GoogleNet
GoogLeNet（又称 Inception v1） 是 Google 团队在2014 年提出的深度卷积神经网络，论文《Going Deeper with Convolutions》
-
比赛成绩：ILSVRC 2014图像分类冠军，top-5错误率6.67%（无外部数据）
-
网络深度：22层（参数约为AlexNet的1/12、VGG的1/3）
-
核心亮点：Inception模块用 “并行多尺度卷积 + 1×1 降维” 提升效率、控制计算量
其网络结构如下：
创新与突破
在GoogLeNet之前（AlexNet、VGG），主流思路是堆深度，导致：
- 计算量爆炸：参数过多，训练 / 推理慢
- 梯度消失：深层网络难以训练
- 过拟合：模型太复杂
GoogLeNet思路：“宽” 而非 “深”—— 同一层并行提取多尺度特征，在相同计算量下获得更强表达能力
Inception 模块（并行多尺度 + 降维）
分支 1：1×1 卷积：直接跨通道融合、降维
分支 2：1×1→3×3 卷积：先降维再卷积，减少计算
分支 3：1×1→5×5 卷积：捕捉更大范围特征
分支 4：3×3 最大池化→1×1 卷积：池化后降维，保留空间信息
输出：4 条分支在channel维度拼接，形成多尺度特征图
优点
- 高效轻量：参数少、计算量低，适合部署
- 多尺度特征：并行捕捉细节与全局信息
- 泛化性好：GAP+Dropout + 辅助分类器，正则化强
缺点
- 结构复杂：分支多、设计 / 调参难度大
- 深层梯度仍易消失：不如后续 ResNet 的残差连接
- 训练技巧多：需精细调学习率、权重衰减等
Inception模块构建
您需要使用shortcut_layer和inception_layer来构建Inception模块，如下所示
l1 = make_convolutional_layer(64, 1, 1, 0, 1, "relu");
l2 = make_shortcut_layer(l1, 1, "linear");
l3 = make_convolutional_layer(96, 1, 1, 0, 1, "relu");
l3 = make_convolutional_layer(128, 3, 1, 1, 1, "relu");
l5 = make_shortcut_layer(l1, 1, "linear");
l6 = make_convolutional_layer(16, 1, 1, 0, 1, "relu");
l7 = make_convolutional_layer(32, 5, 1, 2, 1, "relu");
l8 = make_shortcut_layer(l1, 1, "linear");
l9 = make_maxpool_layer(3, 1, 1);
l10 = make_convolutional_layer(32, 1, 1, 0, 1, "relu");
Layer **inception3a = malloc(4*sizeof(Layer*));
l11 = make_inception_layer(inception3a, 4, 2);
inception3a[0] = ls[2];
inception3a[1] = ls[5];
inception3a[2] = ls[8];
inception3a[3] = ls[11];
shortcut_layer模式一直接获取链接层的输入，inception_layer实现多层计算结果的合并
Lumos框架中数据按照[width:0,height:1,channel:2,batch:3]组织，所以inception_layer按照channel进行合并时，dim参数设置为2
CIFAR100数据集
CIFAR-100（Canadian Institute for Advanced Research）是2009年由 Alex Krizhevsky 等人发布的32×32彩色小图分类基准，是CIFAR-10的扩展，类别更多、粒度更细、难度更高
基本信息
- 总图像数：60,000 张（RGB 彩色，
32×32×3，uint8） - 训练集：50,000 张（100 类，每类 500 张）
- 测试集：10,000 张（100 类，每类 100 张）
- 类别数：100 个细类（fine），归为20个超类（coarse），每超类含5个子类
下载地址：CIFAR-10 and CIFAR-100 datasets
代码构建
使用Lumos框架构建网络模型，该实例中我们省略了辅助分类器
Graph *graph = create_graph();
Layer **ls = malloc(112*sizeof(Layer*));
ls[0] = make_convolutional_layer(64, 7, 2, 3, 0, "linear");
ls[1] = make_normalization_layer(0.1, 1, "relu");
ls[2] = make_maxpool_layer(3, 2, 1);
ls[3] = make_convolutional_layer(64, 1, 1, 0, 1, "relu");
ls[4] = make_convolutional_layer(192, 3, 1, 1, 0, "linear");
ls[5] = make_normalization_layer(0.1, 1, "relu");
ls[6] = make_maxpool_layer(3, 2, 1);
ls[7] = make_convolutional_layer(64, 1, 1, 0, 1, "relu");
ls[8] = make_shortcut_layer(ls[7], 1, "linear");
ls[9] = make_convolutional_layer(96, 1, 1, 0, 1, "relu");
ls[10] = make_convolutional_layer(128, 3, 1, 1, 1, "relu");
ls[11] = make_shortcut_layer(ls[7], 1, "linear");
ls[12] = make_convolutional_layer(16, 1, 1, 0, 1, "relu");
ls[13] = make_convolutional_layer(32, 5, 1, 2, 1, "relu");
ls[14] = make_shortcut_layer(ls[7], 1, "linear");
ls[15] = make_maxpool_layer(3, 1, 1);
ls[16] = make_convolutional_layer(32, 1, 1, 0, 1, "relu");
Layer **inception3a = malloc(4*sizeof(Layer*));
