CSV Type Provider

This article demonstrates how to use the CSV type provider to read CSV files in a statically typed way.

The CSV type provider takes a sample CSV as input and generates a type based on the data present on the columns of that sample. The column names are obtained from the first (header) row, and the types are inferred from the values present on the subsequent rows.

Introducing the provider

The type provider is located in the FSharp.Data.dll assembly. Assuming the package is referenced we can access its namespace as follows:

open FSharp.Data

Parsing stock prices

The Yahoo Finance web site provides daily stock prices in a CSV format that has the following structure (you can find a larger example in the data/MSFT.csv file):

Date,Open,High,Low,Close,Volume,Adj Close
2012-01-27,29.45,29.53,29.17,29.23,44187700,29.23
2012-01-26,29.61,29.70,29.40,29.50,49102800,29.50
2012-01-25,29.07,29.65,29.07,29.56,59231700,29.56
2012-01-24,29.47,29.57,29.18,29.34,51703300,29.34

As usual with CSV files, the first row contains the headers (names of individual columns) and the next rows define the data. We can pass reference to the file to CsvProvider to get a strongly typed view of the file:

[<Literal>]
let ResolutionFolder = __SOURCE_DIRECTORY__

type Stocks = CsvProvider<"../data/MSFT.csv", ResolutionFolder=ResolutionFolder>

The generated type provides two static methods for loading data. The Parse method can be used if we have the data in a string value. The Load method allows reading the data from a file or from a web resource (and there's also an asynchronous AsyncLoad version). We could also have used a web URL instead of a local file in the sample parameter of the type provider. The following sample calls the Load method with an URL that points to a live CSV file on the Yahoo finance web site:

// Download the stock prices
let msft =
    Stocks
        .Load(__SOURCE_DIRECTORY__ + "/../data/MSFT.csv")
        .Cache()

// Look at the most recent row. Note the 'Date' property
// is of type 'DateTime' and 'Open' has a type 'decimal'
let firstRow = msft.Rows |> Seq.head
let lastDate = firstRow.Date
let lastOpen = firstRow.Open

// Print the first 10 prices in the HLOC format
for row in msft.Rows |> Seq.truncate 10 do
    printfn "HLOC: (%A, %A, %A, %A)" row.High row.Low row.Open row.Close

HLOC: (76.55M, 75.86M, 75.97M, 76.29M)
HLOC: (76.03M, 75.54M, 75.67M, 76.00M)
HLOC: (76.12M, 74.96M, 75.22M, 75.97M)
HLOC: (74.72M, 73.71M, 74.09M, 74.69M)
HLOC: (74.88M, 74.20M, 74.67M, 74.26M)
HLOC: (75.01M, 74.30M, 74.71M, 74.61M)
HLOC: (74.54M, 73.88M, 73.94M, 74.49M)
HLOC: (73.97M, 73.31M, 73.54M, 73.87M)
HLOC: (74.17M, 73.17M, 73.55M, 73.85M)
HLOC: (73.81M, 72.99M, 73.67M, 73.26M)
val msft: Runtime.CsvFile<CsvProvider<...>.Row>
val firstRow: CsvProvider<...>.Row =
  (10/9/2017 12:00:00 AM, 75.97M, 76.55M, 75.86M, 76.29M, 11386502)
val lastDate: System.DateTime = 10/9/2017 12:00:00 AM
val lastOpen: decimal = 75.97M
val it: unit = ()

The generated type has a property Rows that returns the data from the CSV file as a collection of rows. We iterate over the rows using a for loop. As you can see the (generated) type for rows has properties such as High, Low and Close that correspond to the columns in the CSV file.

As you can see, the type provider also infers types of individual rows. The Date property is inferred to be a DateTime (because the values in the sample file can all be parsed as dates) while HLOC prices are inferred as decimal.

Using units of measure

The CSV type provider supports F# units of measure: if the header includes the name or symbol of one of the standard SI units, then the generated type returns values annotated with the appropriate unit.

