Skip to content

Latest commit

 

History

History
324 lines (249 loc) · 11.1 KB

README.md

File metadata and controls

324 lines (249 loc) · 11.1 KB
Raw

R interface for MLeap

R-CMD-check MLeap-Tests Coverage status CRAN status

What is MLeap?

MLeap allows us to take Spark pipelines to production. The MLeap runtime can recreate most of Spark’s feature transformers and model predictions. This allows for the ML Pipeline to be deployed with no Spark dependencies.

Figure 1 - Train in Spark

In practice, we can save the ML Pipeline Model (fitted model) as an MLeap bundle (see Figure 1). MLeap serializes the pipeline steps and model. The resulting Zip file can then be used in an external environment that has MLeap. Once the MLeap bundle is loaded in the new environment, new data can be passed to obtain predictions (see Figure 2).

Figure 2 - Deploy with MLeap

The mleap package

The goal of the mleap package is twofold:

  1. Convert an ML Pipeline Model created in sparklyr, into an MLeap bundle file

  2. Load an MLeap bundle file into an R session, and then use the loaded bundle for predictions

Additionally, the mleap package allows us to load an existing MLeap bundle into a Spark session. This would typically be to re-train, or modify a previously created ML Pipeline Model.

The primary functions in mleap are:

  • ml_write_to_bundle_transformed() - Writes an MLeap bundle. It depends on data that has been trained using the pipeline

  • mleap_load_bundle() - Loads an MLeap bundle file into R

  • mleap_transform() - Runs the MLeap bundle steps against new data in R

Additional operational functions in mleap are:

  • ml_read_bundle() - Loads an MLeap bundle file into Spark, via a sparklyr session

  • ml_write_bundle() - Writes an MLeap bundle. It depends on a sample of the training data to re-train the pipeline

Figure 3 - mleap functions

Use Cases

Here are couple of use cases to consider using MLeap, with mleap:

  • It opens the door to collaborate with non-R, and even non-Spark, teams. The resulting MLeap bundle can be used as the integration for those teams to use the model in other environments.

  • Deploy a Shiny app, or a plumber API, with no dependencies on Spark. Using mleap, the model can be loaded into the R environment, and then used for predictions within the R artifact.

Getting started

In order for the R package to work, we will need a local installation of MLeap. Maven is required to install MLeap. mleap contains functions to take care of that.

Steps

  1. Install mleap. For the CRAN version use:

    install.packages("mleap")

    For the development version, use:

    devtools::install_github("rstudio/mleap")

  2. Install Maven. If you already have Maven installed, you can let mleap know by setting an R option:

    options(maven.home = "path/to/maven")`:

    If no installation of Maven exists, use:

    mleap::install_maven()

  3. Install MLeap. There are a couple of considerations regarding the version of MLeap to install:

    • If using Spark, the version of MLeap to install and use will be that closest to the recommended one by the developers of MLeap. The mleap_dep_versions_table() contains the combinations of Spark and MLeap versions as reference.

    • If not using Spark, meaning, that we are using mleap to load an existing bundle, then we would need to match the version of MLeap in which the bundle was originally created.

    mleap::install_mleap(version = "0.20.0")

Example

For the example, we will use the Fine Foods example data. It contains reviews of foods. We will use an ML Pipeline Model to predict if the verbiage in the review can tell us if the customer thinks if the product is “great”.

Create the pipeline

  1. We will use a local version of Spark, version 3.2:

    library(sparklyr)
library(modeldata)

data("small_fine_foods")

sc <- spark_connect(master = "local", version = "3.2")

sff_training_data <- copy_to(sc, training_data)

sff_testing_data <- copy_to(sc, testing_data)

  2. We will create an ML Pipeline. We will index the outcome varaible (score), and then use several text feature transformers to create the features column which will be used as our predictor:

    sff_pipeline <- ml_pipeline(sc) %>% 
  ft_string_indexer(
    input_col = "score",
    output_col = "label",
    handle_invalid = "keep",
    string_order_type = "alphabetDesc"
  ) %>% 
  ft_tokenizer(
    input_col = "review",
    output_col = "word_list"
  ) %>% 
  ft_stop_words_remover(
    input_col = "word_list", 
    output_col = "wo_stop_words"
    ) %>% 
  ft_hashing_tf(
    input_col = "wo_stop_words", 
    output_col = "hashed_features", 
    num_features = 4096,
    binary = TRUE
    ) %>%
  ft_normalizer(
    input_col = "hashed_features", 
    output_col = "features"
    ) %>% 
  ml_logistic_regression(elastic_net_param = 0.05, reg_param = 0.25)

  3. An ML Pipeline Model is now created after running the training data through the pipeline created in the previous step:

    sff_pipeline_model <- ml_fit(sff_pipeline, sff_training_data)

  4. Assuming we are happy with the results. We run the same pipeline using the hold-out set (sff_testing_data). The idea, is that we can use this last transformed data set as a base for our MLeap bundle.

    sff_test_predictions <- sff_pipeline_model %>% 
  ml_transform(sff_testing_data)

  5. Using ml_write_to_bundle_transformed() from mleap, we save the new ML Pipeline Model as an MLeap bundle. We also pass the transformed data set we created with the hold-out test set.

    ml_write_to_bundle_transformed(
  x = sff_pipeline_model,  
  transformed_dataset = sff_test_predictions,  
  path = "sff.zip", 
  overwrite = TRUE
  )
#> Model successfully exported.

