自己打造一个深度学习框架 for java
前言
从16年开始利用空余时间研究深度学习的方面，由于工作的原因，最熟悉的编程语言就是java，所以框架的编程语言自然而然就使用了java。自己打造框架的初衷就是为了更加深入了解各个算法、模型、实现的原理和思路。
框架介绍
Omega-AI：基于java打造的深度学习框架，帮助你快速搭建神经网络，实现训练或测试模型，支持多线程运算，框架目前支持BP神经网络、卷积神经网络、vgg16、resnet、yolo等模型的构建，目前引擎最新版本支持CUDA和CUDNN两种GPU加速方式，关于GPU加速的环境配置与jcuda版本jar包的对应依赖，可添加QQ群(119593195)进行咨询讨论。
源码地址：
https://gitee.com/iangellove/omega-ai
https://github.com/iangellove/Omega-AI
依赖
由于omega-engine-1.0.3加入了jcuda支持，所以1.0.3需要安装与jcuda版本对应的cuda，我在该项目中使用的是jcuda-11.2.0版本的包，那么我cuda需要安装11.2.x版本
系统参数
由于训练vgg16模型的参数比较庞大，所以在部署项目的时候需要对jvm内存进行调整.
调整事例如：-Xmx20480m -Xms20480m -Xmn10240m
Demo展示
功能介绍
支持的网络层类型：
Fullylayer 全连接层
ConvolutionLayer 卷积层
PoolingLayer 池化层(maxpooling,meanpooling)
AVGPooingLayer 全局平均池化层
激活函数层
SoftmaxLayer (softmax激活函)
ReluLayer
LeakyReluLayer
TanhLayer
SigmodLayer
归一化层
BNLayer (Batch Normalization)
DropoutLayer
优化器
Momentum
Adam
Adamw
Sgd (sgd with momentum)
训练器
BGDOptimizer (批量梯度下降法)
MBSGDOptimizer (小批量随机梯度下降)
SGDOptimizer（随机梯度下降算法）
损失函数(loss function)
MSELoss (平方差损失函数)
CrossEntropyLoss (交叉熵损失函数)
CrossEntropyLossWithSoftmax (交叉熵损失 + softmax)
MultiLabelSoftMargin (多标签损失函数)
学习率更新器（LearnRateUpdate）
NONE (固定学习率)
LR_DECAY (decay)
GD_GECAY (gd_decay)
CONSTANT(gd_decay)
RANDOM [Math.pow(RandomUtils.getInstance().nextFloat(), power) * this.lr]
POLY [this.lr * Math.pow((1.0f - (batchIndex * 1.0f / trainTime / dataSize * batchSize)), power)]
STEP [this.lr * Math.pow(this.scale, batchIndex / step)]
EXP [this.lr * Math.pow(this.gama, batchIndex)]
SIG [this.lr / (1 + Math.pow(Math.E, this.gama * (batchIndex - step)))]
数据加载器
.bin (二进制数据文件)
.idx3-ubyte
.txt
使用说明
自带的数据集
iris（鸢尾花数据集）
mnist（手写数字数据集）
cifar_10 （cifar_10数据集）
附加数据集
数据集成绩
iris epoch:5 bp神经网络[3层全连接层] 测试数据集准确率100%
mnist epoch:10 alexnet 测试数据集准确率98.6%
cifar_10 epoch:50 alexnet 测试数据集准确率76.6%
cifar_10 epoch:50 vgg16 测试数据集准确率86.45%
cifar_10 epoch:300 resnet18 [batchSize:128,初始learningRate:0.1,learnRateUpdate:GD_GECAY,optimizer:adamw] 数据预处理[randomCrop,randomHorizontalFilp,cutout,normalize] 测试数据集准确率91.23%
事例代码
bp iris demo
public void bpNetwork_iris() {
// TODO Auto-generated method stub
/**
* 读取训练数据集
*/
String iris_train = "/dataset/iris/iris.txt";
String iris_test = "/dataset/iris/iris_test.txt";
String[] labelSet = new String[] {"1","-1"};
DataSet trainData = DataLoader.loalDataByTxt(iris_train, ",", 1, 1, 4, 2,labelSet);
DataSet testData = DataLoader.loalDataByTxt(iris_test, ",", 1, 1, 4, 2,labelSet);
System.out.println("train_data:"+JsonUtils.toJson(trainData));
BPNetwork netWork = new BPNetwork(new SoftmaxWithCrossEntropyLoss());
InputLayer inputLayer = new InputLayer(1,1,4);
FullyLayer hidden1 = new FullyLayer(4, 40);
ReluLayer active1 = new ReluLayer();
FullyLayer hidden2 = new FullyLayer(40, 20);
ReluLayer active2 = new ReluLayer();