In this section, we use a simple file data/SmallTest.csv which looks as follows:

Name,  Distance (metre), Time (s)
First, 50.0,             3.7

As you can see, the second and third columns are annotated with metre and s, respectively. To use units of measure in our code, we need to open the namespace with standard unit names. Then, we pass the SmallTest.csv file to the type provider as a static argument. Also, note that in this case, we're using the same data at runtime, so we use the GetSample method instead of calling Load and passing the same parameter again.

let small =
    CsvProvider<"../data/SmallTest.csv", ResolutionFolder=ResolutionFolder>.GetSample ()

val small: CsvProvider<...>

We can also use the default constructor instead of the GetSample static method:

let small2 =
    new CsvProvider<"../data/SmallTest.csv", ResolutionFolder=ResolutionFolder>()

val small2: CsvProvider<...>

but the VisualStudio IntelliSense for the type provider parameters doesn't work when we use a default constructor for a type provider, so we'll keep using GetSample instead.

As in the previous example, the small value exposes the rows using the Rows property. The generated properties Distance and Time are now annotated with units. Look at the following simple calculation:

open FSharp.Data.UnitSystems.SI.UnitNames

for row in small.Rows do
    let speed = row.Distance / row.Time

    if speed > 15.0M<metre/second> then
        printfn "%s (%A m/s)" row.Name speed

Second (19.230769230769230769230769231M m/s)
Third (23.4375M m/s)
val it: unit = ()

The numerical values of Distance and Time are both inferred as decimal (because they are small enough). Thus the type of speed becomes decimal<metre/second>. The compiler can then statically check that we're not comparing incompatible values - e.g. number in meters per second against a value in kilometres per hour.

Custom separators and tab-separated files

By default, the CSV type provider uses a comma (,) as a separator. However, CSV files sometimes use a different separator character than ,. In some European countries, , is already used as the numeric decimal separator, so a semicolon (;) is used instead to separate CSV columns. The CsvProvider has an optional Separators static parameter where you can specify what to use as a separator. This means that you can consume any textual tabular format. Here is an example using ; as a separator:

type AirQuality = CsvProvider<"../data/AirQuality.csv", ";", ResolutionFolder=ResolutionFolder>

let airQuality = new AirQuality()

for row in airQuality.Rows |> Seq.truncate 10 do
    if row.Month > 6 then
        printfn "Temp: %i Ozone: %f " row.Temp row.Ozone

type AirQuality = CsvProvider<...>
val airQuality: AirQuality
val it: unit = ()

The air quality dataset (data/AirQuality.csv) is used in many samples for the Statistical Computing language R. A short description of the dataset can be found in the R language manual.

If you are parsing a tab-separated file that uses \t as the separator, you can also specify the separator explicitly. However, if you're using an url or file that has the .tsv extension, the type provider will use \t by default. In the following example, we also set IgnoreErrors static parameter to true so that lines with an incorrect number of elements are automatically skipped (the sample file (data/MortalityNY.csv) contains additional unstructured data at the end):

let mortalityNy =
    CsvProvider<"../data/MortalityNY.tsv", IgnoreErrors=true, ResolutionFolder=ResolutionFolder>.GetSample ()

// Find the name of a cause based on code
// (Pedal cyclist injured in an accident)
let cause =
    mortalityNy.Rows
    |> Seq.find (fun r -> r.``Cause of death Code`` = "V13.4")

// Print the number of injured cyclists
printfn "CAUSE: %s" cause.``Cause of death``

for r in mortalityNy.Rows do
    if r.``Cause of death Code`` = "V13.4" then
        printfn "%s (%d cases)" r.County r.Count