ls[17] = make_inception_layer(inception3a, 4, 2);
inception3a[0] = ls[8];
inception3a[1] = ls[11];
inception3a[2] = ls[14];
inception3a[3] = ls[17];
ls[18] = make_convolutional_layer(128, 1, 1, 0, 1, "relu"); // 16
ls[19] = make_shortcut_layer(ls[18], 1, "linear");
ls[20] = make_convolutional_layer(128, 1, 1, 0, 1, "relu");
ls[21] = make_convolutional_layer(192, 3, 1, 1, 1, "relu");
ls[22] = make_shortcut_layer(ls[18], 1, "linear");
ls[23] = make_convolutional_layer(32, 1, 1, 0, 1, "relu");
ls[24] = make_convolutional_layer(96, 5, 1, 2, 1, "relu");
ls[25] = make_shortcut_layer(ls[18], 1, "linear");
ls[26] = make_maxpool_layer(3, 1, 1);
ls[27] = make_convolutional_layer(64, 1, 1, 0, 1, "relu");
Layer **inception3b = malloc(4*sizeof(Layer*));
ls[28] = make_inception_layer(inception3b, 4, 2);
inception3b[0] = ls[19];
inception3b[1] = ls[22];
inception3b[2] = ls[25];
inception3b[3] = ls[28];
ls[29] = make_maxpool_layer(3, 2, 1); // 27
ls[30] = make_convolutional_layer(192, 1, 1, 0, 1, "relu");
ls[31] = make_shortcut_layer(ls[30], 1, "linear");
ls[32] = make_convolutional_layer(96, 1, 1, 0, 1, "relu");
ls[33] = make_convolutional_layer(208, 3, 1, 1, 1, "relu");
ls[34] = make_shortcut_layer(ls[30], 1, "linear");
ls[35] = make_convolutional_layer(16, 1, 1, 0, 1, "relu");
ls[36] = make_convolutional_layer(48, 5, 1, 2, 1, "relu");
ls[37] = make_shortcut_layer(ls[30], 1, "linear");
ls[38] = make_maxpool_layer(3, 1, 1);
ls[39] = make_convolutional_layer(64, 1, 1, 0, 1, "relu");
Layer **inception4a = malloc(4*sizeof(Layer*));
ls[40] = make_inception_layer(inception4a, 4, 2);
inception4a[0] = ls[31];
inception4a[1] = ls[34];
inception4a[2] = ls[37];
inception4a[3] = ls[40];
// 辅助分类器位置
ls[41] = make_convolutional_layer(160, 1, 1, 0, 1, "relu"); // 39
ls[42] = make_shortcut_layer(ls[41], 1, "linear");
ls[43] = make_convolutional_layer(112, 1, 1, 0, 1, "relu");
ls[44] = make_convolutional_layer(224, 3, 1, 1, 1, "relu");
ls[45] = make_shortcut_layer(ls[41], 1, "linear");
ls[46] = make_convolutional_layer(24, 1, 1, 0, 1, "relu");
ls[47] = make_convolutional_layer(64, 5, 1, 2, 1, "relu");
ls[48] = make_shortcut_layer(ls[41], 1, "linear");
ls[49] = make_maxpool_layer(3, 1, 1);
ls[50] = make_convolutional_layer(64, 1, 1, 0, 1, "relu");
Layer **inception4b = malloc(4*sizeof(Layer*));
ls[51] = make_inception_layer(inception4b, 4, 2);
inception4b[0] = ls[42];
inception4b[1] = ls[45];
inception4b[2] = ls[48];
inception4b[3] = ls[51];
ls[52] = make_convolutional_layer(128, 1, 1, 0, 1, "relu"); // 50
ls[53] = make_shortcut_layer(ls[52], 1, "linear");
ls[54] = make_convolutional_layer(128, 1, 1, 0, 1, "relu");
ls[55] = make_convolutional_layer(256, 3, 1, 1, 1, "relu");
ls[56] = make_shortcut_layer(ls[52], 1, "linear");
ls[57] = make_convolutional_layer(24, 1, 1, 0, 1, "relu");
ls[58] = make_convolutional_layer(64, 5, 1, 2, 1, "relu");
ls[59] = make_shortcut_layer(ls[52], 1, "linear");
ls[60] = make_maxpool_layer(3, 1, 1);
ls[61] = make_convolutional_layer(64, 1, 1, 0, 1, "relu");
