C# Onnx Chinese CLIP 通过一句话从图库中搜出来符合要求的图片

C# Onnx Chinese CLIP 通过一句话从图库中搜出来符合要求的图片

效果

生成图片特征

查找踢足球的小孩

测试图片

模型信息

image_model.onnx

Inputs
-------------------------
name：image
tensor：Float[1, 3, 224, 224]
---------------------------------------------------------------

Outputs
-------------------------
name：unnorm_image_features
tensor：Float[1, 512]
---------------------------------------------------------------

text_model.onnx

Inputs
-------------------------
name：text
tensor：Int64[1, 52]
---------------------------------------------------------------

Outputs
-------------------------
name：unnorm_text_features
tensor：Float[1, 512]
---------------------------------------------------------------

项目

代码

Form1.cs

 
using System;
using System.Collections.Generic;
using System.Drawing;
using System.IO;
using System.Linq;
using System.Text;
using System.Windows.Forms;
 
namespace Onnx_Demo
{
    public partial class Form1 : Form
    {
        public Form1()
        {
            InitializeComponent();
        }
 
        Clip mynet = new Clip("model/image_model.onnx", "model/text_model.onnx", "model/myvocab.txt");
 
        float[] imagedir_features;
        string image_dir = "test_img";
        StringBuilder sb = new StringBuilder();
 
        private void button2_Click(object sender, EventArgs e)
        {
            //特征向量 可以存二进制文件或者向量数据库
            imagedir_features = mynet.generate_imagedir_features(image_dir);
            txtInfo.Text = "生成完成！";
            txtInfo.Text += "有" + mynet.imgnum + "张图片,特征向量长度=" + imagedir_features.Length;
        }
 
        private void button3_Click(object sender, EventArgs e)
        {
            if (imagedir_features == null)
            {
                MessageBox.Show("请先生成图片特征！");
                return;
            }
 
            sb.Clear();
            txtInfo.Text = "";
            lblInfo.Text = "";
            pictureBox1.Image = null;
 
            string input_text = txt_input_text.Text;
            if (string.IsNullOrEmpty(input_text))
            {
                return;
            }
            List<Dictionary<string, float>> top5imglist = mynet.input_text_search_image(input_text, imagedir_features, mynet.imglist);
            sb.AppendLine("top5:");
            foreach (var item in top5imglist)
            {
                sb.AppendLine(Path.GetFileName(item.Keys.First()) + "  相似度:" + item[item.Keys.First()].ToString("F2"));
            }
 
            txtInfo.Text = sb.ToString();
            lblInfo.Text = Path.GetFileName(top5imglist[0].Keys.First());
            pictureBox1.Image = new Bitmap(top5imglist[0].Keys.First());
 
        }
 
        private void Form1_Load(object sender, EventArgs e)
        {
 
        }
    }
}

Clip.cs

public class Clip
    {
        int inpWidth = 224;
        int inpHeight = 224;
        float[] mean = new float[] { 0.48145466f, 0.4578275f, 0.40821073f };
        float[] std = new float[] { 0.26862954f, 0.26130258f, 0.27577711f };
 
        int context_length = 52;
        int len_text_feature = 512;
 
        Net net;
        float[] image_features_input;
 
        SessionOptions options;
        InferenceSession onnx_session;
        Tensor<long> input_tensor;
        List<NamedOnnxValue> input_container;
        IDisposableReadOnlyCollection<DisposableNamedOnnxValue> result_infer;
        DisposableNamedOnnxValue[] results_onnxvalue;
        Tensor<float> result_tensors;
 
        TokenizerBase tokenizer;
 
        int[] text_tokens_input;
        float[,] text_features_input;
 
        public int imgnum = 0;
        public List<string> imglist = new List<string>();
 
        public Clip(string image_modelpath, string text_modelpath, string vocab_path)
        {
            net = CvDnn.ReadNetFromOnnx(image_modelpath);
 