FullyLayer hidden3 = new FullyLayer(20, 2);
SoftmaxWithCrossEntropyLayer hidden4 = new SoftmaxWithCrossEntropyLayer(2);
netWork.addLayer(inputLayer);
netWork.addLayer(hidden1);
netWork.addLayer(active1);
netWork.addLayer(hidden2);
netWork.addLayer(active2);
netWork.addLayer(hidden3);
netWork.addLayer(hidden4);
try {
MBSGDOptimizer optimizer = new MBSGDOptimizer(netWork, 8, 0.00001d, 10, LearnRateUpdate.NONE);
optimizer.train(trainData);
optimizer.test(testData);
} catch (Exception e) {
// TODO Auto-generated catch block
e.printStackTrace();
}
}
cnn mnist demo
public void cnnNetwork_mnist() {
// TODO Auto-generated method stub
try {
/**
* 读取训练数据集
*/
String mnist_train_data = "/dataset/mnist/train-images.idx3-ubyte";
String mnist_train_label = "/dataset/mnist/train-labels.idx1-ubyte";
String mnist_test_data = "/dataset/mnist/t10k-images.idx3-ubyte";
String mnist_test_label = "/dataset/mnist/t10k-labels.idx1-ubyte";
String[] labelSet = new String[] {"0","1","2","3","4","5","6","7","8","9"};
Resource trainDataRes = new ClassPathResource(mnist_train_data);
Resource trainLabelRes = new ClassPathResource(mnist_train_label);
Resource testDataRes = new ClassPathResource(mnist_test_data);
Resource testLabelRes = new ClassPathResource(mnist_test_label);
DataSet trainData = DataLoader.loadDataByUByte(trainDataRes.getFile(), trainLabelRes.getFile(), labelSet, 1, 1 , 784, true);
DataSet testData = DataLoader.loadDataByUByte(testDataRes.getFile(), testLabelRes.getFile(), labelSet, 1, 1 , 784, true);
int channel = 1;
int height = 28;
int width = 28;
CNN netWork = new CNN(new SoftmaxWithCrossEntropyLoss(), UpdaterType.momentum);
netWork.learnRate = 0.001d;
InputLayer inputLayer = new InputLayer(channel, 1, 784);
ConvolutionLayer conv1 = new ConvolutionLayer(channel, 6, width, height, 5, 5, 2, 1, false);
BNLayer bn1 = new BNLayer();
LeakyReluLayer active1 = new LeakyReluLayer();
PoolingLayer pool1 = new PoolingLayer(conv1.oChannel, conv1.oWidth, conv1.oHeight, 2, 2, 2, PoolingType.MAX_POOLING);
ConvolutionLayer conv2 = new ConvolutionLayer(pool1.oChannel, 12, pool1.oWidth, pool1.oHeight, 5, 5, 0, 1, false);
BNLayer bn2 = new BNLayer();
LeakyReluLayer active2 = new LeakyReluLayer();
DropoutLayer drop1 = new DropoutLayer(0.5d);
PoolingLayer pool2 = new PoolingLayer(conv2.oChannel, conv2.oWidth, conv2.oHeight, 2, 2, 2, PoolingType.MAX_POOLING);
int fInputCount = pool2.oChannel * pool2.oWidth * pool2.oHeight;
int inputCount = (int) (Math.sqrt((fInputCount) + 10) + 10);
FullyLayer full1 = new FullyLayer(fInputCount, inputCount, false);
BNLayer bn3 = new BNLayer();
LeakyReluLayer active3 = new LeakyReluLayer();
FullyLayer full2 = new FullyLayer(inputCount, 10);
SoftmaxWithCrossEntropyLayer softmax = new SoftmaxWithCrossEntropyLayer(10);
netWork.addLayer(inputLayer);