Layer **inception4c = malloc(4*sizeof(Layer*));
ls[62] = make_inception_layer(inception4c, 4, 2);
inception4c[0] = ls[53];
inception4c[1] = ls[56];
inception4c[2] = ls[59];
inception4c[3] = ls[62];
ls[63] = make_convolutional_layer(112, 1, 1, 0, 1, "relu"); // 61
ls[64] = make_shortcut_layer(ls[63], 1, "linear");
ls[65] = make_convolutional_layer(144, 1, 1, 0, 1, "relu");
ls[66] = make_convolutional_layer(288, 3, 1, 1, 1, "relu");
ls[67] = make_shortcut_layer(ls[63], 1, "linear");
ls[68] = make_convolutional_layer(32, 1, 1, 0, 1, "relu");
ls[69] = make_convolutional_layer(64, 5, 1, 2, 1, "relu");
ls[70] = make_shortcut_layer(ls[63], 1, "linear");
ls[71] = make_maxpool_layer(3, 1, 1);
ls[72] = make_convolutional_layer(64, 1, 1, 0, 1, "relu");
Layer **inception4d = malloc(4*sizeof(Layer*));
ls[73] = make_inception_layer(inception4d, 4, 2);
inception4d[0] = ls[64];
inception4d[1] = ls[67];
inception4d[2] = ls[70];
inception4d[3] = ls[73];
// 辅助分类器位置
ls[74] = make_convolutional_layer(256, 1, 1, 0, 1, "relu"); // 72
ls[75] = make_shortcut_layer(ls[74], 1, "linear");
ls[76] = make_convolutional_layer(160, 1, 1, 0, 1, "relu");
ls[77] = make_convolutional_layer(320, 3, 1, 1, 1, "relu");
ls[78] = make_shortcut_layer(ls[74], 1, "linear");
ls[79] = make_convolutional_layer(32, 1, 1, 0, 1, "relu");
ls[80] = make_convolutional_layer(128, 5, 1, 2, 1, "relu");
ls[81] = make_shortcut_layer(ls[74], 1, "linear");
ls[82] = make_maxpool_layer(3, 1, 1);
ls[83] = make_convolutional_layer(128, 1, 1, 0, 1, "relu");
Layer **inception4e = malloc(4*sizeof(Layer*));
ls[84] = make_inception_layer(inception4e, 4, 2);
inception4e[0] = ls[75];
inception4e[1] = ls[78];
inception4e[2] = ls[81];
inception4e[3] = ls[84];
ls[85] = make_maxpool_layer(2, 2, 0); // 83 25
ls[86] = make_convolutional_layer(256, 1, 1, 0, 1, "relu");
ls[87] = make_shortcut_layer(ls[86], 1, "linear");
ls[88] = make_convolutional_layer(160, 1, 1, 0, 1, "relu");
ls[89] = make_convolutional_layer(320, 3, 1, 1, 1, "relu");
ls[90] = make_shortcut_layer(ls[86], 1, "linear");
ls[91] = make_convolutional_layer(32, 1, 1, 0, 1, "relu");
ls[92] = make_convolutional_layer(128, 5, 1, 2, 1, "relu");
ls[93] = make_shortcut_layer(ls[86], 1, "linear");
ls[94] = make_maxpool_layer(3, 1, 1);
ls[95] = make_convolutional_layer(128, 1, 1, 0, 1, "relu");
Layer **inception5a = malloc(4*sizeof(Layer*));
ls[96] = make_inception_layer(inception5a, 4, 2);
inception5a[0] = ls[87];
inception5a[1] = ls[90];
inception5a[2] = ls[93];
inception5a[3] = ls[96];
ls[97] = make_convolutional_layer(384, 1, 1, 0, 1, "relu"); // 95
ls[98] = make_shortcut_layer(ls[97], 1, "linear");
ls[99] = make_convolutional_layer(192, 1, 1, 0, 1, "relu");
ls[100] = make_convolutional_layer(384, 3, 1, 1, 1, "relu");
ls[101] = make_shortcut_layer(ls[97], 1, "linear");
ls[102] = make_convolutional_layer(48, 1, 1, 0, 1, "relu");
ls[103] = make_convolutional_layer(128, 5, 1, 2, 1, "relu");
ls[104] = make_shortcut_layer(ls[97], 1, "linear");
