0

Declaring a Keras model as reactiveVal causes Shiny to not recognize when it is mutated during a call to fit(). What is the reason for this and what's the best workaround? Is it possible to manually tell Shiny that some reactive variable has been mutated?

Keras models are generally mutated in-place and I suspect that this circumstance causes Shiny to miss whenever the model's parameters are updated during a training run. Or might reason be that Keras is an interface to TensorFlow (i.e. some portion of the code is executed outside the scope of R and Shiny)?

In this self-contained MWE an arbitrary weight is displayed. Although the model is trained the displayed value never updates (the weight does indeed change and is printed to the shell after each epoch of training).

library(keras)
library(shiny)

# Load data, reshape and normalize inputs and one-hot-encode labels
mnist <- dataset_mnist()
x.train <- array_reshape(mnist$train$x, c(nrow(mnist$train$x), 784)) / 255
y.train <- to_categorical(mnist$train$y, 10)

# Returns untrained Keras model. Don't worry about the details.
initialize_model <- function() {
  keras_model_sequential()                                                    %>%
  layer_dense(units = 5, activation = 'relu', input_shape = c(784))           %>%
  layer_dense(units = 10, activation = 'softmax')                             %>%
  compile(loss = 'categorical_crossentropy', optimizer = optimizer_rmsprop())
}

ui <- fluidPage(
  fluidRow(
    column(width = 12, align = "center",
      actionButton('train.model', label = 'Train Model'),
      br(), br(),
      code(textOutput('random.weight'))
    )
  )
)

server <- function(input, output) {

  MODEL         <- reactiveVal(initialize_model())
  RANDOM.WEIGHT <- reactive({ return(MODEL()$get_weights()[[3]][1,2]) })

  observeEvent(input$train.model, {
    # this causes the model object to be mutated in-place but RANDOM.WEIGHT
    # is never updated
    MODEL() %>% fit(x.train, y.train, epochs = 1, batch_size = 60000)
    cat("The weight's new value is = ", MODEL()$get_weights()[[3]][1,2], "\n")
  })
  output$random.weight <- renderText({ 
    return(paste0("RANDOM.WEIGHT() = ", RANDOM.WEIGHT()))
  })
}

app <- shinyApp(ui = ui, server = server)
runApp(app, port = 1337)

Any ideas? Thanks

Response to comments:

What's the output of dput(initialize_model())?

structure(function (object)
{
    compose_layer(object, x)
}, class = c("keras.engine.sequential.Sequential",        
             "keras.engine.training.Model",
             "keras.engine.network.Network",
             "keras.engine.base_layer.Layer",              
             "tensorflow.python.training.checkpointable.base.CheckpointableBase",
             "python.builtin.object"), py_object = <environment>)

Unfortunately, that doesn't really tell me much.

korolev
  • 46
  • 5
  • 1
    I don't have `keras`, so this is speculation: does `keras_model_sequential()` return a "normal" structure, or does it include an `external pointer` in some form? – r2evans Jun 10 '19 at 12:15
  • @r2evans I don't know. I have just begun my venture into R and neural network hence I don't know anything about the internals of R. If you give me a hint where to look or what to do I'll be glad to come back with an answer to your question. – korolev Jun 10 '19 at 14:32
  • 1
    Shiny can only recognize reactive elements that depend on other reactive elements or things that come from `input$`. Most R objects update with new values but `Keras` seems to be updating via reference. The object is updating "in place" and Shiny can't see the difference. The `data.table` package also updates by reference so you need workaround [like this](https://stackoverflow.com/questions/44074599/r-shiny-use-reactivevalues-with-data-table-assign-by-reference) or [this](https://stackoverflow.com/questions/32536940/shiny-reactivity-fails-with-data-tables) – MrFlick Jun 10 '19 at 14:53
  • Slavistan, do `dput(initialize_model())`, and be prepared for a possible console-filling storm. For instance, if you `dput(data.table::data.table())`, you'll see ``, which is limited in a couple of ways: (1) it does not survive forking; (2) cannot be passed (as-is) to other R sessions; and (3) changes in the internal `struct` of the object are not noticed by `shiny` (or anything in R, for that matter). (This `pointer:` is literally a memory pointer to a C/C++-struct, something that cannot be (easily) stored in a native-R object.) – r2evans Jun 10 '19 at 15:22
  • @MrFlick I updated my question with the output to `dput(initialize_model())`. Does that help in any way? – korolev Jun 12 '19 at 10:05

0 Answers0