Overview

The kerasformula package deal gives a high-level interface for the R interface to Keras. It’s principal interface is the kms operate, a regression-style interface to keras_model_sequential that makes use of formulation and sparse matrices.

The kerasformula package deal is obtainable on CRAN, and might be put in with:

# set up the kerasformula package deal
set up.packages("kerasformula")    
# or devtools::install_github("rdrr1990/kerasformula")

library(kerasformula)

# set up the core keras library (if you have not already completed so)
# see ?install_keras() for choices e.g. install_keras(tensorflow = "gpu")
install_keras()

The kms() operate

Many traditional machine studying tutorials assume that information are available in a comparatively homogenous kind (e.g., pixels for digit recognition or phrase counts or ranks) which may make coding considerably cumbersome when information is contained in a heterogenous information body. kms() takes benefit of the flexibleness of R formulation to clean this course of.

kms builds dense neural nets and, after becoming them, returns a single object with predictions, measures of match, and particulars concerning the operate name. kms accepts various parameters together with the loss and activation features present in keras. kms additionally accepts compiled keras_model_sequential objects permitting for even additional customization. This little demo reveals how kms can help is mannequin constructing and hyperparameter choice (e.g., batch measurement) beginning with uncooked information gathered utilizing library(rtweet).

Let’s have a look at #rstats tweets (excluding retweets) for a six-day interval ending January 24, 2018 at 10:40. This occurs to offer us a pleasant cheap variety of observations to work with when it comes to runtime (and the aim of this doc is to indicate syntax, not construct notably predictive fashions).

rstats <- search_tweets("#rstats", n = 10000, include_rts = FALSE)
dim(rstats)

  [1] 2840   42

Suppose our aim is to foretell how standard tweets shall be based mostly on how typically the tweet was retweeted and favorited (which correlate strongly).

cor(rstats$favorite_count, rstats$retweet_count, technique="spearman")

    [1] 0.7051952

Since few tweeets go viral, the info are fairly skewed in direction of zero.

Getting essentially the most out of formulation

Let’s suppose we’re fascinated by placing tweets into classes based mostly on reputation however we’re unsure how finely-grained we wish to make distinctions. Among the information, like rstats$mentions_screen_name is available in an inventory of various lengths, so let’s write a helper operate to depend non-NA entries.

Let’s begin with a dense neural internet, the default of kms. We will use base R features to assist clear the info–on this case, reduce to discretize the result, grepl to search for key phrases, and weekdays and format to seize completely different facets of the time the tweet was posted.

breaks <- c(-1, 0, 1, 10, 100, 1000, 10000)
reputation <- kms(reduce(retweet_count + favorite_count, breaks) ~ screen_name + 
                  supply + n(hashtags) + n(mentions_screen_name) + 
                  n(urls_url) + nchar(textual content) +
                  grepl('photograph', media_type) +
                  weekdays(created_at) + 
                  format(created_at, '%H'), rstats)
plot(reputation$historical past) 
  + ggtitle(paste("#rstat reputation:", 
            paste0(spherical(100*reputation$evaluations$acc, 1), "%"),
            "out-of-sample accuracy")) 
  + theme_minimal()

reputation$confusion

reputation$confusion

                    (-1,0] (0,1] (1,10] (10,100] (100,1e+03] (1e+03,1e+04]
      (-1,0]            37    12     28        2           0             0
      (0,1]             14    19     72        1           0             0
      (1,10]             6    11    187       30           0             0
      (10,100]           1     3     54       68           0             0
      (100,1e+03]        0     0      4       10           0             0
      (1e+03,1e+04]      0     0      0        1           0             0

The mannequin solely classifies about 55% of the out-of-sample information appropriately and that predictive accuracy doesn’t enhance after the primary ten epochs. The confusion matrix means that mannequin does greatest with tweets which might be retweeted a handful of occasions however overpredicts the 1-10 stage. The historical past plot additionally means that out-of-sample accuracy shouldn’t be very steady. We will simply change the breakpoints and variety of epochs.

breaks <- c(-1, 0, 1, 25, 50, 75, 100, 500, 1000, 10000)
reputation <- kms(reduce(retweet_count + favorite_count, breaks) ~  
                  n(hashtags) + n(mentions_screen_name) + n(urls_url) +
                  nchar(textual content) +
                  screen_name + supply +
                  grepl('photograph', media_type) +
                  weekdays(created_at) + 
                  format(created_at, '%H'), rstats, Nepochs = 10)

plot(reputation$historical past) 
  + ggtitle(paste("#rstat reputation (new breakpoints):",
            paste0(spherical(100*reputation$evaluations$acc, 1), "%"),
            "out-of-sample accuracy")) 
  + theme_minimal()

That helped some (about 5% extra predictive accuracy). Suppose we wish to add somewhat extra information. Let’s first retailer the enter system.

pop_input <- "reduce(retweet_count + favorite_count, breaks) ~  
                          n(hashtags) + n(mentions_screen_name) + n(urls_url) +
                          nchar(textual content) +
                          screen_name + supply +
                          grepl('photograph', media_type) +
                          weekdays(created_at) + 
                          format(created_at, '%H')"