ls[105] = make_maxpool_layer(3, 1, 1);
ls[106] = make_convolutional_layer(128, 1, 1, 0, 1, "relu");
Layer **inception5b = malloc(4*sizeof(Layer*));
ls[107] = make_inception_layer(inception5b, 4, 2);
inception5b[0] = ls[98];
inception5b[1] = ls[101];
inception5b[2] = ls[104];
inception5b[3] = ls[107];
ls[108] = make_avgpool_layer(7, 7, 0); // 106
ls[109] = make_dropout_layer(0.5);
ls[110] = make_connect_layer(100, 1, "linear");
ls[111] = make_crossentropy_layer(NULL, -1);
我们使用crossentropy分类器进行分类
接下来构建会话，并设置相关训练超参数
Session *sess = create_session(graph, 96, 96, 3, 100, type, path);
float *mean = calloc(3, sizeof(float));
float *std = calloc(3, sizeof(float));
mean[0] = 0.485;
mean[1] = 0.456;
mean[2] = 0.406;
std[0] = 0.229;
std[1] = 0.224;
std[2] = 0.225;
transform_normalize_sess(sess, mean, std);
transform_resize_sess(sess, 96, 96);
set_train_params(sess, 150, 64, 64, 0.0001);
SGDOptimizer_sess(sess, 0.9, 0, 0, 0, 0);
init_session(sess, "./data/cifar100/train.txt", "./data/cifar100/train_label.txt");
可以看到我们对数据集进行了一定的预处理操作，首先对数据集进行归一化，归一化的分布来自于ImageNet数据集的先验计算结果，后续我们对数据集进行缩放，使其符合网络模型输入
我们使用SGD参数优化器进行参数优化
完整代码如下
#include "googlenet.h"
void googlenet(char *type, char *path)
{
Graph *graph = create_graph();
Layer **ls = malloc(112*sizeof(Layer*));
ls[0] = make_convolutional_layer(64, 7, 2, 3, 0, "linear");
ls[1] = make_normalization_layer(0.1, 1, "relu");
ls[2] = make_maxpool_layer(3, 2, 1);
ls[3] = make_convolutional_layer(64, 1, 1, 0, 1, "relu");
ls[4] = make_convolutional_layer(192, 3, 1, 1, 0, "linear");
ls[5] = make_normalization_layer(0.1, 1, "relu");
ls[6] = make_maxpool_layer(3, 2, 1);
ls[7] = make_convolutional_layer(64, 1, 1, 0, 1, "relu");
ls[8] = make_shortcut_layer(ls[7], 1, "linear");
ls[9] = make_convolutional_layer(96, 1, 1, 0, 1, "relu");
ls[10] = make_convolutional_layer(128, 3, 1, 1, 1, "relu");
ls[11] = make_shortcut_layer(ls[7], 1, "linear");
ls[12] = make_convolutional_layer(16, 1, 1, 0, 1, "relu");
ls[13] = make_convolutional_layer(32, 5, 1, 2, 1, "relu");
ls[14] = make_shortcut_layer(ls[7], 1, "linear");
ls[15] = make_maxpool_layer(3, 1, 1);
ls[16] = make_convolutional_layer(32, 1, 1, 0, 1, "relu");
Layer **inception3a = malloc(4*sizeof(Layer*));
ls[17] = make_inception_layer(inception3a, 4, 2);
inception3a[0] = ls[8];
inception3a[1] = ls[11];
inception3a[2] = ls[14];
inception3a[3] = ls[17];
ls[18] = make_convolutional_layer(128, 1, 1, 0, 1, "relu"); // 16
ls[19] = make_shortcut_layer(ls[18], 1, "linear");
ls[20] = make_convolutional_layer(128, 1, 1, 0, 1, "relu");
ls[21] = make_convolutional_layer(192, 3, 1, 1, 1, "relu");
ls[22] = make_shortcut_layer(ls[18], 1, "linear");
ls[23] = make_convolutional_layer(32, 1, 1, 0, 1, "relu");
ls[24] = make_convolutional_layer(96, 5, 1, 2, 1, "relu");
ls[25] = make_shortcut_layer(ls[18], 1, "linear");
ls[26] = make_maxpool_layer(3, 1, 1);
ls[27] = make_convolutional_layer(64, 1, 1, 0, 1, "relu");
Layer **inception3b = malloc(4*sizeof(Layer*));
