I want to train a regression model by "keras_model_sequential" in R. For finding the best parameters over the grid search I have used "tuning_run". But I got this error:
training run 1/128 (flags = list(0.05, 66, "relu", 8, 10, 0.001, 0.2))
Error in sink(file = output_file, type = "output", split = TRUE) :
sink stack is full
Calls: tuning_run ... with_changed_file_copy -> force -> sink -> .handleSimpleError -> h
I need to mention, that a folder named "runs" was created in the data and script path which has a lot of subfolders whose name is like a date format. maybe this is the reason.
This is my code
library(plyr)
library(boot)
library(keras)
library(tensorflow)
library(kerasR)
library(tidyverse)
library(tfruns)
library(MLmetrics)
df= mainlist[[1]] # data which is a 33*31 dataframe (33 samples and 31 features which last column is target)
x = (length(df))-1
print(x)
df1 = df[, 2:x]
#normalization
df2 = df[, 2:length(df)]
normalize <- function(x) {
return ((x - min(x)) / (max(x) - min(x)))
}
maxmindf <- as.data.frame(lapply(df2, normalize))
attach(maxmindf)
df_norm<-as.matrix(maxmindf)
# Determine sample size
ind <- sample(2, nrow(df_norm), replace=TRUE, prob=c(0.80, 0.20))
# Split the data(peaks)
training <- df_norm[ind==1, 1:ncol(df_norm)-1]
test1 <- df_norm[ind==2, 1:ncol(df_norm)-1]
training_target <- df_norm[ind==1, ncol(df_norm)]
test1_target <- df_norm[ind==2, ncol(df_norm)]
#number of nodes in the first hidden layer
u1_1 = ceiling((1/2) * (ncol(training)+1))
u2_1 = ceiling(1* (ncol(training)+1))
u3_1 = ceiling((2/3) * (ncol(training)+1))
u4_1 = ceiling(2*(ncol(training)))
####a) Declaring the flags for hyperparameters
FLAGS = flags(
flag_numeric("dropout1", 0.05),
flag_integer("units",u1_1),
flag_string("activation1", "relu"),
flag_integer("batchsize1",8),
flag_integer("Epoch1",50),
flag_numeric("learning_rate", 0.01),
flag_numeric("val_split",0.2),
flag_numeric("reg_l1",0.001)
)
# ####b) Defining the DNN model
build_model<-function() {
model <- keras_model_sequential()
model %>%
layer_dense(units = FLAGS$units, activation = FLAGS$activation1, input_shape = c(dim(training)[2])) %>%
layer_dropout(rate = FLAGS$dropout1) %>%
layer_dense(units=1, activation ="linear")
#####c) Compiling the DNN model
model %>% compile(
loss = "mse",
optimizer =optimizer_adam(FLAGS$learning_rate),
metrics = c("mse"))
model
}
model<-build_model()
model %>% summary()
print_dot_callback <- callback_lambda(
on_epoch_end = function(epoch, logs) {
if (epoch %% 80 == 0) cat("\n")
cat(".")})
early_stop <- callback_early_stopping(monitor = "val_loss", mode='min',patience =20)
###########d) Fitting the DNN model#################
model_Final<-build_model()
model_fit_Final<-model_Final %>% fit(
training,
training_target,
epochs =FLAGS$Epoch1, batch_size = FLAGS$batchsize1,
shuffled=F,
validation_split = FLAGS$val_split,
verbose=0,
callbacks = list(early_stop, print_dot_callback)
)
################a) Inner cross-validation##########################
nCVI=5
Hyperpar = data.frame() #the results of each combination of hyperparameters resulting from each inner partition will be saved
for (i in 1:nCVI){ #do it to choose best parameters
print("I is:")
print(i)
Sam_per=sample(1:nrow(training),nrow(training))
