after dusk, before dawn: [C#] Test MNIST Classifier with image using CNTK

CNTK - The Microsoft Cognitive Toolkit - is a unified deep-learning toolkit by Microsoft. This time we will using CNTK to run up a MNIST Classifier and test it by our selected image.
CNTK - 微軟認知工具(?) - 是一套微軟的整合性深度學習工具。這次我們將使用CNTK來執行一個MNIST分類器，並使用我們所選擇的圖片進行測試。

Due to Microsoft didn't provide much CNTK support for C#, so the best way is start from the CNTK C#/.NET API training examples, in here you can get some code example for training/validate model and the dataset that you need for training/validate. Copy example code to get a new post not my sytle, so I will skip it, assume you already get all things complete and let the MNISTClassifier model trained done.
由於微軟沒有提供很多C#的CNTK支援，所以最佳的方式是從CNTK C#/.NET API training examples這範例開始，在這裡你可以取得一些訓練/驗證模型的程式範例，還有訓練/驗證所需的資料。拷貝範例程式就生出一篇新的文章不是我的風格，所以我會跳過這段，假設你已經取得所有東西而且訓練好MNISTClassifier的模型了。

In Microsoft's example, you can't find how to load one image and evaluate it by pre-trained model, it only show up using MinibatchSource to validate it. After very long time search and study, finally I found a way to load image to test model without make MinibatchSource.
在微軟的範例裡，你找不到如何讀取一張圖片然後用訓練好的模型去算出結果，範例只有教怎麼用MinibatchSource去做驗證。經過很長時間的搜尋跟研究，我終於找到一個方法可以不透過MinibatchSource就讀入圖片並用模型測試。

Here is my example code, at first, choose GPU as our calculate device and set model and input image's file location.
接下來就是我的範例，首先，選擇GPU作為計算裝置，然後填好模型跟輸入圖片的檔案位置。

            DeviceDescriptor device = DeviceDescriptor.GPUDevice(0);

            string modelFile = "MNISTConvolution.model";
            string imageFile = "3.png";

Then, create a CNTK Value object from our input image. To do it, we need convert image to float array, it is not hard, but note I used Cv2.BitwiseNot to inverse image, because the dataset that we training MNISTClassifier model are black background, but our input image is white background.
然後，將我們輸入的圖片轉成CNTK Value物件。要做到這個，我們必須將圖片轉成浮點數陣列，這不難，但注意我有用Cv2.BitwiseNot將圖像反轉，因為訓練MNISTClassifier模型的資料是黑色背景的，而我們輸入的圖片是白色背景。

            int height = 28, width = 28;
            int[] imageDim = {height, width, 1};
            var org = Cv2.ImRead(imageFile, ImreadModes.GrayScale);
            var image = new Mat();
            Cv2.BitwiseNot(org,image);
            var img = new float[height * width];
            for (int i = 0; i < height; i++)
            {
                for (int j = 0; j < width; j++)
                {
                    img[i * height + j] = image.At(i, j);
                }
            }
            var imageBatch = Value.CreateBatch(imageDim, img, device);

After input image ready, we can setup model and evaluate the result.
輸入圖片準備好後，我們就可以設定模型然後計算結果了。

            //load model
            Function model = Function.Load(modelFile, device);

            //setting
            var inputDataMap = new Dictionary {{model.Arguments[0], imageBatch}};
            Variable outputVar = model.Output;
            var outputDataMap = new Dictionary {{outputVar, null}};

            //evaluate result
            model.Evaluate(inputDataMap, outputDataMap, device);

The result will be a list of 10 value (0~9), higher value mean higher chance, we only print out top 3 result.
結果將是10個數值（0〜9）的列表，值越高意味著越高的可能性，這裡我們只印出前3個結果。

            //get result
            var outputVal = outputDataMap[outputVar];
            var actual = outputVal.GetDenseData(outputVar).First();

            //get top 3 label
            var actualLabels = actual.Select((x, index) => new {label = index, value = x}).ToList();
            var list = actualLabels.OrderByDescending(x => x.value).Take(3);
            foreach (var v in list)
            {
                Console.WriteLine($"Label:{v.label} Value:{v.value}");
            }

Here is other test result.
這裡是其他的一些測試結果。

Full code here.
完整程式碼。

using System;
using System.Collections.Generic;
using System.Diagnostics;
using System.Linq;
using CNTK;
using OpenCvSharp;

namespace CntkMnistTrain
{
    class Program
    {
        static void Main()
        {
            DeviceDescriptor device = DeviceDescriptor.GPUDevice(0);

            string modelFile = "MNISTConvolution.model";
            string imageFile = "3.png";

            #region prepare test image
            int height = 28, width = 28;
            int[] imageDim = {height, width, 1};
            var org = Cv2.ImRead(imageFile, ImreadModes.GrayScale);
            var image = new Mat();
            Cv2.BitwiseNot(org,image);
            var img = new float[height * width];
            for (int i = 0; i < height; i++)
            {
                for (int j = 0; j < width; j++)
                {
                    img[i * height + j] = image.At(i, j);
                }
            }
            var imageBatch = Value.CreateBatch(imageDim, img, device);
            #endregion

            //load model
            Function model = Function.Load(modelFile, device);

            //setting
            var inputDataMap = new Dictionary {{model.Arguments[0], imageBatch}};
            Variable outputVar = model.Output;
            var outputDataMap = new Dictionary {{outputVar, null}};

            Stopwatch sw = new Stopwatch();
            sw.Start();

            //evaluate result
            model.Evaluate(inputDataMap, outputDataMap, device);

            sw.Stop();
            Console.WriteLine($"Evaluate cost {sw.ElapsedMilliseconds} ms");

            //get result
            var outputVal = outputDataMap[outputVar];
            var actual = outputVal.GetDenseData(outputVar).First();

            //get top 3 label
            var actualLabels = actual.Select((x, index) => new {label = index, value = x}).ToList();
            var list = actualLabels.OrderByDescending(x => x.value).Take(3);
            foreach (var v in list)
            {
                Console.WriteLine($"Label:{v.label} Value:{v.value}");
            }

            //display image
            using (new Window("image", org))
            {
                Cv2.WaitKey();
            }
        }
    }
}

Hope you enjoy it.

BTW, if using Visual Studio, you can get CNTK from nuget.
順帶一提，如果用Visual Studio的話，你可以從nuget取得CNTK。