Directly below is an example that works followed by ones that don't.
library(keras)
# placeholder data
Y <- data.frame(y1=1:100,y2=1:100)
X <- data.frame(x1=1:100,x2=1:00,x3=1:100)
# add covariates
input <- layer_input(shape=dim(X)[2],name="covars")
# add hidden layers
base_model <- input %>%
layer_dense(units = 3, activation='relu') %>%
layer_dense(units = 2, activation='relu')
# add outputs
y1 <- base_model %>%
layer_dense(units = 1, name="y1")
y2 <- base_model %>%
layer_dense(units = 1, name="y2")
# combine
model <- keras_model(input,list(y1,y2))
This is simple case where there are only two outputs. What about the case with many outputs and you don't want to script each one like I did above for y1 and y2? This adds the outputs in a loop:
# add outputs in loop
for(i in 1:dim(Y)[2]){
y <- colnames(Y)[i]
outstring <- paste0(
sprintf("%s <- base_model %%>%%", y),
sprintf(" layer_dense(units = 1, name='%s')",y)
)
eval(parse(text=outstring))
}
But I cannot figure out how to pass a list of the outputs to the compile function. This attempt:
Ylist <- do.call(c, apply(Y, 2, list))
model <- keras_model(input,Ylist)
Returns the following error:
Error in py_call_impl(callable, dots$args, dots$keywords) :
TypeError: unhashable type: 'list'
I also tried keras_array():
model <- keras_model(input,keras_array(Ylist))
Which returned:
Error in py_call_impl(callable, dots$args, dots$keywords) :
TypeError: unhashable type: 'numpy.ndarray'
I am okay with not naming the outputs if there is a way around my for-loop that uses sprintf(). The problem I am working on has over 20 outputs I want to predict simultaneously.
