在深度学习推理应用阶段,涉及到提速,绕不开一个关键词“预热”。
在其他地方的“预热”,预先加热到指定的温度。通常指预习准备做某一样事时,为此做好准备。
而在深度学习推理应用阶段涉及的预热通常是指GPU预热,GPU在不用的时候是低功耗状态,它会把有些高性能的功能暂时关闭或降低性能,这时候如果把模型放上面处理,能明显感觉到有点慢,甚至从点击程序运行以后要等个几秒钟才出结果,因为这个阶段GPU要完成很多初始化工作【当然了,这也和显卡好坏有关系】。
接下来的Demo案例主要是呈现预热与否的运行时间差异
首先呈现的是未经预热的第一次测试
未经预热的第二次测试
未经预热的第三次测试
预热后第一次应用
预热后第二次应用
预热后第N次应用
Demo案例预热方式仅采用部分内存释放与否的方式进行测试,仅为展示效果
附上C#UI代码
using System;
using System.Collections.Generic;
using System.ComponentModel;
using System.Data;
using System.Drawing;
using System.Linq;
using System.Text;
using System.Threading.Tasks;
using System.Windows.Forms;
using ViewControl;
using HalconDotNet;
using System.Reflection.Emit;
using static System.Net.Mime.MediaTypeNames;
namespace DeepLearningTest1
{public partial class Form1 : Form{HalconView HW; //dll调用public static DeepLearning Dl = new DeepLearning();//类调用//全局变量HObject HIMage = new HObject();HTuple hv_DLDataset = new HTuple(), hv_DLPreprocessParam = new HTuple();HTuple hv_DLModelHandle = new HTuple(), hv_ImageFiles = new HTuple();public Form1(){InitializeComponent();HW = new HalconView();HW.HWindowControl.BackColor = Color.White;splitContainer1.Panel1.Controls.Add(HW);HW.Dock = DockStyle.Fill;}/// <summary>/// 加载图片/// </summary>/// <param name="sender"></param>/// <param name="e"></param>private void button1_Click(object sender, EventArgs e){try{OpenFileDialog openFileDialog = new OpenFileDialog();//openFileDialog.InitialDirectory = AppDomain.CurrentDomain.BaseDirectory;openFileDialog.Filter = "图片文件(*.bmp;*.jpg;*.gif;*.png;*.tiff;*.tif)|*.bmp;*.jpg;*.gif;*.png;*.tiff;*.tif";openFileDialog.RestoreDirectory = true;openFileDialog.FilterIndex = 1;if (openFileDialog.ShowDialog() == DialogResult.OK){label3.Text = openFileDialog.FileName;HOperatorSet.ReadImage(out HIMage, label3.Text);HW.DispImage(HIMage, true);} }catch (Exception ex){MessageBox.Show("加载图片失败 " + ex.ToString());}}/// <summary>/// 模型预热/// </summary>Detect detectResult = new Detect();private Detect detectTest(HObject Hobj, string mode){detectResult = new Detect();HObject ho_Image = new HObject(); ;HTuple hv_WindowDict = new HTuple(), hv_Index = new HTuple();HTuple hv_SampleIndex = new HTuple(), hv_DLSampleBatch = new HTuple();HTuple hv_DLResult = new HTuple(), hv_ClaID = new HTuple();HTuple hv_ClassSorce = new HTuple(), hv_ClassName = new HTuple();HTuple hv_ID = new HTuple(), hv_Sorce = new HTuple(), hv_N = new HTuple();HOperatorSet.GenEmptyObj(out ho_Image);try{HTuple t = new HTuple(), t1 = new HTuple(), T = new HTuple();hv_DLDataset.Dispose();HOperatorSet.ReadDict("E:/Halcon数据/测量程序/深度学习/1109/xin1109.hdict",new HTuple(), new HTuple(), out hv_DLDataset);hv_DLPreprocessParam.Dispose();HOperatorSet.ReadDict("E:/Halcon数据/测量程序/深度学习/xin110900/训练-241109-153815/dl_preprocess_param.hdict",new HTuple(), new HTuple(), out hv_DLPreprocessParam);hv_DLModelHandle.Dispose();HOperatorSet.ReadDlModel("E:/Halcon数据/测量程序/深度学习/xin110900/训练-241109-153815/best_model.hdl",out hv_DLModelHandle);ho_Image.Dispose();ho_Image = Hobj;HW.DispImage(ho_Image, true);hv_DLSampleBatch.Dispose();Dl.gen_dl_samples_from_images(ho_Image, out hv_DLSampleBatch);Dl.preprocess_dl_samples(hv_DLSampleBatch, hv_DLPreprocessParam);HOperatorSet.CountSeconds(out t);hv_DLResult.Dispose();HOperatorSet.ApplyDlModel(hv_DLModelHandle, hv_DLSampleBatch, new HTuple(),out hv_DLResult);HOperatorSet.CountSeconds(out t1);T = t1 - t;label11.Text = T.ToString();hv_ClaID.Dispose();HOperatorSet.GetDictTuple(hv_DLResult, "classification_class_ids", out hv_ClaID);hv_ClassSorce.Dispose();HOperatorSet.GetDictTuple(hv_DLResult, "classification_confidences", out hv_ClassSorce);hv_ClassName.Dispose();HOperatorSet.GetDictTuple(hv_DLResult, "classification_class_names", out hv_ClassName);hv_ID.Dispose();using (HDevDisposeHelper dh = new