Mapping a keras NN over data lists in R - r

Question

I am trying to figure out the correct method of going about applying a keras model to each of my lists. I have used the iris dataset and created 4 lists and the goal is to correctly predict versicolor or virginica (I omit setosa because I want a binary classification model).
data(iris)
iris <- iris %>%
mutate(
splt = sample(4, size = nrow(.), replace = TRUE),
binary = case_when(
Species == "versicolor" ~ 0,
Species == "virginica" ~ 1
)
) %>%
filter(Species != "setosa") %>%
split(., .$splt)
iris_x_train <- iris %>%
map(., ~select(., Sepal.Length, Sepal.Width, Petal.Length, Petal.Width) %>%
as.matrix())
iris_y_train <- iris %>%
map(., ~select(., binary) %>%
to_categorical(2))
NN_model <- keras_model_sequential() %>%
layer_dense(units = 4, activation = 'relu', input_shape = 4) %>%
layer_dense(units = 2, activation = 'softmax')
NN_model %>%
summary
NN_model %>%
compile(
loss = 'binary_crossentropy',
optimizer_sgd(lr = 0.01, momentum = 0.9),
metrics = c('accuracy')
)
My problem occurs here. When I apply the below code:
NN_model %>%
future_map(., ~future_map2(
.x = iris_x_train,
.y = iris_y_train,
~fit(
x = .x,
y = .y,
epochs = 5,
batch_size = 20,
validation_split = 0
)
)
)
I get this error:
Error in py_get_item_impl(x, key, FALSE) : TypeError: 'Sequential'
object does not support indexing
When I apply this code:
NN_model %>%
future_map2(
.x = iris_x_train,
.y = iris_y_train,
~fit(
x = .x,
y = .y,
epochs = 5,
batch_size = 20,
validation_split = 0
)
)
I get this error:
~fit(x = .x, y = .y, epochs = 5, batch_size = 20, validation_split =
0) Error in py_call_impl(callable,
dots$args, dots$keywords) : Evaluation error: Unable to convert R
object to Python type.
How can I map a keras model to each of the 4 datasets?
library(keras)
library(tensorflow)
library(furrr)
library(purrr)
The following works for the first list:
NN_model %>%
fit(
x = iris_x_train[[1]],
y = iris_y_train[[1]],
epochs = 50,
batch_size = 20,
validation_split = 0
)
EDIT: I seem to have solved it.
Putting the NN_model inside the fit() function appears to work.
future_map2(
.x = iris_x_train,
.y = iris_y_train,
~fit(NN_model,
.x,
.y,
epochs = 5,
batch_size = 20,
validation_split = 0
)
)