ConvNetSharp性能优化及EEG数据网络设计技术问询

一、性能优化需求 #

我在时间敏感型应用中使用ConvNetSharp包，测试net.Forward功能性能时，本机运行Classify2DDemo程序单输入预测耗时约6-7毫秒。本机配置：

• CPU：2.6 GHz
• 内存：16 GB
• 系统：Win11 64位
• 显卡：NVIDIA GeForce GTX 1650 Ti

希望将预测时间从7毫秒降至2毫秒，提出以下问题：

1. 要使Classify2DDemo程序的net.Forward功能最优运行，需采用哪些标准配置？
2. 有哪些方案可将处理时间从7毫秒降至2毫秒？

二、网络设计细节确认（EEG信号处理场景） #

补充说明：我实际用ConvNetSharp做EEG信号处理，头带含4个电极，每个电极采集500个样本的epoch，共3000+段epoch用于训练，网络设计代码如下：

using ConvNetSharp.Core.Training;
using ConvNetSharp.Core;
using ConvNetSharp.Core.Layers.Double;
using System;
using ConvNetSharp.Volume;
using ConvNetSharp.Volume.Double;


public class CNNModelTraining
{
    public const int iterations = 40;
    public const int featureNumber = 2000;
    public const int featuresWithLabels = 2001;

    public CNNModelTraining()
    {
        BuilderInstance<double>.Volume = new VolumeBuilder();
    }
    public ConvNetSharp.Core.Net<double> CreateCNNNetwork()
    {
        ConvNetSharp.Core.Net<double> net1 = new Net<double>();

        net1.AddLayer(new InputLayer(500, 4, 1));
        net1.AddLayer(new ConvLayer(5, 2, 20) { Stride = 2, Pad = 0 });
        net1.AddLayer(new ReluLayer());
        net1.AddLayer(new PoolLayer(2, 1));
        net1.AddLayer(new ConvLayer(3, 1, 30) { Stride = 2, Pad = 0 });
        net1.AddLayer(new ReluLayer());
        //net1.AddLayer(new PoolLayer(3, 1));
        net1.AddLayer(new FullyConnLayer(3));
        net1.AddLayer(new SoftmaxLayer(3));

        return net1;
    }

    public TrainerBase<double> GetTrainerForNetwork(Net<double> paramNet)
    {
        return new AdamTrainer<double>(paramNet) { LearningRate = 0.00003, BatchSize = 30 };
    }

    public static void TrainModel(double[][] jaggedBalancedData, double[][] outputArray, TrainerBase<double> trainerParam, ConvNetSharp.Core.Net<double> net1Param)
    {
        Console.WriteLine("Starting to train the network ...");

        Console.WriteLine(outputArray.GetLength(0));
        Console.WriteLine(jaggedBalancedData.GetLength(0));

        var netx = BuilderInstance.Volume.SameAs(new Shape(500, 4, 1, jaggedBalancedData.GetLength(0)));
        var hotLabels = BuilderInstance.Volume.SameAs(new Shape(1, 1, 1, outputArray.GetLength(0)));

        for (var ix = 0; ix < outputArray.GetLength(0); ix++)   //Swati
        {
            Console.WriteLine($"{ix} {(int)outputArray[ix][0]}");

            hotLabels.Set(0, 0, 0, ix, (int)outputArray[ix][0]);

            for (var featureGrp = 0; featureGrp < 4; ++featureGrp)
            {
                for (int features = 0; features < 500; ++features)
                    netx.Set(features, featureGrp, 0, ix, jaggedBalancedData[ix][features + (featureGrp * 500)]);
            }
        }

        for (var iters = 0; iters < CNNModelTraining.iterations; iters++)
        {
            trainerParam.Train(netx, hotLabels);
            Console.WriteLine(trainerParam.Loss);
        }

    }

}

需确认以下问题：

1. 单样本输入层InputLayer(500, 4, 1)设计是否正确？库内部会按“4组500样本”还是“500组4样本”处理？若为后者，训练和预测结果会出错。
2. 第一个ConvLayer(5, 2, 20)是否会处理所有4个电极？还是必须设为(5, 4, 20)才能覆盖所有电极？
3. 代码中的PoolLayer是平均池化层还是最大池化层？

内容的提问来源于stack exchange，提问作者Swati Shah