Right here we use paste0 so as to add to the system by looping over consumer IDs including one thing like:

grepl("12233344455556", mentions_user_id)

mentions <- unlist(rstats$mentions_user_id)
mentions <- distinctive(mentions[which(table(mentions) > 5)]) # take away rare
mentions <- mentions[!is.na(mentions)] # drop NA

for(i in mentions)
  pop_input <- paste0(pop_input, " + ", "grepl(", i, ", mentions_user_id)")

reputation <- kms(pop_input, rstats)

That helped a contact however the predictive accuracy remains to be pretty unstable throughout epochs…

Customizing layers with kms()

We might add extra information, maybe add particular person phrases from the textual content or another abstract stat (imply(textual content %in% LETTERS) to see if all caps explains reputation). However let’s alter the neural internet.

The enter.system is used to create a sparse mannequin matrix. For instance, rstats$supply (Twitter or Twitter-client utility sort) and rstats$screen_name are character vectors that shall be dummied out. What number of columns does it have?

    [1] 1277

Say we wished to reshape the layers to transition extra regularly from the enter form to the output.

reputation <- kms(pop_input, rstats,
                  layers = record(
                    models = c(1024, 512, 256, 128, NA),
                    activation = c("relu", "relu", "relu", "relu", "softmax"), 
                    dropout = c(0.5, 0.45, 0.4, 0.35, NA)
                  ))

kms builds a keras_sequential_model(), which is a stack of linear layers. The enter form is set by the dimensionality of the mannequin matrix (reputation$P) however after that customers are free to find out the variety of layers and so forth. The kms argument layers expects an inventory, the primary entry of which is a vector models with which to name keras::layer_dense(). The primary ingredient the variety of models within the first layer, the second ingredient for the second layer, and so forth (NA as the ultimate ingredient connotes to auto-detect the ultimate variety of models based mostly on the noticed variety of outcomes). activation can be handed to layer_dense() and will take values comparable to softmax, relu, elu, and linear. (kms additionally has a separate parameter to manage the optimizer; by default kms(... optimizer="rms_prop").) The dropout that follows every dense layer price prevents overfitting (however in fact isn’t relevant to the ultimate layer).

Selecting a Batch Dimension

By default, kms makes use of batches of 32. Suppose we have been proud of our mannequin however didn’t have any explicit instinct about what the scale must be.

Nbatch <- c(16, 32, 64)
Nruns <- 4
accuracy <- matrix(nrow = Nruns, ncol = size(Nbatch))
colnames(accuracy) <- paste0("Nbatch_", Nbatch)

est <- record()
for(i in 1:Nruns){
  for(j in 1:size(Nbatch)){
   est[[i]] <- kms(pop_input, rstats, Nepochs = 2, batch_size = Nbatch[j])
   accuracy[i,j] <- est[[i]][["evaluations"]][["acc"]]
  }
}
  
colMeans(accuracy)

    Nbatch_16 Nbatch_32 Nbatch_64 
    0.5088407 0.3820850 0.5556952 

For the sake of curbing runtime, the variety of epochs was set arbitrarily brief however, from these outcomes, 64 is the most effective batch measurement.

Making predictions for brand new information

To this point, we’ve been utilizing the default settings for kms which first splits information into 80% coaching and 20% testing. Of the 80% coaching, a sure portion is put aside for validation and that’s what produces the epoch-by-epoch graphs of loss and accuracy. The 20% is barely used on the finish to evaluate predictive accuracy.
However suppose you wished to make predictions on a brand new information set…

reputation <- kms(pop_input, rstats[1:1000,])
predictions <- predict(reputation, rstats[1001:2000,])
predictions$accuracy

    [1] 0.579

As a result of the system creates a dummy variable for every display identify and point out, any given set of tweets is all however assured to have completely different columns. predict.kms_fit is an S3 technique that takes the brand new information and constructs a (sparse) mannequin matrix that preserves the unique construction of the coaching matrix. predict then returns the predictions together with a confusion matrix and accuracy rating.

In case your newdata has the identical noticed ranges of y and columns of x_train (the mannequin matrix), you may as well use keras::predict_classes on object$mannequin.

Utilizing a compiled Keras mannequin

This part reveals methods to enter a mannequin compiled within the vogue typical to library(keras), which is beneficial for extra superior fashions. Right here is an instance for lstm analogous to the imbd with Keras instance.

okay <- keras_model_sequential()
okay %>%
  layer_embedding(input_dim = reputation$P, output_dim = reputation$P) %>% 
  layer_lstm(models = 512, dropout = 0.4, recurrent_dropout = 0.2) %>% 
  layer_dense(models = 256, activation = "relu") %>%
  layer_dropout(0.3) %>%
  layer_dense(models = 8, # variety of ranges noticed on y (end result)  
              activation = 'sigmoid')

okay %>% compile(
  loss = 'categorical_crossentropy',
  optimizer = 'rmsprop',
  metrics = c('accuracy')
)

popularity_lstm <- kms(pop_input, rstats, okay)

Drop me a line by way of the challenge’s Github repo. Particular due to @dfalbel and @jjallaire for useful strategies!!