HDevDisposeHelper()){hv_ID = hv_ClaID.TupleSelect(0);}hv_Sorce.Dispose();using (HDevDisposeHelper dh = new HDevDisposeHelper()){hv_Sorce = hv_ClassSorce.TupleSelect(0);}label1.Text = hv_Sorce.ToString();hv_N.Dispose();using (HDevDisposeHelper dh = new HDevDisposeHelper()){hv_N = hv_ClassName.TupleSelect(0);}label2.Text = hv_N.ToString();}catch{detectResult = new Detect();detectResult.finishJudge_CCD2 = false;}ho_Image.Dispose();hv_WindowDict.Dispose();hv_Index.Dispose();hv_SampleIndex.Dispose();hv_DLSampleBatch.Dispose();hv_DLResult.Dispose();hv_ClaID.Dispose();hv_ClassSorce.Dispose();hv_ClassName.Dispose();hv_ID.Dispose();hv_Sorce.Dispose();hv_N.Dispose();return detectResult;}/// <summary>/// 图片检测/// </summary>/// <param name="Hobj"></param>/// <param name="mode"></param>/// <returns></returns>private Detect detectTest1(HObject Hobj, string mode){detectResult = new Detect();HObject ho_Image = new HObject(); ;HTuple hv_WindowDict = new HTuple(), hv_Index = new HTuple();HTuple hv_SampleIndex = new HTuple(), hv_DLSampleBatch = new HTuple();HTuple hv_DLResult = new HTuple(), hv_ClaID = new HTuple();HTuple hv_ClassSorce = new HTuple(), hv_ClassName = new HTuple();HTuple hv_ID = new HTuple(), hv_Sorce = new HTuple(), hv_N = new HTuple();HOperatorSet.GenEmptyObj(out ho_Image);try{HTuple t = new HTuple(), t1 = new HTuple(), T = new HTuple();ho_Image.Dispose();ho_Image = Hobj;HW.DispImage(ho_Image, true);hv_DLSampleBatch.Dispose();Dl.gen_dl_samples_from_images(ho_Image, out hv_DLSampleBatch);Dl.preprocess_dl_samples(hv_DLSampleBatch, hv_DLPreprocessParam);HOperatorSet.CountSeconds(out t);hv_DLResult.Dispose();HOperatorSet.ApplyDlModel(hv_DLModelHandle, hv_DLSampleBatch, new HTuple(),out hv_DLResult);HOperatorSet.CountSeconds(out t1);T = t1 - t;label13.Text = T.ToString();hv_ClaID.Dispose();HOperatorSet.GetDictTuple(hv_DLResult, "classification_class_ids", out hv_ClaID);hv_ClassSorce.Dispose();HOperatorSet.GetDictTuple(hv_DLResult, "classification_confidences", out hv_ClassSorce);hv_ClassName.Dispose();HOperatorSet.GetDictTuple(hv_DLResult, "classification_class_names", out hv_ClassName);hv_ID.Dispose();using (HDevDisposeHelper dh = new HDevDisposeHelper()){hv_ID = hv_ClaID.TupleSelect(0);}hv_Sorce.Dispose();using (HDevDisposeHelper dh = new HDevDisposeHelper()){hv_Sorce = hv_ClassSorce.TupleSelect(0);}label21.Text = hv_Sorce.ToString();hv_N.Dispose();using (HDevDisposeHelper dh = new HDevDisposeHelper()){hv_N = hv_ClassName.TupleSelect(0);}label20.Text = hv_N.ToString();}catch{detectResult = new Detect();detectResult.finishJudge_CCD2 = false;}ho_Image.Dispose();hv_WindowDict.Dispose();hv_Index.Dispose();hv_SampleIndex.Dispose();hv_DLSampleBatch.Dispose();hv_DLResult.Dispose();hv_ClaID.Dispose();hv_ClassSorce.Dispose();hv_ClassName.Dispose();hv_ID.Dispose();hv_Sorce.Dispose();hv_N.Dispose();return detectResult;}/// <summary>/// 模型预热/// </summary>/// <param name="sender"></param>/// <param name="e"></param>private void button3_Click(object sender, EventArgs e){if (!HIMage.IsInitialized()) { MessageBox.Show("图片为空"); return; }Detect detectTemp = new Detect();HTuple t = new HTuple(), t1 = new HTuple(), T = new HTuple();HOperatorSet.CountSeconds(out t);detectTemp = detectTest(HIMage, "test");HOperatorSet.CountSeconds(out t1);T = t1 - t;label7.Text = T.ToString();}/// <summary>/// 图片检测/// </summary>/// <param name="sender"></param>/// <param name="e"></param>private void button2_Click(object sender, EventArgs e){if (!HIMage.IsInitialized()) { MessageBox.Show("图片为空"); return; }Detect detectTemp = new Detect();HTuple t = new HTuple(), t1 = new HTuple(), T = new HTuple();HOperatorSet.CountSeconds(out t);detectTemp = detectTest1(HIMage, "test");HOperatorSet.CountSeconds(out t1);T = t1 - t;label17.Text = T.ToString();}}
}