netWork.addLayer(conv1);
netWork.addLayer(bn1);
netWork.addLayer(active1);
netWork.addLayer(pool1);
netWork.addLayer(conv2);
netWork.addLayer(bn2);
netWork.addLayer(active2);
netWork.addLayer(drop1);
netWork.addLayer(pool2);
netWork.addLayer(full1);
netWork.addLayer(bn3);
netWork.addLayer(active3);
netWork.addLayer(full2);
netWork.addLayer(softmax);
MBSGDOptimizer optimizer = new MBSGDOptimizer(netWork, 10, 0.0001d, 96, LearnRateUpdate.NONE);
long start = System.currentTimeMillis();
optimizer.train(trainData);
optimizer.test(testData);
System.out.println(((System.currentTimeMillis() - start) / 1000) + "s.");
} catch (Exception e) {
// TODO: handle exception
e.printStackTrace();
}
}
resnet cifar10 demo
public void resnet18_cifar10() {
// TODO Auto-generated method stub
try {
String[] labelSet = new String[] {"airplane","automobile","bird","cat","deer","dog","frog","horse","ship","truck"};
String[] train_data_filenames = new String[] {
"H:/dataset/cifar-10/data_batch_1.bin",
"H:/dataset/cifar-10/data_batch_2.bin",
"H:/dataset/cifar-10/data_batch_3.bin",
"H:/dataset/cifar-10/data_batch_4.bin",
"H:/dataset/cifar-10/data_batch_5.bin"
};
String test_data_filename = "H:/dataset/cifar-10/test_batch.bin";
float[] mean = new float[] {0.491f, 0.482f, 0.446f};
float[] std = new float[] {0.247f, 0.243f, 0.261f};
DataSet trainData = DataLoader.getImagesToDataSetByBin(train_data_filenames, 10000, 3, 32, 32, 10, labelSet, true);
DataSet testData = DataLoader.getImagesToDataSetByBin(test_data_filename, 10000, 3, 32, 32, 10, labelSet, true, mean, std);
System.out.println("data is ready.");
int channel = 3;
int height = 32;
int width = 32;
CNN netWork = new CNN(LossType.softmax_with_cross_entropy, UpdaterType.adamw);
netWork.CUDNN = true;
netWork.learnRate = 0.1f;
InputLayer inputLayer = new InputLayer(channel, height, width);
ConvolutionLayer conv1 = new ConvolutionLayer(channel, 64, width, height, 3, 3, 1, 1, false);
BNLayer bn1 = new BNLayer();
ReluLayer active1 = new ReluLayer();
/**
* block1 64 * 32 * 32
*/
BasicBlockLayer bl1 = new BasicBlockLayer(conv1.oChannel, 64, conv1.oHeight, conv1.oWidth, 1, netWork);
ReluLayer active2 = new ReluLayer();
/**
* block2 64 * 32 * 32
*/
BasicBlockLayer bl2 = new BasicBlockLayer(bl1.oChannel, 64, bl1.oHeight, bl1.oWidth, 1, netWork);
ReluLayer active3 = new ReluLayer();
/**
* block3 128 * 16 * 16
* downSample 32 / 2 = 16
*/
BasicBlockLayer bl3 = new BasicBlockLayer(bl2.oChannel, 128, bl2.oHeight, bl2.oWidth, 2, netWork);
ReluLayer active4 = new ReluLayer();
/**
* block4 128 * 16 * 16
*/
BasicBlockLayer bl4 = new BasicBlockLayer(bl3.oChannel, 128, bl3.oHeight, bl3.oWidth, 1, netWork);
ReluLayer active5 = new ReluLayer();
/**
* block5 256 * 8 * 8
* downSample 16 / 2 = 8
*/
