 1 ... 67 > dim(Sonar) 68 [1] 208 61 69 70 >Sonar[,61] = as.numeric(Sonar[,61])-1 71 >train.ind = c(1:50, 100:150) 72 >train.x = data.matrix(Sonar[train.ind, 1:60]) 73 >train.y = Sonar[train.ind, 61] 74 >test.x = data.matrix(Sonar[-train.ind, 1:60]) 75 >test.y = Sonar[-train.ind, 61]
采用mx.mlp函数来训练模型，这里给出对该函数参数的介绍：
1 mx.mlp(data, label, hidden_node = 1, out_node, dropout = NULL, 2 activation = "tanh", out_activation = "softmax", 3 device = mx.ctx.default(), ...)
data :为数据
label :标签
hidden_node: 隐藏层节点数，默认为1
out_node ： 输出成节点数
dropout ： 丢失率，为[0,1)
activation : 激活函数
out_activation：输出成激活函数，默认为softmax
device = mx.ctx.default() : 这里用于设置是GPU还是CPU来训练
其他的参数 ： 用于传递给 mx.model.FeedForward.create
eval_metric ： 评估矩阵
1 #设置种子 2 > mx.set.seed(0) 3 # 模型，num.round表示迭代数，array.batch.size表示批规模 4 > model <- mx.mlp(train.x, train.y, hidden_node=10, out_node=2,out_activation="softmax", num.round=20, array.batch.size=15, learning.rate=0.07, momentum=0.9, eval.metric=mx.metric.accuracy) 5 6 Start training with 1 devices 7 [1] Train-accuracy=0.488888888888889 8 [2] Train-accuracy=0.514285714285714 9 [3] Train-accuracy=0.514285714285714 10 [4] Train-accuracy=0.514285714285714 11 [5] Train-accuracy=0.514285714285714 12 [6] Train-accuracy=0.523809523809524 13 [7] Train-accuracy=0.619047619047619 14 [8] Train-accuracy=0.695238095238095 15 [9] Train-accuracy=0.695238095238095 16 [10] Train-accuracy=0.761904761904762 17 [11] Train-accuracy=0.828571428571429 18 [12] Train-accuracy=0.771428571428571 19 [13] Train-accuracy=0.742857142857143 20 [14] Train-accuracy=0.733333333333333 21 [15] Train-accuracy=0.771428571428571 22 [16] Train-accuracy=0.847619047619048 23 [17] Train-accuracy=0.857142857142857 24 [18] Train-accuracy=0.838095238095238 25 [19] Train-accuracy=0.838095238095238 26 [20] Train-accuracy=0.838095238095238 27 28 > preds = predict(model, test.x) 29 > pred.label = max.col(t(preds))-1 30 > table(pred.label, test.y) 31 test.y 32 pred.label 0 1 33 0 24 14 34 1 36 33 35 > (24+33) / (24+14+36+33) 36 [1] 0.5327103
准确率= 0.5327103
完整的代码：
1 ###########分类 2 require(mlbench) 3 require(mxnet) 4 data(Sonar, package="mlbench") 5 Sonar[,61] = as.numeric(Sonar[,61])-1 6 train.ind = c(1:50, 100:150) 7 train.x = data.matrix(Sonar[train.ind, 1:60]) 8 train.y = Sonar[train.ind, 61] 9 test.x = data.matrix(Sonar[-train.ind, 1:60]) 10 test.y = Sonar[-train.ind, 61] 11 12 13 mx.set.seed(0) 14 model <- mx.mlp(train.x, train.y, hidden_node=10, out_node=2,out_activation="softmax", num.round=20, array.batch.size=15, learning.rate=0.07, momentum=0.9, eval.metric=mx.metric.accuracy) 15 16 17 preds = predict(model, test.x) 18 19 pred.label = max.col(t(preds))-1 20 table(pred.label, test.y) 21 22 (24+33) / (24+14+36+33)
回归
下面给一个回归的模型例子：
数据的介绍