CAUSE: Pedal cyclist injured in collision with car, pick-up truck or van, driver injured in traffic accident
Albany County, NY (1 cases)
Bronx County, NY (1 cases)
Broome County, NY (1 cases)
Cayuga County, NY (1 cases)
Chemung County, NY (1 cases)
Dutchess County, NY (1 cases)
Kings County, NY (3 cases)
Monroe County, NY (1 cases)
Nassau County, NY (8 cases)
New York County, NY (1 cases)
Niagara County, NY (2 cases)
Oneida County, NY (2 cases)
Onondaga County, NY (2 cases)
Orange County, NY (2 cases)
Oswego County, NY (2 cases)
Queens County, NY (1 cases)
Rensselaer County, NY (2 cases)
Saratoga County, NY (2 cases)
Schenectady County, NY (1 cases)
Seneca County, NY (1 cases)
Steuben County, NY (1 cases)
Suffolk County, NY (9 cases)
Sullivan County, NY (1 cases)
Ulster County, NY (1 cases)
Westchester County, NY (3 cases)
val mortalityNy: CsvProvider<...>
val cause: CsvProvider<...>.Row =
  ("", "Albany County, NY", 36001,
   "Pedal cyclist injured in collision with car, pick-up truck or"+[40 chars],
   "V13.4", 1, 2072701, "0.0 (Unreliable)")
val it: unit = ()

Finally, note that it is also possible to specify multiple different separators for the CsvProvider. This might be useful if a file is irregular and contains rows separated by either a semicolon or a colon. You can use: CsvProvider<"../data/AirQuality.csv", Separators=";,", ResolutionFolder=ResolutionFolder>.

Missing values

It is quite common in statistical datasets for some values to be missing. If you open the data/AirQuality.csv file, you will see that some values for the ozone observations are marked #N/A. Such values are parsed as float and will be marked with Double.NaN in F#. The values NaN, NA, N/A, #N/A, :, -, TBA, and TBD are recognized as missing values by default, but you can customize it by specifying the MissingValues static parameter of CsvProvider as a comma-separated string. For example, to ignore this and that we could do:

CsvProvider<"X,Y,Z\nthis,that,1.0", MissingValues="this,that">
    .GetSample()
    .Rows

val it: CsvProvider<...>.Row seq = seq [(nan, nan, 1.0M {Scale = 1uy;})]

The following snippet calculates the mean of the ozone observations excluding the Double.NaN values. We first obtain the Ozone property for each row, then remove missing values and then use the standard Seq.average function:

open System

let mean =
    airQuality.Rows
    |> Seq.toArray
    |> Array.map (fun row -> row.Ozone)
    |> Array.filter (fun elem -> not (Double.IsNaN elem))
    |> Array.average

val mean: float = 42.12931034

If the sample doesn't have missing values on all columns, but at runtime missing values could appear anywhere, you can set the static parameter AssumeMissingValues to true in order to force CsvProvider to assume missing values can occur in any column.

Controlling the column types

By default, the CSV type provider checks the first 1000 rows to infer the types, but you can customize it by specifying the InferRows static parameter of CsvProvider. If you specify 0 the entire file will be used.

Columns with only 0, 1, Yes, No, True, or False will be set to bool. Columns with numerical values will be set to either int, int64, decimal, or float, in that order of preference.

If a value is missing in any row, by default the CSV type provider will infer a nullable (for int and int64) or an optional (for bool, DateTime and Guid). When a decimal would be inferred but there are missing values, we will infer a float instead, and use Double.NaN to represent those missing values. The string type is already inherently nullable, so by default, we won't generate a string option. If you prefer to use optionals in all cases, you can set the static parameter PreferOptionals to true. In that case, you'll never get an empty string or a Double.NaN and will always get a None instead.

If you have other preferences, e.g. if you want a column to be a float instead of a decimal, you can override the default behaviour by specifying the types in the header column between braces, similar to what can be done to specify the units of measure. This will override both AssumeMissingValues and PreferOptionals. The valid types are:

• int
• int?
• int option
• int64
• int64?
• int64 option
• bool
• bool?
• bool option
• float
• float?
• float option
• decimal
• decimal?
• decimal option
• date
• date?
• date option
• datetimeoffset
• datetimeoffset?
• datetimeoffset option
• guid
• guid?
• guid option
• string
• string option.

