Neptune-R Integration

You can interact freely with Neptune from R and enjoy the same functionality that is available in Python.

Installation

The Neptune integration with R is available as a CRAN package and a Github project. You can install it by running:

install.packages('neptune')

Python environment setup

Under the hood, Neptune uses reticulate, which in turn calls Python. That means you must:

  • Install Python in your environment (venv, conda, miniconda and pure Python environments are supported).

  • Point to that environment when you initialize Neptune, as shown below.

Initialize Neptune

Toward the top of your script, you must import the neptune library and initialize the connection with Neptune.

Specify:

  • project_name: the name of your project in Neptune

  • api_token: a key associated with your Neptune account

library('neptune')
init_neptune(project_name = 'shared/r-integration',
             api_token = 'ANONYMOUS')

Once this is added, you can start logging your experiment data to Neptune.

Note

As an example, we are using the ‘ANONYMOUS’ user token and a public project named ‘shared/r-integration’.

Warning

The suggested way to pass in your api_token is to store your key in an environment variable and pass it using Sys.getenv('MY_NEPTUNE_KEY').

Sys.setenv('NEPTUNE_API_TOKEN'='eyJhcGlfYWRkcmVzcyI6Imh0dHBz')

init_neptune(project_name = 'my_organization/my_project',
             api_token = Sys.getenv('NEPTUNE_API_TOKEN')
             )

Specifying Python environment

  • Python

    init_neptune(project_name = 'shared/r-integration',
             api_token = 'ANONYMOUS'
             python='python',
             python_path='/usr/bin/python3')

  • venv

    init_neptune(project_name = 'shared/r-integration',
             api_token = 'ANONYMOUS'
             python='venv',
             python_path='my_venv')

  • conda

    init_neptune(project_name = 'shared/r-integration',
             api_token = 'ANONYMOUS'
             python='conda',
             python_path='my_conda_env')

  • miniconda

    init_neptune(project_name = 'shared/r-integration',
             api_token = 'ANONYMOUS'
             python='miniconda',
             python_path='my_miniconda_env')

Create experiment

To start tracking, you must create an experiment.

You can:

  • Name your experiments

  • Tag them to keep your work organized

  • Specify params to keep track of hyperparameters of your experiments

For example:

create_experiment(name='training on Sonar',
                  tags=c('random-forest','sonar'),
                  params = list(ntree=100,
                                mtry=10,
                                maxnodes=20)
                  )

Track data versions and other properties

Keeping track of your data is an important part of the job. With Neptune, you can log a fingerprint (hash) of your data for every experiment. By doing so, you will make sure that you are “comparing apples to apples.”

Add a property to your experiment:

library(digest)

# Load data
data(Sonar)
dataset <- Sonar

set_property(property = 'data-version', value = digest(dataset))

Track metrics

Tracking evaluation metrics is as simple as logging. You can track a single metric by using the log_metric() method. Just define your logging channel name and metric value.

For example:

log_metric('mean OOB error', mean(model$err.rate[,1]))
log_metric('error class M', model$confusion[1,3])
log_metric('error class R', model$confusion[2,3])

If you want to log multiple values to a single logging channel, just call the log_metric() method multiple times for the same channel name. Neptune will automatically create charts for you!

for (err in (model$err.rate[,1])) {
  log_metric('OOB errors', err)
}

You can sort your experiments by the metrics you care about and take a look at the charts in the application.

Track artifacts

You can also save your model weights, PDF report files or other objects in Neptune. All you need to do is pass the filepath to the neptune$send_artifact() method and it will be attached to your experiment.

save(model, file="model.Rdata")
log_artifact('model.Rdata')

Once it is logged, sharing it with your colleagues or collaborators is super easy.

Track images and charts

Logging images and charts to Neptune is very simple, as well. Just use the neptune$send_image() method that takes the name of the logging channel and a path to image as arguments. You can log more than one chart to the same channel to organize things - just send another image to the same channel.

for (t in c(1,2)){
  jpeg('importance_plot.jpeg')
  varImpPlot(model,type=t)
  dev.off()
  log_image('feature_importance', 'importance_plot.jpeg')
}

Once it is logged, you can view it in the app.

Stop experiment

Once you are finished tracking, you should stop the experiment

stop_experiment()

Explore your experiment in Neptune

Thanks to all the logging you’ve done, you can now see all your experiments in the Neptune app.

Explore the shared/r-integration public project to see how it looks.

Experiment Tracking in R

Full tracking script

Install dependencies

# install neptune
install.packages('neptune', dependencies = TRUE)

# install other packages for this tutorial
install.packages(c('digest', 'mlbench', 'randomForest'), dependencies = TRUE)

Run experiment with tracking script

# load libraries,
library(neptune)
library(digest)
library(mlbench)
library(randomForest)

SEED=1234
set.seed(SEED)

# load dataset
data(Sonar)
dataset <- Sonar
x <- dataset[,1:60]   # predictors
y <- dataset[,61]     # labels

# Initialize Neptune
init_neptune(project_name = 'shared/r-integration',
             api_token = 'ANONYMOUS',
             python='miniconda',
             python_path='py_37'
             )

# Start an experiment and track hyperparameters

params = list(ntree=625,
              mtry=13,
              maxnodes=50
              )

create_experiment(name='training on Sonar',
                  tags=c('random-forest','sonar'),
                  params = params
)

# track data version and SEED
set_property(property = 'data-version', value = digest(dataset))
set_property(property = 'seed', value = SEED)

# train your model
model <- randomForest(x = x, y = y,
  ntree=params$ntree, mtry = params$mtry, maxnodes = params$maxnodes,
  importance = TRUE
  )

# Log metrics

log_metric('mean OOB error', mean(model$err.rate[,1]))
log_metric('error class M', model$confusion[1,3])
log_metric('error class R', model$confusion[2,3])

for (err in (model$err.rate[,1])) {
  log_metric('OOB errors', err)
}

# Log artifact
save(model, file="model.Rdata")
log_artifact('model.Rdata')

# Log image
for (t in c(1,2)){
  jpeg('importance_plot.jpeg')
  varImpPlot(model,type=t)
  dev.off()
  log_image('feature_importance', 'importance_plot.jpeg')
}

# Stop Experiment
stop_experiment()