X_trII=training[Sam_per,]
y_trII=training_target[Sam_per]
# print(head(X_trII, 3))
print("----------------------")
print(head(y_trII,3))
############b) Grid search using the tuning_run() function of tfruns package########
runs.sp<-tuning_run(paste0("train.R")
,runs_dir = '_tuningE1'
,flags=list(dropout1 = c(0,0.05),
units = c(u1_1, u2_1),
activation1 = ("relu"),
batchsize1 = c(8, 16),
Epoch1 = c(10,50),
learning_rate = c(0.001),
val_split = c(0.2)),
sample = 0.2,
confirm = FALSE,
echo =F)
# clean_runs(ls_runs(completed == FALSE))
#####c) Saving each combination of hyperparameters in the Hyperpar data.frame
runs.sp = runs.sp[order(runs.sp$flag_units,runs.sp$flag_dropout1, runs.sp$flag_batchsize1, runs.sp$flag_Epoch1),]
runs.sp$grid_length = 1:nrow(runs.sp) #we save the grid lenght and also important parameters
Parameters = data.frame(grid_length=runs.sp$grid_length,
metric_val_mse=runs.sp$metric_val_mse,
flag_dropout1=runs.sp$flag_dropout1,
flag_units=runs.sp$flag_units,
flag_batchsize1=runs.sp$flag_batchsize1,
epochs_completed=runs.sp$epochs_completed,
flag_learning_rate=runs.sp$flag_learning_rate,
flag_activation1=runs.sp$flag_activation1)
Hyperpar = rbind(Hyperpar,data.frame(Parameters)) #we saved the important parameters
}
#####d) Summarizing the five inner fold by hyperparameter combination
#the average prediction performance is obtained for each inner fold
Hyperpar %>%
group_by(grid_length) %>%
summarise(val_mse=mean(metric_val_mse),
dropout1=mean(flag_dropout1),
units=mean(flag_units),
batchsize1=mean(flag_batchsize1),
learning_rate=mean(flag_learning_rate),
epochs=mean( epochs_completed)) %>%
select(grid_length,val_mse,dropout1,units,batchsize1,
learning_rate, epochs) %>%
mutate_if(is.numeric, funs(round(., 3)))
Hyperpar_Opt = Hyperpar
######e) ############ select the best combinition of hyperparameters
Min = min(Hyperpar_Opt$val_mse)
pos_opt = which(Hyperpar_Opt$val_mse==Min)
pos_opt=pos_opt[1]
Optimal_Hyper=Hyperpar_Opt[pos_opt,]
#####Selecting the best hyperparameters
Drop_O = Optimal_Hyper$dropout1
Epoch_O = round(Optimal_Hyper$epochs,0)
Units_O = round(Optimal_Hyper$units,0)
activation_O = unique(Hyperpar$flag_activation1)
batchsize_O = round(Optimal_Hyper$batchsize1,0)
lr_O = Optimal_Hyper$learning_rate
print_dot_callback <- callback_lambda(
on_epoch_end = function(epoch, logs) {
if (epoch %% 20 == 0) cat("\n")
cat(".")})
#refitting the model with optimal values
model_Sec<-keras_model_sequential()
model_Sec %>%
layer_dense(units =Units_O , activation =activation_O, input_shape =
c(dim(training)[2])) %>%
layer_dropout(rate =Drop_O) %>%
layer_dense(units =1, activation =activation_O)
model_Sec %>% compile(
loss = "mean_squared_error",
optimizer = optimizer_adam(lr=lr_O),
metrics = c("mean_squared_error"))
# fit the model with our data
ModelFited<-model_Sec %>% fit(
X_trII, y_trII,
epochs=Epoch_O, batch_size =batchsize_O, #####validation_split=0.2,
early_stop,
verbose=0
,callbacks=list(print_dot_callback)
)
#############g) Prediction of testing set ##########################
Yhat=model_Sec%>% predict(test1)
y_p=Yhat
y_p_tst =as.numeric(y_p)
#y_tst=y[tst_set]
plot(test1_target,y_p_tst)
MSE=mean((test1_target - y_p_tst)^2)
How can I fix the error?
Thanks in advance