You can also specify both the type and a unit (e.g float<metre>). Example:

Name,  Distance (decimal?<metre>), Time (float)
First, 50,                        3

Additionally, you can also specify some or all the types in the Schema static parameter of CsvProvider. Valid formats are:

• Type
• Type<Measure>
• Name (Type)
• Name (Type<Measure>)

What's specified in the Schema static parameter will always take precedence to what's specified in the column headers.

If the first row of the file is not a header row, you can specify the HasHeaders static parameter to false in order to consider that row as a data row. In that case, the columns will be named Column1, Column2, etc..., unless the names are overridden using the Schema parameter. Note that you can override only the name in the Schema parameter and still have the provider infer the type for you. Example:

type OneTwoThree = CsvProvider<"1,2,3", HasHeaders=false, Schema="Duration (float<second>),foo,float option">

let csv = OneTwoThree.GetSample()

for row in csv.Rows do
    printfn "%f %d %f" (row.Duration / 1.0<second>) row.Foo (defaultArg row.Column3 1.0)

1.000000 2 3.000000
type OneTwoThree = CsvProvider<...>
val csv: CsvProvider<...>
val it: unit = ()

You don't need to override all the columns, you can skip the ones to leave as default. For example, in the titanic training dataset from Kaggle (data/Titanic.csv), if you want to rename the 3rd column (the PClass column) to Passenger Class and override the 6th column (the Fare column) to be a float instead of a decimal, you can define only that, and leave the other columns as blank in the schema (you also don't need to add all the trailing commas).

type Titanic1 =
    CsvProvider<"../data/Titanic.csv", Schema=",,Passenger Class,,,float", ResolutionFolder=ResolutionFolder>

let titanic1 = Titanic1.GetSample()

for row in titanic1.Rows |> Seq.truncate 10 do
    printfn "%s Class = %d Fare = %g" row.Name row.``Passenger Class`` row.Fare

Braund, Mr. Owen Harris Class = 3 Fare = 7.25
Cumings, Mrs. John Bradley (Florence Briggs Thayer) Class = 1 Fare = 71.2833
Heikkinen, Miss. Laina Class = 3 Fare = 7.925
Futrelle, Mrs. Jacques Heath (Lily May Peel) Class = 1 Fare = 53.1
Allen, Mr. William Henry Class = 3 Fare = 8.05
Moran, Mr. James Class = 3 Fare = 8.4583
McCarthy, Mr. Timothy J Class = 1 Fare = 51.8625
Palsson, Master. Gosta Leonard Class = 3 Fare = 21.075
Johnson, Mrs. Oscar W (Elisabeth Vilhelmina Berg) Class = 3 Fare = 11.1333
Nasser, Mrs. Nicholas (Adele Achem) Class = 2 Fare = 30.0708
type Titanic1 = CsvProvider<...>
val titanic1: CsvProvider<...>
val it: unit = ()

Alternatively, you can rename and override the type of any column by name instead of by position:

type Titanic2 =
    CsvProvider<"../data/Titanic.csv", Schema="Fare=float,PClass->Passenger Class", ResolutionFolder=ResolutionFolder>

let titanic2 = Titanic2.GetSample()

for row in titanic2.Rows |> Seq.truncate 10 do
    printfn "%s Class = %d Fare = %g" row.Name row.``Passenger Class`` row.Fare

