ML.NET CV repro

repro.fsx

#r "nuget:FSharp.Stats"
#r "nuget: Microsoft.ML, 1.7.*"
#r "nuget: Microsoft.ML.FastTree"
#r "nuget: FSharp.Data"
#r "nuget: Plotly.NET, 2.0.*"


open System
open System.IO
open System.IO.Compression
open System.Text.Json
open System.Net
open System
open FSharp.Data
open FSharp.Stats
open Plotly.NET
open Microsoft.ML
open Microsoft.ML.Data
open Microsoft.ML.Transforms.Text

Environment.CurrentDirectory <- __SOURCE_DIRECTORY__

let download (inputUrl:string) (outputFile:string) =
    Directory.CreateDirectory(Path.GetDirectoryName(outputFile)) |> ignore
    if IO.File.Exists(outputFile) then
        printfn $"The file {outputFile} already exists. Skipping download" 
    else
        let web = Http.RequestStream(inputUrl)
        use fileStream = IO.File.Create(outputFile)
        web.ResponseStream.CopyTo(fileStream)
        fileStream.Close()

// Decompress a gzip file
let gunzip (inputFile:string) (outputFile:string) =
    Directory.CreateDirectory(Path.GetDirectoryName(outputFile)) |> ignore
    if File.Exists(outputFile) then File.Delete(outputFile)
    use inputStream = File.OpenRead(inputFile)
    use outputStream = File.Create(outputFile)
    use gzipStream = new GZipStream(inputStream, CompressionMode.Decompress)
    gzipStream.CopyTo(outputStream)

let nq100FullUrl = "https://www.dropbox.com/s/izcsjp06lgwbauu/Nasdaq100CallFull.json.gz?dl=1"
let dataFolder = "data"
let nqFullFile = Path.Combine(dataFolder, "Nasdaq100CallFull.json")
let nq100FullFileGz = nqFullFile.Replace(".json", ".json.gz")

download nq100FullUrl nq100FullFileGz
gunzip nq100FullFileGz nqFullFile


type CallId =
    { Ticker: string
      Exchange: string
      FiscalQuarter: int
      Date: DateTime }

type CallFull = 
    { CallId: CallId
      Header: string
      PreparedRemarks: string
      QuestionsAndAnswers: string
      Label: float }

let nq100Full = 
    File.ReadAllText(nqFullFile)
    |> JsonSerializer.Deserialize<List<CallFull>>



[<CLIMutable>]
type BinarySentimentInput =
    { Label: bool
      Text: string }

[<CLIMutable>]
type BinarySentimentOutput =
    { PredictedLabel: bool
      Probability: single
      Score: single }

let ctx = new MLContext(seed = 1)


let nq100FullSentiment =
    nq100Full
    |> Seq.map (fun x ->
        { Label = x.Label > 0.0
          Text =  x.QuestionsAndAnswers })
    |> ctx.Data.LoadFromEnumerable 

let featurizePipeline = 
    ctx.Transforms.Text.FeaturizeText(
        outputColumnName = "Features", 
        inputColumnName = "Text")

let treeTrainer = 
    ctx.BinaryClassification.Trainers.FastTree(
        labelColumnName = "Label",
        featureColumnName = "Features")


let treePipeline = featurizePipeline.Append(treeTrainer)

let downcastPipeline (pipeline : IEstimator<'a>) =
    match pipeline with
    | :? IEstimator<ITransformer> as p -> p
    | _ -> failwith "The pipeline has to be an instance of IEstimator<ITransformer>."


let cvResults = 
    ctx.BinaryClassification
        .CrossValidate(data = nq100FullSentiment, 
                       estimator = downcastPipeline treePipeline,
                       numberOfFolds=5,
                       seed = 1)