H2O deep learning model results with dropout scaled down - r

Question

I am having the following figure when training an H2O Deep Learning model with dropout
Misaligned predictions
The code used to train the net is
m.nn <- h2o.deeplearning(x = 1:(nc-1),
y = nc,
training_frame = datTra,
#validation_frame = datTst,
nfolds = 5,
activation = 'RectifierWithDropout',
#input_dropout_ratio = 0.2,
hidden_dropout_ratios = c(dro, dro, dro),
hidden = c(120,30,8),
#hidden = 20,
epochs = 999,
#mini_batch_size = 100,
#variable_importances = TRUE,
standardize = TRUE,
regression_stop = 1e-3,
stopping_metric = "MSE",
stopping_tolerance = 1e-6,
stopping_rounds = 10)
Figure corresponds to dro=0.1
Why am I having that misalignment? Is there any option i am missing out?
You can find a piece of code to try below
(download 'SampleData.csv' from here)
library(h2o)
library(readr)
library(ggplot2)
df <- as.data.frame(read_delim(file = 'SampleData.csv', delim = ";"))
localH2O <- h2o.init(ip = "localhost", startH2O = TRUE, nthreads = 2, max_mem_size = '4g')
dat_h2o <- as.h2o(x = df)
model.ref <- h2o.deeplearning(x = 1:(ncol(df)-1), y = ncol(df),
training_frame = dat_h2o,
hidden = c(120,30,8),
activation = 'Rectifier',
epochs = 199,
mini_batch_size = 10,
regression_stop = 0.1,
stopping_metric = "MSE",
stopping_tolerance = 1e-6,
stopping_rounds = 10)
model.dro <- h2o.deeplearning(x = 1:(ncol(df)-1), y = ncol(df),
training_frame = dat_h2o,
hidden = c(120,30,8),
activation = 'RectifierWithDropout',
hidden_dropout_ratios = c(0.2, 0.2, 0.2),
epochs = 199,
mini_batch_size = 10,
regression_stop = 0.1,
stopping_metric = "MSE",
stopping_tolerance = 1e-6,
stopping_rounds = 10)
pred.ref <- as.data.frame(h2o.predict(object = model.ref, newdata = dat_h2o))
pred.dro <- as.data.frame(h2o.predict(object = model.dro, newdata = dat_h2o))
dfRes <- data.frame(cbind(df$SeqF, pred.ref$predict, pred.dro$predict))
colnames(dfRes) <- c('act', 'pred', 'pred2')
ggplot(data = dfRes) + geom_point(aes(x=act, y=pred), color='blue') +
geom_point(aes(x=act, y=pred2), color='red') + geom_abline()