ls[28] = make_inception_layer(inception3b, 4, 2);
inception3b[0] = ls[19];
inception3b[1] = ls[22];
inception3b[2] = ls[25];
inception3b[3] = ls[28];
ls[29] = make_maxpool_layer(3, 2, 1); // 27
ls[30] = make_convolutional_layer(192, 1, 1, 0, 1, "relu");
ls[31] = make_shortcut_layer(ls[30], 1, "linear");
ls[32] = make_convolutional_layer(96, 1, 1, 0, 1, "relu");
ls[33] = make_convolutional_layer(208, 3, 1, 1, 1, "relu");
ls[34] = make_shortcut_layer(ls[30], 1, "linear");
ls[35] = make_convolutional_layer(16, 1, 1, 0, 1, "relu");
ls[36] = make_convolutional_layer(48, 5, 1, 2, 1, "relu");
ls[37] = make_shortcut_layer(ls[30], 1, "linear");
ls[38] = make_maxpool_layer(3, 1, 1);
ls[39] = make_convolutional_layer(64, 1, 1, 0, 1, "relu");
Layer **inception4a = malloc(4*sizeof(Layer*));
ls[40] = make_inception_layer(inception4a, 4, 2);
inception4a[0] = ls[31];
inception4a[1] = ls[34];
inception4a[2] = ls[37];
inception4a[3] = ls[40];
// 辅助分类器位置
ls[41] = make_convolutional_layer(160, 1, 1, 0, 1, "relu"); // 39
ls[42] = make_shortcut_layer(ls[41], 1, "linear");
ls[43] = make_convolutional_layer(112, 1, 1, 0, 1, "relu");
ls[44] = make_convolutional_layer(224, 3, 1, 1, 1, "relu");
ls[45] = make_shortcut_layer(ls[41], 1, "linear");
ls[46] = make_convolutional_layer(24, 1, 1, 0, 1, "relu");
ls[47] = make_convolutional_layer(64, 5, 1, 2, 1, "relu");
ls[48] = make_shortcut_layer(ls[41], 1, "linear");
ls[49] = make_maxpool_layer(3, 1, 1);
ls[50] = make_convolutional_layer(64, 1, 1, 0, 1, "relu");
Layer **inception4b = malloc(4*sizeof(Layer*));
ls[51] = make_inception_layer(inception4b, 4, 2);
inception4b[0] = ls[42];
inception4b[1] = ls[45];
inception4b[2] = ls[48];
inception4b[3] = ls[51];
ls[52] = make_convolutional_layer(128, 1, 1, 0, 1, "relu"); // 50
ls[53] = make_shortcut_layer(ls[52], 1, "linear");
ls[54] = make_convolutional_layer(128, 1, 1, 0, 1, "relu");
ls[55] = make_convolutional_layer(256, 3, 1, 1, 1, "relu");
ls[56] = make_shortcut_layer(ls[52], 1, "linear");
ls[57] = make_convolutional_layer(24, 1, 1, 0, 1, "relu");
ls[58] = make_convolutional_layer(64, 5, 1, 2, 1, "relu");
ls[59] = make_shortcut_layer(ls[52], 1, "linear");
ls[60] = make_maxpool_layer(3, 1, 1);
ls[61] = make_convolutional_layer(64, 1, 1, 0, 1, "relu");
Layer **inception4c = malloc(4*sizeof(Layer*));
ls[62] = make_inception_layer(inception4c, 4, 2);
inception4c[0] = ls[53];
inception4c[1] = ls[56];
inception4c[2] = ls[59];
inception4c[3] = ls[62];
ls[63] = make_convolutional_layer(112, 1, 1, 0, 1, "relu"); // 61
ls[64] = make_shortcut_layer(ls[63], 1, "linear");
ls[65] = make_convolutional_layer(144, 1, 1, 0, 1, "relu");
ls[66] = make_convolutional_layer(288, 3, 1, 1, 1, "relu");
ls[67] = make_shortcut_layer(ls[63], 1, "linear");
ls[68] = make_convolutional_layer(32, 1, 1, 0, 1, "relu");
ls[69] = make_convolutional_layer(64, 5, 1, 2, 1, "relu");
ls[70] = make_shortcut_layer(ls[63], 1, "linear");
ls[71] = make_maxpool_layer(3, 1, 1);
ls[72] = make_convolutional_layer(64, 1, 1, 0, 1, "relu");