BasicBlockLayer bl5 = new BasicBlockLayer(bl4.oChannel, 256, bl4.oHeight, bl4.oWidth, 2, netWork);
ReluLayer active6 = new ReluLayer();
/**
* block6 256 * 8 * 8
*/
BasicBlockLayer bl6 = new BasicBlockLayer(bl5.oChannel, 256, bl5.oHeight, bl5.oWidth, 1, netWork);
ReluLayer active7 = new ReluLayer();
/**
* block7 512 * 4 * 4
* downSample 8 / 2 = 4
*/
BasicBlockLayer bl7 = new BasicBlockLayer(bl6.oChannel, 512, bl6.oHeight, bl6.oWidth, 2, netWork);
ReluLayer active8 = new ReluLayer();
/**
* block8 512 * 4 * 4
*/
BasicBlockLayer bl8 = new BasicBlockLayer(bl7.oChannel, 512, bl7.oHeight, bl7.oWidth, 1, netWork);
ReluLayer active9 = new ReluLayer();
AVGPoolingLayer pool2 = new AVGPoolingLayer(bl8.oChannel, bl8.oWidth, bl8.oHeight);
/**
* fully 512 * 1 * 1
*/
int fInputCount = pool2.oChannel * pool2.oWidth * pool2.oHeight;
FullyLayer full1 = new FullyLayer(fInputCount, 10);
netWork.addLayer(inputLayer);
netWork.addLayer(conv1);
netWork.addLayer(bn1);
netWork.addLayer(active1);
/**
* block1 64
*/
netWork.addLayer(bl1);
netWork.addLayer(active2);
netWork.addLayer(bl2);
netWork.addLayer(active3);
/**
* block2 128
*/
netWork.addLayer(bl3);
netWork.addLayer(active4);
netWork.addLayer(bl4);
netWork.addLayer(active5);
/**
* block3 256
*/
netWork.addLayer(bl5);
netWork.addLayer(active6);
netWork.addLayer(bl6);
netWork.addLayer(active7);
/**
* block4 512
*/
netWork.addLayer(bl7);
netWork.addLayer(active8);
netWork.addLayer(bl8);
netWork.addLayer(active9);
netWork.addLayer(pool2);
netWork.addLayer(full1);
MBSGDOptimizer optimizer = new MBSGDOptimizer(netWork, 250, 0.001f, 128, LearnRateUpdate.GD_GECAY, false);
long start = System.currentTimeMillis();
optimizer.train(trainData, testData, mean, std);
optimizer.test(testData);
System.out.println(((System.currentTimeMillis() - start) / 1000) + "s.");
} catch (Exception e) {
// TODO: handle exception
e.printStackTrace();
}finally {
try {
CUDAMemoryManager.freeAll();
} catch (Exception e) {
// TODO Auto-generated catch block
e.printStackTrace();
}
}
}
yolo banana-detection demo
public void yolov1_tiny() {
try {
String cfg_path = "H:/voc/train/yolov1-tiny.cfg";
String trainPath = "H:\\voc\\banana-detection\\bananas_train\\images";
String trainLabelPath = "H:\\voc\\banana-detection\\bananas_train\\label.csv";
String testPath = "H:\\voc\\banana-detection\\bananas_val\\images";
String testLabelPath = "H:\\voc\\banana-detection\\bananas_val\\label.csv";
YoloDataLoader trainData = new YoloDataLoader(trainPath, trainLabelPath, 1000, 3, 256, 256, 5, LabelType.csv, true);
YoloDataLoader vailData = new YoloDataLoader(testPath, testLabelPath, 100, 3, 256, 256, 5, LabelType.csv, true);
DataSet trainSet = formatToYolo(trainData.getDataSet());
DataSet vailSet = formatToYolo(vailData.getDataSet());
