issue400.cs

using System;
using System.Collections.Generic;
using System.Diagnostics;
using System.Diagnostics.CodeAnalysis;
using System.Drawing;
using System.Drawing.Imaging;
using System.IO;
using System.Linq;
using System.Runtime.CompilerServices;
using System.Text;
using System.Threading.Tasks;
using NumSharp;
using NumSharp.Backends;
using NumSharp.Backends.Unmanaged;
using Tensorflow;
using static Tensorflow.Binding;
using Buffer = System.Buffer;

namespace ConsoleApp2
{
    class Program
    {
        static void Main(string[] args)
        {
            var dncnn = new DnCNN();
            var l = new DirectoryInfo("./github_dataset").GetFiles().Select(f => new Bitmap(f.FullName)).ToList();
            dncnn.Train(l, l.Select(b=>(Bitmap)b.Clone()).ToList(), 100);
        }
    }

    public class DnCNN
    {
        private const int batch_size = 128;

        Tensor X, Y_, Y;
        Tensor loss;
        Operation optimizer;

        Session sess;

        public DnCNN()
        {
            X = tf.placeholder(tf.float32, shape: (-1, -1, -1, 3), name: "input_image");
            Y_ = tf.placeholder(tf.float32, shape: (-1, -1, -1, 3), name: "clean_image");

            Y = BuildModel(X);

            loss = (1.0 / batch_size) * tf.nn.relu(Y_ - Y);
            optimizer = tf.train.AdamOptimizer(0.001f, name: "AdamOptimizer").minimize(loss);

            sess = new Session();
            var init = tf.global_variables_initializer();
            sess.run(init);
        }

        private Tensor BuildModel(Tensor input, bool is_training = true)
        {
            var output = tf.layers.conv2d(input, 64, new int[] {3, 3}, name: "conv1", padding: "same");
            for (int i = 2; i < 20; i++)
            {
                output = tf.layers.conv2d(output, 64, new int[] {3, 3}, name: "conv" + i, padding: "same", use_bias: false);
            }

            output = tf.layers.conv2d(output, 3, new int[] {3, 3}, name: "conv20", padding: "same", use_bias: false);

            return input - output;
        }

        public void Train(List<Bitmap> inputImages, List<Bitmap> outputImages, int epochs = 1)
        {
            var sw = new Stopwatch();
            sw.Start();
            NDArray x_train = GenerateDataset(inputImages);
            NDArray y_train = GenerateDataset(outputImages);

            var saver = new Saver();
            print($"Dataset created in {sw.ElapsedMilliseconds}ms");
            sw.Restart();
            for (int i = 0; i < epochs; i++)
            {
                sess.run(optimizer, (X, (x_train)), (Y_, (y_train)));
                // Calculate and display the batch loss and accuracy
                var result = sess.run(new[] {loss}, new FeedItem(X, x_train), new FeedItem(Y_, y_train));
                print($"iter {i.ToString("000")}: {sw.ElapsedMilliseconds}ms");
                sw.Restart();

                saver.save(sess, @"E:\Downloads\ciao.ckpt");
            }
        }

        public Bitmap Evaluate(List<Bitmap> inputImages)
        {
            var sw = new Stopwatch();
            sw.Start();
            NDArray x_eval = GenerateDataset(inputImages);
            var output = sess.run(Y, new FeedItem(X, x_eval));
            sw.Stop();
            print($"Inference done in {sw.ElapsedMilliseconds}ms");

            output = 255 * output;
            return image(output[0].astype(NPTypeCode.Byte));
        }

        public NDArray GenerateDataset(List<Bitmap> imgs)
        {
            return np.vstack(imgs.Select(img => img.ToNDArray(false, false)).ToArray());
        }

        public Bitmap image(NDArray nd)
        {
            return nd.ToBitmap(nd.shape[2], nd.shape[1]);
        }
    }

}