7 Improving an operator

Now that we have deployed our R operator for Tercen, we can always improve it! In this chapter we will see how to catch errors, add input parameters (properties), and prepare testing for our linear regression operator.

Error catching

1. Check the presence of input projection

For the lm_operator, we can check if both x and y axes have been set in the projection, and return an error message to help the user. We can do that using the try() function.

ctx <- tercenCtx()

if(inherits(try(ctx$select(".x")), 'try-error')) stop("x axis is missing.")
if(inherits(try(ctx$select(".y")), 'try-error')) stop("y axis is missing.")

2. Catch errors in key processes

We can also use the try() function to catch potential errors in key processes of our operator. In our example, we can for example catch error occuring while running the lm() function. Then, we return coefficients only if the model ran successfully. Otherwise, NaN are returned.

do.lm <- function(df) {
  out <- data.frame(
    .ri = df$.ri[1],
    .ci = df$.ci[1],
    intercept = NaN,
    slope = NaN
  )
  mod <- try(lm(.y ~ .x, data = df))    # try-catch
  if(!inherits(mod, 'try-error')) {     # get coefficients if lm ran succesfully
    out$intercept <- mod$coefficients[1]
    out$slope <- mod$coefficients[2]
  }
  return(out)
}

Adding properties

Tercen operators can take input parameters (called properties). They can be of different types (boolean, enumerated, numeric).

As an example, let’s see how to add a property to the lm_operator. We will add a parameter to decide whether to omit the intercept in the model or not.

1. Modify the operator’s JSON file to add properties.

For each property, we have to set values to different attributes:

• kind: property kind (BooleanProperty, DoubleProperty, or EnumeratedProperty)

• name: name that will be displayed in Tercen

• defaultValue: default value taken by the property

• description: description to be displayed in Tercen

In our case, the JSON file looks now like this:

{
  "name": "Linear regression",
  "description": "Returns the intercept and slope of a linear regression in a cell",
  "tags": ["linear model"],
  "authors": ["tercen"],
  "urls": ["https://github.com/agouy/lm_operator"],
  "properties": [
  {
    "kind": "BooleanProperty",
    "name": "intercept.omit",
    "defaultValue": false,
    "description": "A logical value indicating whether the intercept should be omitted in the model."
  }
  ]
}

2. Use the property in the R code

This property can be called in the R operator code as follows:

  intercept.omit <- as.logical(ctx$op.value('intercept.omit'))
  if(intercept.omit) {
    mod <- try(lm(.y ~ .x - 1, data = df))
  } else {
    mod <- try(lm(.y ~ .x, data = df))
  }

Preparing operator testing

It’s always good to prepare some tests that could be ran when a new version of Tercen is released.

To include a test, you need to create a test subdirectory in your project directory. It must include:

• a test input file

• an expected output file

• a JSON file containing information about the test

For example, a test for our lm operator can be created as follow:

# Simulate tercen input based on the CO2 dataset 
# with an x and y-axis, rows and columns
data(CO2)
df <- data.frame(.x = CO2$conc, .y = CO2$uptake, .ri = CO2$Plant, .ci = CO2$Treatment)

# Run the do.lm() function created above to generate the expected output
out <- df %>% select(.ci, .ri, .x, .y) %>%
  group_by(.ri, .ci) %>%
  do(do.lm(.))

# write input and expected output in the test subdirectory
write.csv(CO2, file="./test/CO2.csv", row.names = FALSE, quote = FALSE)
write.csv(out, file="./test/output.csv", row.names = FALSE, quote = FALSE)

Now that we have our input and ouput files, we can create the JSON file that shall include the following information:

{
  "kind": "OperatorUnitTest",
  "name": "testlm1",
  "namespace": "test",
  "inputDataUri": "CO2.csv",
  "outputDataUri": ["output.csv"],
  "columns": ["Treatment"],
  "rows": ["Plant"],
  "colors": [],
  "labels": [],
  "yAxis": "uptake",
  "xAxis": "conc"
}

The name attribute is free. Input and output test files names must be assigned to the inputDataUri and inputDataUri, respectively. Variable names of the input file must be assigned to the columns, rows, colors, labels, yAxis, xAxis attributes. In our case, colors and labels are left empty as they are not part of our input.