Braund, Mr. Owen Harris Class = 3 Fare = 7.25
Cumings, Mrs. John Bradley (Florence Briggs Thayer) Class = 1 Fare = 71.2833
Heikkinen, Miss. Laina Class = 3 Fare = 7.925
Futrelle, Mrs. Jacques Heath (Lily May Peel) Class = 1 Fare = 53.1
Allen, Mr. William Henry Class = 3 Fare = 8.05
Moran, Mr. James Class = 3 Fare = 8.4583
McCarthy, Mr. Timothy J Class = 1 Fare = 51.8625
Palsson, Master. Gosta Leonard Class = 3 Fare = 21.075
Johnson, Mrs. Oscar W (Elisabeth Vilhelmina Berg) Class = 3 Fare = 11.1333
Nasser, Mrs. Nicholas (Adele Achem) Class = 2 Fare = 30.0708
type Titanic2 = CsvProvider<...>
val titanic2: CsvProvider<...>
val it: unit = ()

You can even mix and match the two syntaxes like this Schema="int64,DidSurvive,PClass->Passenger Class=string"

Transforming CSV files

In addition to reading, CsvProvider also has support for transforming the row collection of CSV files. The operations available are Filter, Take, TakeWhile, Skip, SkipWhile, and Truncate. All these operations preserve the schema, so after transforming, you can save the results by using one of the overloads of the Save method. You can also use the SaveToString() to get the output directly as a string.

// Saving the first 10 rows that don't have missing values to a new csv file
airQuality
    .Filter(fun row ->
        not (Double.IsNaN row.Ozone)
        && not (Double.IsNaN row.``Solar.R``))
    .Truncate(10)
    .SaveToString()

val it: string =
  "Ozone;Solar.R;Wind;Temp;Month;Day
41;190;7.4;67;5;1
36;118;8;72;5;2
12;149;12.6;74;5;3
18;313;11.5;62;5;4
23;299;8.6;65;5;7
19;99;13.8;59;5;8
8;19;20.1;61;5;9
16;256;9.7;69;5;12
11;290;9.2;66;5;13
14;274;10.9;68;5;14
"

It's also possible to transform the columns themselves by using Map and the constructor for the Row type.

let doubleOzone =
    airQuality.Map(fun row -> AirQuality.Row(row.Ozone * 2.0, row.``Solar.R``, row.Wind, row.Temp, row.Month, row.Day))

val doubleOzone: Runtime.CsvFile<CsvProvider<...>.Row>

You can also append new rows, either by creating them directly as in the previous example, or by parsing them from a string.

let newRows =
    AirQuality.ParseRows(
        """41;190;7.4;67;5;1
        36;118;8;72;5;2"""
    )

let airQualityWithExtraRows = airQuality.Append newRows

val newRows: CsvProvider<...>.Row array =
  [|(41.0, 190.0, 7.4M, 67, 5, 1); (36.0, 118.0, 8M, 72, 5, 2)|]
val airQualityWithExtraRows: Runtime.CsvFile<CsvProvider<...>.Row>

It's even possible to create csv files without parsing at all:

type MyCsvType = CsvProvider<Schema="A (int), B (string), C (date option)", HasHeaders=false>

let myRows =
    [ MyCsvType.Row(1, "a", None)
      MyCsvType.Row(2, "B", Some System.DateTime.Now) ]

let myCsv = new MyCsvType(myRows)
myCsv.SaveToString()

type MyCsvType = CsvProvider<...>
val myRows: CsvProvider<...>.Row list =
  [(1, "a", None); (2, "B", Some 3/21/2025 3:22:31 PM)]
val myCsv: MyCsvType
val it: string = "1,a,
2,B,2025-03-21T15:22:31.4907854+00:00
"

Handling big datasets

By default, the rows are cached so you can iterate over the Rows property multiple times without worrying. But if you will only iterate once, you can disable caching by setting the CacheRows static parameter of CsvProvider to false. If the number of rows is very big, you have to do this otherwise you may exhaust the memory. You can still cache the data at some point by using the Cache method, but only do that if you have already transformed the dataset to be smaller.

Related articles

• Using JSON provider in a library also applies to CSV type provider
• CSV Parser - provides more information about working with CSV documents dynamically.
• API Reference: CsvProvider type provider