  6. We can now close the Spark connection

    spark_disconnect(sc)

Loading an MLeap bundle to R (without Spark dependencies)

  1. We can use the same bundle created in the previous section to load into R. Simply pass the path to the Zip file to ml_load_bundle():

    sff_mleap_model <- mleap_load_bundle("sff.zip")

sff_mleap_model
#> MLeap Transformer
#> <d04e078a-2786-4e5e-923b-cea0ba0ca392> 
#>   Name: pipeline__4ac614f9_18a1_48f0_ac53_d4d03ca86464 
#>   Format: json 
#>   MLeap Version: 0.20.0

  2. We can use mleap_model_schema() to view more information about the contents of the bundle:

    mleap_model_schema(sff_mleap_model)
#> # A tibble: 10 × 5
#>    name            type   nullable dimension io    
#>    <chr>           <chr>  <lgl>    <chr>     <chr> 
#>  1 review          string FALSE    <NA>      input 
#>  2 score           string TRUE     <NA>      input 
#>  3 wo_stop_words   string TRUE     <NA>      output
#>  4 word_list       string TRUE     <NA>      output
#>  5 features        double TRUE     (4096)    output
#>  6 label           double FALSE    <NA>      output
#>  7 hashed_features double TRUE     (4096)    output
#>  8 prediction      double FALSE    <NA>      output
#>  9 rawPrediction   double TRUE     (3)       output
#> 10 probability     double TRUE     (3)       output

  3. mleap_transform() can process the model and new data. Pass a tibble with the expected input variables:

    tibble(review = "worst bad thing I will never buy again", score = "") %>% 
  mleap_transform(sff_mleap_model, .) %>% 
  glimpse()
#> Rows: 1
#> Columns: 10
#> $ review          <chr> "worst bad thing I will never buy again"
#> $ score           <chr> ""
#> $ label           <dbl> 2
#> $ word_list       <list> ["worst", "bad", "thing", "i", "will", "never", "buy",…
#> $ wo_stop_words   <list> ["worst", "bad", "thing", "never", "buy"]
#> $ hashed_features <list> [[[433], [768], [2020], [3081], [4092]], [1, 1, 1, 1, …
#> $ features        <list> [[[433], [768], [2020], [3081], [4092]], [0.4472136, 0…
#> $ rawPrediction   <list> [[6.00653, 5.469893, -11.47642], [3]]
#> $ probability     <list> [[0.6310298, 0.3689702, 1.611768e-08], [3]]
#> $ prediction      <dbl> 0
    tibble(review = "I really loved the proudct best product", score = "") %>% 
  mleap_transform(sff_mleap_model, .) %>% 
  dplyr::glimpse()
#> Rows: 1
#> Columns: 10
#> $ review          <chr> "I really loved the proudct best product"
#> $ score           <chr> ""
#> $ label           <dbl> 2
#> $ word_list       <list> ["i", "really", "loved", "the", "proudct", "best", "pr…
#> $ wo_stop_words   <list> ["really", "loved", "proudct", "best", "product"]
#> $ hashed_features <list> [[[2187], [2365], [3229], [3727], [3984]], [1, 1, 1, 1…
#> $ features        <list> [[[2187], [2365], [3229], [3727], [3984]], [0.4472136,…
#> $ rawPrediction   <list> [[4.768167, 6.708236, -11.47642], [3]]
#> $ probability     <list> [[0.1256402, 0.8743598, 1.107122e-08], [3]]
#> $ prediction      <dbl> 1

Known limitations

MLeap translates the feature transformer and models into its own code base. Not everything available in Spark is translated.

This means two layering things:

  1. No dplyr transformation is available. Only models and feature transformers are available. In sparklyr, feature transformers are functions that start with ft_.

  2. Not every Spark Feature Transformer and model are supported. Please refer to the MLeap documentation to see a concise view of what is available: MLeap Supported Transformers & Models.

Most notably, the following three transformers are not supported:

  • ft_dplyr_transformer()
  • ft_sql_transformer()
  • ft_r_formula()

There is a workaround for ft_r_formula(). It involves using the ML Pipeline “way” of setting up the outcome and predictors. For the predictors, use ft_vector_assembler() if the final stage of the predictors is not a single vectorized variable. For outcomes, anything numeric works fine, but anything categorical will not. Use ft_string_indexer() on top of the outcome variable, before passing it to the modeling step (See the Example section).