Caffe任务池GPU模型图像识别

一开始我在网上找demo没有找到，在群里寻求帮助也没有得到结果，索性将网上的易语言模块反编译之后，提取出对应的dll以及代码，然后对照官方的c++代码，写出了下面的c#版本

/***
 * @pName caffe_task_pool_demo
 * @name CC
 * @user wadezh
 * @date 2018/6/16
 * @desc
 */
using System;
using System.Collections;
using System.Collections.Generic;
using System.IO;
using System.Linq;
using System.Runtime.InteropServices;
using System.Text;
using System.Threading.Tasks;

namespace caffe_task_pool_demo
{
    class CC
    {

        public static int taskPool { get; set; } = ;
        public static string prototxt { get; set; }
        public static ArrayList map { get; set; }
        public static int timeStep { get; set; }
        public static int alphabetSize { get; set; }

        /*Caffe_API TaskPool* __stdcall createTaskPoolByData(

        const void* prototxt_data,

        int prototxt_data_length,

        const void* caffemodel_data,

        int caffemodel_data_length,

        float scale_raw = 1,

        const char* mean_file = 0,

        int num_means = 0,

        float* means = 0,

        int gpu_id = -1,

        int batch_size = 3);*/

        [DllImport("classification_dll.dll", EntryPoint = "createTaskPoolByData", CallingConvention = CallingConvention.StdCall)]
        public static extern int CreateTaskPoolByData(byte[] prototxt_data,
        int prototxt_data_length,
        byte[] caffemodel_data,
        int caffemodel_data_length,
        float scale_raw = ,
        string mean_file = "",
        int num_means = ,
        float means = ,
        int gpu_id = -,
        int cach_size = );

        /*Caffe_API BlobData* __stdcall forwardByTaskPool(TaskPool* pool, const void* img, int len, const char* blob_name);*/

        [DllImport("classification_dll.dll", EntryPoint = "forwardByTaskPool", CallingConvention = CallingConvention.StdCall)]
        public static extern int ForwardByTaskPool(int poolHandle, byte[] image, int imageLen, string printBlobName);

        /*Caffe_API int __stdcall getBlobLength(BlobData* feature);*/
        [DllImport("classification_dll.dll", EntryPoint = "getBlobLength", CallingConvention = CallingConvention.StdCall)]
        public static extern int GetBlobLength(int feature);

        /*Caffe_API void __stdcall cpyBlobData(void* buffer, BlobData* feature);*/
        [DllImport("classification_dll.dll", EntryPoint = "cpyBlobData", CallingConvention = CallingConvention.StdCall)]
        public static extern int CpyBlobData(float[] buffer, int feature);

        /*Caffe_API void  __stdcall releaseBlobData(BlobData* ptr);*/
        [DllImport("classification_dll.dll", EntryPoint = "releaseBlobData", CallingConvention = CallingConvention.StdCall)]
        public static extern int ReleaseBlobData(int ptr);

        private static int Argmax(float[] arr, int begin, int end, ref float acc)
        {
            acc = -;
            int mxInd = ;
            for (int i = begin; i < end; i++)
            {
                if (acc < arr[i])
                {
                    mxInd = i;
                    acc = arr[i];
                }
            }
            return mxInd - begin;
        }

        public static bool InitCaptcha(string prototxtPath, string modelPath, string mapPath, int gpuId, int batchSize) {
            byte[] deploy = Util.GetFileStream(prototxtPath);
            byte[] model = Util.GetFileStream(modelPath);
            CC.taskPool = CC.CreateTaskPoolByData(deploy, deploy.Length, model, model.Length, 1F, "", , 0F, gpuId, batchSize);
            CC.prototxt = System.Text.Encoding.Default.GetString(deploy);
            string[] mapFile = Util.LoadStringFromFile(mapPath).Trim().Split("\r\n".ToArray());
            CC.map = new ArrayList();
            for (int i = ; i < mapFile.Length; i++)
            {
                if (mapFile[i].Length > )
                {
                    CC.map.Add(mapFile[i]);
                }
            }
            string time_step = Util.GetMiddleString(CC.prototxt, "time_step:", "\r\n");
            string layer = Util.GetMiddleString(CC.prototxt, "inner_product_param {", "{");
            string alphabet_size = Util.GetMiddleString(layer, "num_output:", "\r\n");
            CC.timeStep = int.Parse(time_step);
            CC.alphabetSize = int.Parse(alphabet_size);
            return CC.taskPool != ;
        }

        public static string GetCaptcha(byte[] image) {
            // Get the prediction result handle
            int poolHandle = CC.ForwardByTaskPool(taskPool, image, image.Length, "premuted_fc");

            // Get the tensor handle
            float[] permute_fc = new float[CC.GetBlobLength(poolHandle)];

            // Copy the tensor data
            CpyBlobData(permute_fc, poolHandle);
            string code = string.Empty;

            if (permute_fc.Length > )
            {
                int o = ;
                float acc = 0F;
                int emptyLabel = alphabetSize - ;
                int prev = emptyLabel;
                for (int i = ; i < timeStep; i++)
                {
                    o = Argmax(permute_fc, (i - ) * alphabetSize + , i * alphabetSize, ref acc);
                    if (o != emptyLabel && prev != o) code += map[o + ];
                    prev = o;
                }
                code = code.Replace("_", "").Trim();
            }

            ReleaseBlobData(poolHandle);
            return code;
        }

        protected class Util
        {

            public static byte[] GetFileStream(string path)
            {
                FileStream fs = new FileStream(path, FileMode.Open);
                long size = fs.Length;
                byte[] array = new byte[size];
                fs.Read(array, , array.Length);
                fs.Close();
                return array;
            }

            public static string LoadStringFromFile(string fileName)
            {
                string content = string.Empty;

                StreamReader sr = null;
                try
                {
                    sr = new StreamReader(fileName, System.Text.Encoding.UTF8);
                    content = sr.ReadToEnd();
                }
                catch (Exception ex)
                {
                    throw ex;
                }

                if (sr != null)
                    sr.Close();

                return content;
            }

            public static string GetMiddleString(string SumString, string LeftString, string RightString)
            {
                if (string.IsNullOrEmpty(SumString)) return "";
                if (string.IsNullOrEmpty(LeftString)) return "";
                if (string.IsNullOrEmpty(RightString)) return "";

                int LeftIndex = SumString.IndexOf(LeftString);
                if (LeftIndex == -) return "";
                LeftIndex = LeftIndex + LeftString.Length;
                int RightIndex = SumString.IndexOf(RightString, LeftIndex);
                if (RightIndex == -) return "";
                return SumString.Substring(LeftIndex, RightIndex - LeftIndex);
            }

        }

    }

}

项目中我已经将caffemodel以及prototxt等文件都打包，可以直接运行

我封装的这个CC类只支持GPU任务池识别，识别速度比较快

链接：https://pan.baidu.com/s/17tSh3IE3Xv_YlJhSOhKddg 密码：ct5z