            // 创建输出会话，用于输出模型读取信息
            options = new SessionOptions();
            options.LogSeverityLevel = OrtLoggingLevel.ORT_LOGGING_LEVEL_INFO;
            options.AppendExecutionProvider_CPU(0);// 设置为CPU上运行
            // 创建推理模型类，读取本地模型文件
            onnx_session = new InferenceSession(text_modelpath, options);//model_path 为onnx模型文件的路径
            // 创建输入容器
            input_container = new List<NamedOnnxValue>();
 
            load_tokenizer(vocab_path);
 
        }
 
        void load_tokenizer(string vocab_path)
        {
 
            tokenizer = new TokenizerClipChinese();
            tokenizer.load_tokenize(vocab_path);
            text_tokens_input = new int[1024 * context_length];
        }
 
        Mat normalize_(Mat src)
        {
            Cv2.CvtColor(src, src, ColorConversionCodes.BGR2RGB);
 
            Mat[] bgr = src.Split();
            for (int i = 0; i < bgr.Length; ++i)
            {
                bgr[i].ConvertTo(bgr[i], MatType.CV_32FC1, 1.0 / (255.0 * std[i]), (0.0 - mean[i]) / std[i]);
            }
 
            Cv2.Merge(bgr, src);
 
            foreach (Mat channel in bgr)
            {
                channel.Dispose();
            }
 
            return src;
        }
 
        unsafe void generate_image_feature(Mat srcimg)
        {
            Mat temp_image = new Mat();
            Cv2.Resize(srcimg, temp_image, new Size(inpWidth, inpHeight), 0, 0, InterpolationFlags.Cubic);
            Mat normalized_mat = normalize_(temp_image);
            Mat blob = CvDnn.BlobFromImage(normalized_mat);
            net.SetInput(blob);
            //模型推理，读取推理结果
            Mat[] outs = new Mat[1] { new Mat() };
            string[] outBlobNames = net.GetUnconnectedOutLayersNames().ToArray();
            net.Forward(outs, outBlobNames);
            float* ptr_feat = (float*)outs[0].Data;
            int len_image_feature = outs[0].Size(1);  //忽略第0维batchsize=1, len_image_feature是定值512,跟len_text_feature相等的, 也可以写死在类成员变量里
            image_features_input = new float[len_image_feature];
            float norm = 0.0f;
            for (int i = 0; i < len_image_feature; i++)
            {
                norm += ptr_feat[i] * ptr_feat[i];
            }
            norm = (float)Math.Sqrt(norm);
            for (int i = 0; i < len_image_feature; i++)
            {
                image_features_input[i] = ptr_feat[i] / norm;
            }
        }
 
        unsafe void generate_text_feature(List<string> texts)
        {
            List<List<int>> text_token = new List<List<int>>(texts.Count);
            for (int i = 0; i < texts.Count; i++)
            {
                text_token.Add(new List<int>());
            }
 
            for (int i = 0; i < texts.Count; i++)
            {
                tokenizer.encode_text(texts[i], text_token[i]);
            }
 
            if (text_token.Count * context_length > text_tokens_input.Length)
            {
                text_tokens_input = new int[text_token.Count * context_length];
            }
 
            foreach (int i in text_tokens_input) { text_tokens_input[i] = 0; }
 
            for (int i = 0; i < text_token.Count; i++)
            {
                if (text_token[i].Count > context_length)
                {
                    Console.WriteLine("text_features index " + i + " ,bigger than " + context_length + "\n");
                    continue;
                }
                for (int j = 0; j < text_token[i].Count; j++)
                {
                    text_tokens_input[i * context_length + j] = text_token[i][j];
                }
 
            }
 
            int[] text_token_shape = new int[] { 1, context_length };
 
            text_features_input = new float[text_token.Count, len_text_feature];
 
            long[] text_tokens_input_64 = new long[texts.Count * context_length];
            for (int i = 0; i < text_tokens_input_64.Length; i++)
            {
                text_tokens_input_64[i] = text_tokens_input[i];
            }
 
            for (int i = 0; i < text_token.Count; i++)
            {
                input_tensor = new DenseTensor<long>(text_tokens_input_64, new[] { 1, 52 });
                input_container.Clear();
                input_container.Add(NamedOnnxValue.CreateFromTensor("text", input_tensor));
 