1 data(BostonHousing, package="mlbench") 2 3 > str(BostonHousing) 4 'data.frame': 506 obs. of 14 variables: 5 $ crim : num 0.00632 0.02731 0.02729 0.03237 0.06905 ... 6 $ zn : num 18 0 0 0 0 0 12.5 12.5 12.5 12.5 ... 7 $ indus : num 2.31 7.07 7.07 2.18 2.18 2.18 7.87 7.87 7.87 7.87 ... 8 $ chas : Factor w/ 2 levels "0","1": 1 1 1 1 1 1 1 1 1 1 ... 9 $ nox : num 0.538 0.469 0.469 0.458 0.458 0.458 0.524 0.524 0.524 0.524 ... 10 $ rm : num 6.58 6.42 7.18 7 7.15 ... 11 $ age : num 65.2 78.9 61.1 45.8 54.2 58.7 66.6 96.1 100 85.9 ... 12 $ dis : num 4.09 4.97 4.97 6.06 6.06 ... 13 $ rad : num 1 2 2 3 3 3 5 5 5 5 ... 14 $ tax : num 296 242 242 222 222 222 311 311 311 311 ... 15 $ ptratio: num 15.3 17.8 17.8 18.7 18.7 18.7 15.2 15.2 15.2 15.2 ... 16 $ b : num 397 397 393 395 397 ... 17 $ lstat : num 4.98 9.14 4.03 2.94 5.33 ... 18 $ medv : num 24 21.6 34.7 33.4 36.2 28.7 22.9 27.1 16.5 18.9 ... 19 > dim(BostonHousing) 20 [1] 506 14
数据的划分，训练集合验证集
1 > train.ind = seq(1, 506, 3) 2 > train.x = data.matrix(BostonHousing[train.ind, -14]) 3 > train.y = BostonHousing[train.ind, 14] 4 > test.x = data.matrix(BostonHousing[-train.ind, -14]) 5 > test.y = BostonHousing[-train.ind, 14]
定义节点之间的连接方式
1 > # 定义输入数据 2 > data <- mx.symbol.Variable("data") 3 > # 完整连接的隐藏层 4 > # data: 输入源 5 > # num_hidden: 该层的节点数 6 > fc1 <- mx.symbol.FullyConnected(data, num_hidden=1)
定义损失函数
1 # 针对回归任务，定义损失函数 2 > lro <- mx.symbol.LinearRegressionOutput(fc1)
模型
1 >mx.set.seed(0) 2 #模型的定义，基于cpu训练，迭代50次，每批次为20，学习率为2e-6，评估矩阵为相对误差 3 >model <- mx.model.FeedForward.create(lro, X=train.x, y=train.y, ctx=mx.cpu(), num.round=50, array.batch.size=20, learning.rate=2e-6, momentum=0.9, eval.metric=mx.metric.rmse) 4 5 6 Start training with 1 devices 7 [1] Train-rmse=16.063282524034 8 [2] Train-rmse=12.2792375712573 9 [3] Train-rmse=11.1984634005885 10 [4] Train-rmse=10.2645236892904 11 [5] Train-rmse=9.49711005504284 12 [6] Train-rmse=9.07733734175182 13 [7] Train-rmse=9.07884450847991 14 [8] Train-rmse=9.10463850277417 15 [9] Train-rmse=9.03977049028532 16 [10] Train-rmse=8.96870685004475 17 [11] Train-rmse=8.93113287361574 18 [12] Train-rmse=8.89937257821847 19 [13] Train-rmse=8.87182096922953 20 [14] Train-rmse=8.84476075083586 21 [15] Train-rmse=8.81464673014974 22 [16] Train-rmse=8.78672567900196 23 [17] Train-rmse=8.76265872846474 24 [18] Train-rmse=8.73946101419974 25 [19] Train-rmse=8.71651926303267 26 [20] Train-rmse=8.69457600919277 27 [21] Train-rmse=8.67354928674563 28 [22] Train-rmse=8.65328755392436 29 [23] Train-rmse=8.63378039680078 30 [24] Train-rmse=8.61488162586984 31 [25] Train-rmse=8.5965105183022 32 [26] Train-rmse=8.57868133563275 33 [27] Train-rmse=8.56135851937663 34 [28] Train-rmse=8.5444819772098 35 [29] Train-rmse=8.52802114610432 36 [30] Train-rmse=8.5119504512622 37 [31] Train-rmse=8.49624261719241 38 [32] Train-rmse=8.48087453238701 39 [33] Train-rmse=8.46582689119887 40 [34] Train-rmse=8.45107881002491 41 [35] Train-rmse=8.43661331401712 42 [36] Train-rmse=8.42241575909639 43 [37] Train-rmse=8.40847217331365 44 [38] Train-rmse=8.39476931796395 45 [39] Train-rmse=8.38129658373974 46 [40] Train-rmse=8.36804269059018 47 [41] Train-rmse=8.35499817678397 48 [42] Train-rmse=8.34215505742154 49 [43] Train-rmse=8.32950441908131 50 [44] Train-rmse=8.31703985777311 51 [45] Train-rmse=8.30475363906755 52 [46] Train-rmse=8.29264031506106 53 [47] Train-rmse=8.28069372820073 54 [48] Train-rmse=8.26890902770415 55 [49] Train-rmse=8.25728089053853 56 [50] Train-rmse=8.24580511500735
或自定义评估矩阵
1 #自定义评估矩阵 2 >demo.metric.mae <- mx.metric.custom("mae", function(label, pred) { 3 res <- mean(abs(label-pred)) 4 return(res) 5 })