System.out.println("load data finish.");
CNN netWork = new CNN(LossType.yolo3, UpdaterType.adamw);
netWork.CUDNN = true;
netWork.learnRate = 0.001f;
ModelLoader.loadConfigToModel(netWork, cfg_path);
MBSGDOptimizer optimizer = new MBSGDOptimizer(netWork, 1000, 0.001f, 64, LearnRateUpdate.CONSTANT, false);
long start = System.currentTimeMillis();
optimizer.trainObjectRecognition(trainSet, vailSet);
/**
* 处理测试预测结果
*/
float[][][] draw_bbox = optimizer.showObjectRecognition(vailSet, 64);
YoloDataLoader testData = new YoloDataLoader(testPath, testLabelPath, 1000, 3, 256, 256, 5, LabelType.csv, false);
String outputPath = "H:\\voc\\banana-detection\\test\\";
showImg(outputPath, testData.getDataSet(), 1, draw_bbox, false);
System.out.println(((System.currentTimeMillis() - start) / 1000) + "s.");
} catch (Exception e) {
// TODO: handle exception
e.printStackTrace();
}finally {
try {
CUDAMemoryManager.freeAll();
} catch (Exception e) {
// TODO Auto-generated catch block
e.printStackTrace();
}
}
}
未来可期
实现rcnn、rnn、ssd、transform等算法
训练情况可视化
支持动态调参，可视化训练
彩蛋
基于神经网络+遗传算法实现AI赛车游戏
http://119.3.123.193:8011/AICar
版本更新
omega-engine-v3
2022-06-20
1.添加gup支持，使用jcuda调用cuda的cublasSgemm矩阵乘法，参考了caffe的卷积操作已将卷积操作优化成im2col+gemm实现，计算效率得到大大提高
2.添加vgg16 demo，该模型在cifar10数据集上表现为测试数据集准确率86.45%
3.利用jdk ForkJoin框架实现任务拆分，充分利用cpu多线程，提高对数组操作与计算速度
4.参考darknet对学习率更新机制进行升级，目前已支持RANDOM、POLY、STEP、EXP、SIG等多种学习率更新方法，并且实现学习率warmup功能
5.添加basicblock模块，新增resnet模型支持，目前该模型在cifar10数据集上的表现，epoch:300，测试数据集准确率为91.23%
omega-engine-v3-gpu
2022-07-02
1.开启omega-engine-v3-gpu版本开发，该版本将实现对omega-engine的gpu全面支持
2.全面优化卷积层计算，包括前向传播与反向传播.
2022-08-17
1.初步完成卷积层的gpu改造，使得卷积神经网络计算速度整体提升，增加im2col与col2im两个经典的核函数（Im2colKernel.cu，Col2imKernel.cu）
2.添加cuda内存管理器，用于管理整体显存的生命周期，减少频繁申请显存的操作，减少主机与显卡之间的数据传输.
2022-09-02
1.修改bn层计算dmean公式,减少计算量
2.更换数据存储方式，以便使用gpu计算，减少4维数组与1维数组之间的转换，获得成倍的计算效率提升
3.全面优化gpu计算，更新cuda核函数实现，使得训练与预测计算效获得大大提升
4.后续版本将进一步优化gpu版本，预计将整个计算过程搬迁入gpu计算，从而减少主机与设备(显卡)之间传输，希望进一步获得更快的计算速度
omega-engine-v4-gpu
2023-01-10
1.开启omega-engine-v4-gpu版本开发，该版本将实现对omega-engine的CUDNN全面支持
2.新增全局平均池化层实现
3.将softmax与cross_entropy结合成softmax_with_cross_entropy作为损失函数使用(注意:使用softmax_with_cross_entropy损失函数,将不需要额外添加SoftmaxLayer)
4.新增BN层对CUDNN支持，实现源码请移步(实现源码请移步BNCudnnKernel.java)
5.后续版本将逐渐实现引擎对CUDNN支持
2023-04-13
1.omega-engine-v4-gpu版本添加cudnn支持，整体推理与训练效率提升4倍
2.优化bn层，激活函数层内存使用，整体内存显存占用减少30%~40%
3.新增yolo目标识别实现，当前实现的yolo版本为yolov1版本(实现源码请移步YoloV1Test.java)
4.新增图片绘制工具，帮助绘制预测框与回显图片
5.后续版本将逐渐实现引擎对yolov3,yolov5等模型
2023-08-02
1.新增自动求导功能(包含cpu，gpu版本).
2.新增multiLabel_soft_margin loss损失函数，yolo loss（Yolov3Loss）.
3.新增yolov3目标识别实现，当前实现的yolo版本为yolov3版本(实现源码请移步YoloV3Test.java) .
4.新增目标识别数据增强功能(随机裁剪边缘，随机上下反转，hsv变换等).
5.使用自动求导功能实现MSN损失函数，代替原有的MSN loss.
6.后续版本将逐渐实现引擎对yolov5,GAN,transformer等模型支持.
2016
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