                //运行 Inference 并获取结果
                result_infer = onnx_session.Run(input_container);
 
                // 将输出结果转为DisposableNamedOnnxValue数组
                results_onnxvalue = result_infer.ToArray();
 
                // 读取第一个节点输出并转为Tensor数据
                result_tensors = results_onnxvalue[0].AsTensor<float>();
 
                float[] text_feature_ptr = results_onnxvalue[0].AsTensor<float>().ToArray();
 
                float norm = 0.0f;
                for (int j = 0; j < len_text_feature; j++)
                {
                    norm += text_feature_ptr[j] * text_feature_ptr[j];
                }
                norm = (float)Math.Sqrt(norm);
                for (int j = 0; j < len_text_feature; j++)
                {
                    text_features_input[i, j] = text_feature_ptr[j] / norm;
                }
 
            }
        }
 
        void softmax(float[] input)
        {
            int length = input.Length;
            float[] exp_x = new float[length];
            float maxVal = input.Max();
            float sum = 0;
            for (int i = 0; i < length; i++)
            {
                float expval = (float)Math.Exp(input[i] - maxVal);
                exp_x[i] = expval;
                sum += expval;
            }
            for (int i = 0; i < length; i++)
            {
                input[i] = exp_x[i] / sum;
            }
        }
 
        int[] argsort_ascend(float[] array)
        {
            int array_len = array.Length;
            int[] array_index = new int[array_len];
            for (int i = 0; i < array_len; ++i)
            {
                array_index[i] = i;
            }
 
            var temp = array_index.ToList();
 
            temp.Sort((pos1, pos2) =>
             {
 
                 if (array[pos1] < array[pos2])
                 {
                     return -1;
                 }
                 else if (array[pos1] == array[pos2])
                 {
                     return 0;
                 }
                 else
                 {
                     return 0;
                 }
 
             });
 
            return temp.ToArray();
        }
 
        public List<Dictionary<string, float>> input_text_search_image(string text, float[] image_features, List<string> imglist)
        {
 
            int imgnum = imglist.Count;
            List<string> texts = new List<string> { text };
 
            generate_text_feature(texts);
 
            float[] logits_per_image = new float[imgnum];
 
            for (int i = 0; i < imgnum; i++)
            {
                float sum = 0;
                for (int j = 0; j < len_text_feature; j++)
                {
                    sum += image_features[i * len_text_feature + j] * text_features_input[0, j]; //图片特征向量跟文本特征向量做内积
                }
                logits_per_image[i] = 100 * sum;
            }
 
            softmax(logits_per_image);
 
            int[] index = argsort_ascend(logits_per_image);
 
            List<Dictionary<string, float>> top5imglist = new List<Dictionary<string, float>>(5);
 
            for (int i = 0; i < 5; i++)
            {
                int ind = index[imgnum - 1 - i];
                Dictionary<string, float> result = new Dictionary<string, float>();
                result.Add(imglist[ind], logits_per_image[ind]);
                top5imglist.Add(result);
            }
            return top5imglist;
        }
 
        public float[] generate_imagedir_features(string image_dir)
        {
 
            imglist = Common.listdir(image_dir);
            imgnum = imglist.Count;
            Console.WriteLine("遍历到" + imgnum + "张图片");
 
            float[] imagedir_features = new float[0];
 
            for (int i = 0; i < imgnum; i++)
            {
                string imgpath = imglist[i];
 
                Mat srcimg = Cv2.ImRead(imgpath);
 
                generate_image_feature(srcimg);
 
                imagedir_features = imagedir_features.Concat(image_features_input).ToArray();
 
                srcimg.Dispose();
            }
 
            return imagedir_features;
 
        }
 
    }