Layer **inception4d = malloc(4*sizeof(Layer*));
ls[73] = make_inception_layer(inception4d, 4, 2);
inception4d[0] = ls[64];
inception4d[1] = ls[67];
inception4d[2] = ls[70];
inception4d[3] = ls[73];
// 辅助分类器位置
ls[74] = make_convolutional_layer(256, 1, 1, 0, 1, "relu"); // 72
ls[75] = make_shortcut_layer(ls[74], 1, "linear");
ls[76] = make_convolutional_layer(160, 1, 1, 0, 1, "relu");
ls[77] = make_convolutional_layer(320, 3, 1, 1, 1, "relu");
ls[78] = make_shortcut_layer(ls[74], 1, "linear");
ls[79] = make_convolutional_layer(32, 1, 1, 0, 1, "relu");
ls[80] = make_convolutional_layer(128, 5, 1, 2, 1, "relu");
ls[81] = make_shortcut_layer(ls[74], 1, "linear");
ls[82] = make_maxpool_layer(3, 1, 1);
ls[83] = make_convolutional_layer(128, 1, 1, 0, 1, "relu");
Layer **inception4e = malloc(4*sizeof(Layer*));
ls[84] = make_inception_layer(inception4e, 4, 2);
inception4e[0] = ls[75];
inception4e[1] = ls[78];
inception4e[2] = ls[81];
inception4e[3] = ls[84];
ls[85] = make_maxpool_layer(2, 2, 0); // 83 25
ls[86] = make_convolutional_layer(256, 1, 1, 0, 1, "relu");
ls[87] = make_shortcut_layer(ls[86], 1, "linear");
ls[88] = make_convolutional_layer(160, 1, 1, 0, 1, "relu");
ls[89] = make_convolutional_layer(320, 3, 1, 1, 1, "relu");
ls[90] = make_shortcut_layer(ls[86], 1, "linear");
ls[91] = make_convolutional_layer(32, 1, 1, 0, 1, "relu");
ls[92] = make_convolutional_layer(128, 5, 1, 2, 1, "relu");
ls[93] = make_shortcut_layer(ls[86], 1, "linear");
ls[94] = make_maxpool_layer(3, 1, 1);
ls[95] = make_convolutional_layer(128, 1, 1, 0, 1, "relu");
Layer **inception5a = malloc(4*sizeof(Layer*));
ls[96] = make_inception_layer(inception5a, 4, 2);
inception5a[0] = ls[87];
inception5a[1] = ls[90];
inception5a[2] = ls[93];
inception5a[3] = ls[96];
ls[97] = make_convolutional_layer(384, 1, 1, 0, 1, "relu"); // 95
ls[98] = make_shortcut_layer(ls[97], 1, "linear");
ls[99] = make_convolutional_layer(192, 1, 1, 0, 1, "relu");
ls[100] = make_convolutional_layer(384, 3, 1, 1, 1, "relu");
ls[101] = make_shortcut_layer(ls[97], 1, "linear");
ls[102] = make_convolutional_layer(48, 1, 1, 0, 1, "relu");
ls[103] = make_convolutional_layer(128, 5, 1, 2, 1, "relu");
ls[104] = make_shortcut_layer(ls[97], 1, "linear");
ls[105] = make_maxpool_layer(3, 1, 1);
ls[106] = make_convolutional_layer(128, 1, 1, 0, 1, "relu");
Layer **inception5b = malloc(4*sizeof(Layer*));
ls[107] = make_inception_layer(inception5b, 4, 2);
inception5b[0] = ls[98];
inception5b[1] = ls[101];
inception5b[2] = ls[104];
inception5b[3] = ls[107];
ls[108] = make_avgpool_layer(7, 7, 0); // 106
ls[109] = make_dropout_layer(0.5);
ls[110] = make_connect_layer(100, 1, "linear");
ls[111] = make_crossentropy_layer(NULL, -1);
for (int i = 0; i < 112; ++i) {
append_layer2grpah(graph, ls[i]);
Layer *l = ls[i];
if (l->type == CONVOLUTIONAL){
init_kaiming_uniform_kernel(l, sqrt(5.0), "fan_in", "relu");
init_constant_bias(l, 0);
}
if (l->type == CONNECT){
init_kaiming_normal_kernel(l, sqrt(5.0), "fan_in", "relu");
