插入符号：如何设置自定义模型 deepnet

Question

我想使用原deepnet包的一些参数，所以我设置了一个自定义模型。我阅读了 Caret 的文档 (Custom Model)，但它不起作用。

这是我设置自定义模型的代码：

# set up custom model
dnn_custom <- list(label = "Stacked AutoEncoder Deep Neural Network",
                  library = "deepnet",
                  loop = NULL,
                  type = c("Classification", "Regression"),
                  parameters = data.frame(parameter = c("layer1", "layer2", "layer3", "hidden_dropout", "visible_dropout", "learningrate"),
                                          class = rep("numeric", 6),
                                          label = c("Hidden Layer 1", "Hidden Layer 2", "Hidden Layer 3", 
                                                    "Hidden Dropouts", "Visible Dropout", "Learning rate")),
                  grid = function(x, y, len = NULL, search = "grid") {
                    if(search == "grid") {
                      out <- expand.grid(layer1 = 1:len, layer2 = 0:(len -1), layer3 = 0:(len -1),
                                         hidden_dropout = seq(0, .7, length = len), 
                                         visible_dropout = seq(0, .7, length = len),
                                         learningrate = seq(0.0001,1, length = len))
                    } else {
                      out <- data.frame(layer1 = sample(2:20, replace = TRUE, size = len),
                                        layer2 = sample(2:20, replace = TRUE, size = len),
                                        layer3 = sample(2:20, replace = TRUE, size = len),
                                        hidden_dropout = runif(len, min = 0, max = .7),
                                        visible_dropout = runif(len, min = 0, max = .7),
                                        learningrate = runif(len, min = 0.0001, max = 1))
                    }
                    out
                  },
                  fit = function(x, y, wts, param, lev, last, classProbs, ...) {
                    if(!is.matrix(x)) x <- as.matrix(x)
                    is_class <- is.factor(y)
                    if (is_class) y <- caret:::class2ind(y)
                    layers <- c(param$layer1, param$layer2, param$layer3)
                    layers <- layers[layers > 0]
                    deepnet::sae.dnn.train(x, y, hidden = layers,
                                           output = if(is_class) "sigm" else "linear",
                                           hidden_dropout = param$hidden_dropout,
                                           visible_dropout = param$visible_dropout,
                                           learningrate = param$learningrate,
                                           ...)
                  },
                  predict = function(modelFit, newdata, submodels = NULL) {
                    pred <- deepnet::nn.predict(modelFit, as.matrix(newdata))
                    if(ncol(pred) > 1)
                      pred <- modelFit$obsLevels[apply(pred, 1, which.max)]
                    pred
                  },
                  prob = function(modelFit, newdata, submodels = NULL) {
                    out <- exp(deepnet::nn.predict(modelFit, as.matrix(newdata)))
                    out <- apply(out, 1, function(x) x/sum(x))
                    t(out)
                  },
                  predictors = function(x, ...) {
                    NULL
                  },
                  varImp = NULL,
                  levels = function(x) x$classes,
                  tags = c("Neural Network"),
                  sort = function(x) x[order(x[,1]),])

这里我尝试使用模型：

library(caret)
library(MASS)
library(deepnet)

set.seed(132, kind = "Mersenne-Twister", normal.kind = "Inversion")

# create data
dat <- Boston

preds <- dat[1:100, -which(names(dat) == "medv")]
response <- dat[1:100, which(names(dat) == "medv")]

# specifiy trainControl for tuning mtry and with specified folds
control <- caret::trainControl(search = "grid", method="repeatedcv", number=3, 
                               repeats=2, 
                               savePred = T)

# tune hyperparameters and build final model
tunegrid <- expand.grid(layer1 = c(50,100),
                        layer2 = c(0,50,100), 
                        layer3 = c(0,50,100),
                        hidden_dropout = c(0, 0.1), 
                        visible_dropout = c(0, 0.1),
                        learningrate = c(0.9, 1)) 

model <- caret::train(x = preds,
                      y = response, 
                      method="dnn_custom",
                      metric= "RMSE", 
                      tuneGrid=tunegrid, 
                      trControl= control,
                      preProc = c("scale", "center") 
)

我收到错误 dnn_custom is not in caret's built-in library。我做错了什么？

Answer 1

我自己找到了答案...

这是一个简单的错误：我在应用自定义模型时不得不删除 method 中的引号：

 model <- caret::train(x = preds,
                  y = response, 
                  method = dnn_custom,          # No quotation marks here!
                  metric = "RMSE", 
                  tuneGrid =tunegrid, 
                  trControl = control,
                  preProc = c("scale", "center")

)

现在一切正常。引号似乎仅指 Caret 中的内置模型。

插入符号：如何设置自定义模型 deepnet

Caret: How to set up custom model deepnet

r

deep-learning

r-caret