//

open TensorFlow

open System.IO

open System

open System.Net

open System.IO.Compression

let FetchModelFiles =

let inceptionUrl = "https://storage.googleapis.com/download.tensorflow.org/models/inception5h.zip"

let dir = Path.Combine (Path.GetTempPath (), "Inception")

let modelFile = Path.Combine (dir, "tensorflow_inception_graph.pb")

let labelsFile = Path.Combine (dir, "imagenet_comp_graph_label_strings.txt")

let zipfile = Path.Combine (dir, "inception5h.zip")

if (not (File.Exists (modelFile) && File.Exists (labelsFile))) then

printfn "Downloading inception5h model and labels"

Directory.CreateDirectory (dir) |> ignore

let wc = new WebClient ()

wc.DownloadFile (inceptionUrl, zipfile)

ZipFile.ExtractToDirectory (zipfile, dir)

File.Delete (zipfile)

(modelFile, labelsFile)

let iconst (graph:TFGraph) (v:int) (label:string) =

graph.Const (new TFTensor (v), label)

let fconst (graph:TFGraph) (v:float32) (label:string) =

graph.Const (new TFTensor (v), label)

let constructGraphToNormalizeImage =

// Some constants specific to the pre-trained model at:

// https://storage.googleapis.com/download.tensorflow.org/models/inception5h.zip

//

// - The model was trained after with images scaled to 224x224 pixels.

// - The colors, represented as R, G, B in 1-byte each were converted to

// float using (value - Mean)/Scale.

let width = 224

let height = 224

let mean = 117.f

let scale = 1.f

let graph = new TFGraph ()

let input = graph.Placeholder TFDataType.String

let image = graph.Cast (graph.DecodeJpeg (contents = input, channels = System.Nullable 3L), DstT = TFDataType.Float)

let expand = graph.ExpandDims (input = image, dim = graph.Const (TFTensor.op_Implicit (0), "make_batch"))

let size = graph.Const (TFTensor.op_Implicit [|width;height |], "size")

let resize = graph.ResizeBilinear (images = expand, size = size)

let mean = fconst graph mean "Mean"

let scale = fconst graph scale "Scale"

let output = graph.Div (x = graph.Sub (x = resize, y = mean), y = scale)

(graph, input, output)

let createTensorFromImageFile f =

let tensor = File.ReadAllBytes f |> TFTensor.CreateString

let (graph, input, output) = constructGraphToNormalizeImage

use session = new TFSession (graph)

let normalized = session.Run (runOptions = null, inputs = [| input |], inputValues = [| tensor |], outputs = [| output |])

normalized.[0]

let main argv =

let (modelFile, labelFile) = FetchModelFiles

let labels = File.ReadAllLines (labelFile)

use graph = new TFGraph()

graph.Import (File.ReadAllBytes (modelFile), "")

use session = new TFSession (graph)

let files = if argv.Length = 0 then [| "/tmp/demo.jpg" |] else argv

for file in files do

let tensor = createTensorFromImageFile file

// Run the session

let runner = session.GetRunner ();

runner.AddInput (graph.["input"].[0], tensor)

runner.Fetch (graph.["output"].[0]);

let sessionOutput = runner.Run ();

//

// Other style of invoking the runner:

// let sessionOutput = session.Run (inputs = [| graph.["input"].[0] |], inputValues= [| tensor |], outputs = [| graph.["output"].[0] |])

//

let resultTensor = sessionOutput.[0]

// The GetValue method returns an 'object' type, we need to cast this

// to a float[] [] array, and the results will be on the first row.

let probabilities = (resultTensor.GetValue (jagged = true) :?> float32 [] []).[0]

let mutable bestIdx = 0

let mutable best = 0.f

for i in [0..probabilities.Length-1] do

if probabilities.[i] > best then

best <- probabilities.[i]

bestIdx <- i

printfn "%s - best match [%d]=%g %s" file bestIdx best labels.[bestIdx]

0 // return an integer exit code