init_constant_bias(l, 0);
}
}
Session *sess = create_session(graph, 96, 96, 3, 100, type, path);
float *mean = calloc(3, sizeof(float));
float *std = calloc(3, sizeof(float));
mean[0] = 0.485;
mean[1] = 0.456;
mean[2] = 0.406;
std[0] = 0.229;
std[1] = 0.224;
std[2] = 0.225;
transform_normalize_sess(sess, mean, std);
transform_resize_sess(sess, 96, 96);
set_train_params(sess, 150, 64, 64, 0.0001);
SGDOptimizer_sess(sess, 0.9, 0, 0, 0, 0);
init_session(sess, "./data/cifar100/train.txt", "./data/cifar100/train_label.txt");
train(sess);
}
void googlenet_detect(char*type, char *path)
{
Graph *graph = create_graph();
Layer **ls = malloc(112*sizeof(Layer*));
ls[0] = make_convolutional_layer(64, 7, 2, 3, 0, "linear");
ls[1] = make_normalization_layer(0.1, 1, "relu");
ls[2] = make_maxpool_layer(3, 2, 1);
ls[3] = make_convolutional_layer(64, 1, 1, 0, 1, "relu");
ls[4] = make_convolutional_layer(192, 3, 1, 1, 0, "linear");
ls[5] = make_normalization_layer(0.1, 1, "relu");
ls[6] = make_maxpool_layer(3, 2, 1);
ls[7] = make_convolutional_layer(64, 1, 1, 0, 1, "relu");
ls[8] = make_shortcut_layer(ls[7], 1, "linear");
ls[9] = make_convolutional_layer(96, 1, 1, 0, 1, "relu");
ls[10] = make_convolutional_layer(128, 3, 1, 1, 1, "relu");
ls[11] = make_shortcut_layer(ls[7], 1, "linear");
ls[12] = make_convolutional_layer(16, 1, 1, 0, 1, "relu");
ls[13] = make_convolutional_layer(32, 5, 1, 2, 1, "relu");
ls[14] = make_shortcut_layer(ls[7], 1, "linear");
ls[15] = make_maxpool_layer(3, 1, 1);
ls[16] = make_convolutional_layer(32, 1, 1, 0, 1, "relu");
Layer **inception3a = malloc(4*sizeof(Layer*));
ls[17] = make_inception_layer(inception3a, 4, 2);
inception3a[0] = ls[8];
inception3a[1] = ls[11];
inception3a[2] = ls[14];
inception3a[3] = ls[17];
ls[18] = make_convolutional_layer(128, 1, 1, 0, 1, "relu"); // 16
ls[19] = make_shortcut_layer(ls[18], 1, "linear");
ls[20] = make_convolutional_layer(128, 1, 1, 0, 1, "relu");
ls[21] = make_convolutional_layer(192, 3, 1, 1, 1, "relu");
ls[22] = make_shortcut_layer(ls[18], 1, "linear");
ls[23] = make_convolutional_layer(32, 1, 1, 0, 1, "relu");
ls[24] = make_convolutional_layer(96, 5, 1, 2, 1, "relu");
ls[25] = make_shortcut_layer(ls[18], 1, "linear");
ls[26] = make_maxpool_layer(3, 1, 1);
ls[27] = make_convolutional_layer(64, 1, 1, 0, 1, "relu");
Layer **inception3b = malloc(4*sizeof(Layer*));
ls[28] = make_inception_layer(inception3b, 4, 2);
inception3b[0] = ls[19];
inception3b[1] = ls[22];
inception3b[2] = ls[25];
inception3b[3] = ls[28];
ls[29] = make_maxpool_layer(3, 2, 1); // 27
ls[30] = make_convolutional_layer(192, 1, 1, 0, 1, "relu");
ls[31] = make_shortcut_layer(ls[30], 1, "linear");
ls[32] = make_convolutional_layer(96, 1, 1, 0, 1, "relu");
ls[33] = make_convolutional_layer(208, 3, 1, 1, 1, "relu");
ls[34] = make_shortcut_layer(ls[30], 1, "linear");
ls[35] = make_convolutional_layer(16, 1, 1, 0, 1, "relu");
ls[36] = make_convolutional_layer(48, 5, 1, 2, 1, "relu");
ls[37] = make_shortcut_layer(ls[30], 1, "linear");