模型结果
> mx.set.seed(0) > model <- mx.model.FeedForward.create(lro, X=train.x, y=train.y, ctx=mx.cpu(), num.round=50, array.batch.size=20, learning.rate=2e-6, momentum=0.9, eval.metric=demo.metric.mae) Start training with 1 devices [1] Train-mae=13.1889538083225 [2] Train-mae=9.81431959337658 [3] Train-mae=9.21576419870059 [4] Train-mae=8.38071537613869 [5] Train-mae=7.45462437611487 [6] Train-mae=6.93423301743136 [7] Train-mae=6.91432357016537 [8] Train-mae=7.02742733055105 [9] Train-mae=7.00618194618469 [10] Train-mae=6.92541576984028 [11] Train-mae=6.87530243690643 [12] Train-mae=6.84757369098564 [13] Train-mae=6.82966501611388 [14] Train-mae=6.81151759574811 [15] Train-mae=6.78394182841811 [16] Train-mae=6.75914719419347 [17] Train-mae=6.74180388773481 [18] Train-mae=6.725853071279 [19] Train-mae=6.70932178215848 [20] Train-mae=6.6928868798746 [21] Train-mae=6.6769521329138 [22] Train-mae=6.66184809505939 [23] Train-mae=6.64754504809777 [24] Train-mae=6.63358514060577 [25] Train-mae=6.62027640889088 [26] Train-mae=6.60738245232238 [27] Train-mae=6.59505546771818 [28] Train-mae=6.58346195800437 [29] Train-mae=6.57285477783945 [30] Train-mae=6.56259003960424 [31] Train-mae=6.5527790788975 [32] Train-mae=6.54353428422991 [33] Train-mae=6.5344172368447 [34] Train-mae=6.52557652526432 [35] Train-mae=6.51697905850079 [36] Train-mae=6.50847898812758 [37] Train-mae=6.50014844106303 [38] Train-mae=6.49207674844397 [39] Train-mae=6.48412070125341 [40] Train-mae=6.47650500999557 [41] Train-mae=6.46893867486053 [42] Train-mae=6.46142131653097 [43] Train-mae=6.45395035048326 [44] Train-mae=6.44652914123403 [45] Train-mae=6.43916216409869 [46] Train-mae=6.43183777381976 [47] Train-mae=6.42455544223388 [48] Train-mae=6.41731406417158 [49] Train-mae=6.41011292926139 [50] Train-mae=6.40312503493494
完整的代码：
1 ########回归 2 data(BostonHousing, package="mlbench") 3 str(BostonHousing) 4 dim(BostonHousing) 5 train.ind = seq(1, 506, 3) 6 train.x = data.matrix(BostonHousing[train.ind, -14]) 7 train.y = BostonHousing[train.ind, 14] 8 test.x = data.matrix(BostonHousing[-train.ind, -14]) 9 test.y = BostonHousing[-train.ind, 14] 10 11 12 # 定义输入数据 13 data <- mx.symbol.Variable("data") 14 # 完整连接的隐藏层 15 # data: 输入源 16 # num_hidden: 该层的节点数 17 fc1 <- mx.symbol.FullyConnected(data, num_hidden=1) 18 19 # 针对回归任务，定义损失函数 20 lro <- mx.symbol.LinearRegressionOutput(fc1) 21 22 23 24 mx.set.seed(0) 25 model <- mx.model.FeedForward.create(lro, X=train.x, y=train.y, ctx=mx.cpu(), num.round=50, array.batch.size=20, learning.rate=2e-6, momentum=0.9, eval.metric=mx.metric.rmse) 26 27 28 29 demo.metric.mae <- mx.metric.custom("mae", function(label, pred) { 30 res <- mean(abs(label-pred)) 31 return(res) 32 }) 33 mx.set.seed(0) 34 model <- mx.model.FeedForward.create(lro, X=train.x, y=train.y, ctx=mx.cpu(), num.round=50, array.batch.size=20, learning.rate=2e-6, momentum=0.9, eval.metric=demo.metric.mae)
总结： 代码很直接明白，简介清晰，容易上手
转自：http://blog.csdn.net/xxzhangx/article/details/54729055
----------------------------------------------------------------------------------
数据和特征决定了效果上限，模型和算法决定了逼近这个上限的程度
----------------------------------------------------------------------------------