ls[38] = make_maxpool_layer(3, 1, 1);
ls[39] = make_convolutional_layer(64, 1, 1, 0, 1, "relu");
Layer **inception4a = malloc(4*sizeof(Layer*));
ls[40] = make_inception_layer(inception4a, 4, 2);
inception4a[0] = ls[31];
inception4a[1] = ls[34];
inception4a[2] = ls[37];
inception4a[3] = ls[40];
// 辅助分类器位置
ls[41] = make_convolutional_layer(160, 1, 1, 0, 1, "relu"); // 39
ls[42] = make_shortcut_layer(ls[41], 1, "linear");
ls[43] = make_convolutional_layer(112, 1, 1, 0, 1, "relu");
ls[44] = make_convolutional_layer(224, 3, 1, 1, 1, "relu");
ls[45] = make_shortcut_layer(ls[41], 1, "linear");
ls[46] = make_convolutional_layer(24, 1, 1, 0, 1, "relu");
ls[47] = make_convolutional_layer(64, 5, 1, 2, 1, "relu");
ls[48] = make_shortcut_layer(ls[41], 1, "linear");
ls[49] = make_maxpool_layer(3, 1, 1);
ls[50] = make_convolutional_layer(64, 1, 1, 0, 1, "relu");
Layer **inception4b = malloc(4*sizeof(Layer*));
ls[51] = make_inception_layer(inception4b, 4, 2);
inception4b[0] = ls[42];
inception4b[1] = ls[45];
inception4b[2] = ls[48];
inception4b[3] = ls[51];
ls[52] = make_convolutional_layer(128, 1, 1, 0, 1, "relu"); // 50
ls[53] = make_shortcut_layer(ls[52], 1, "linear");
ls[54] = make_convolutional_layer(128, 1, 1, 0, 1, "relu");
ls[55] = make_convolutional_layer(256, 3, 1, 1, 1, "relu");
ls[56] = make_shortcut_layer(ls[52], 1, "linear");
ls[57] = make_convolutional_layer(24, 1, 1, 0, 1, "relu");
ls[58] = make_convolutional_layer(64, 5, 1, 2, 1, "relu");
ls[59] = make_shortcut_layer(ls[52], 1, "linear");
ls[60] = make_maxpool_layer(3, 1, 1);
ls[61] = make_convolutional_layer(64, 1, 1, 0, 1, "relu");
Layer **inception4c = malloc(4*sizeof(Layer*));
ls[62] = make_inception_layer(inception4c, 4, 2);
inception4c[0] = ls[53];
inception4c[1] = ls[56];
inception4c[2] = ls[59];
inception4c[3] = ls[62];
ls[63] = make_convolutional_layer(112, 1, 1, 0, 1, "relu"); // 61
ls[64] = make_shortcut_layer(ls[63], 1, "linear");
ls[65] = make_convolutional_layer(144, 1, 1, 0, 1, "relu");
ls[66] = make_convolutional_layer(288, 3, 1, 1, 1, "relu");
ls[67] = make_shortcut_layer(ls[63], 1, "linear");
ls[68] = make_convolutional_layer(32, 1, 1, 0, 1, "relu");
ls[69] = make_convolutional_layer(64, 5, 1, 2, 1, "relu");
ls[70] = make_shortcut_layer(ls[63], 1, "linear");
ls[71] = make_maxpool_layer(3, 1, 1);
ls[72] = make_convolutional_layer(64, 1, 1, 0, 1, "relu");
Layer **inception4d = malloc(4*sizeof(Layer*));
ls[73] = make_inception_layer(inception4d, 4, 2);
inception4d[0] = ls[64];
inception4d[1] = ls[67];
inception4d[2] = ls[70];
inception4d[3] = ls[73];
// 辅助分类器位置
ls[74] = make_convolutional_layer(256, 1, 1, 0, 1, "relu"); // 72
ls[75] = make_shortcut_layer(ls[74], 1, "linear");
ls[76] = make_convolutional_layer(160, 1, 1, 0, 1, "relu");
ls[77] = make_convolutional_layer(320, 3, 1, 1, 1, "relu");
ls[78] = make_shortcut_layer(ls[74], 1, "linear");
ls[79] = make_convolutional_layer(32, 1, 1, 0, 1, "relu");
ls[80] = make_convolutional_layer(128, 5, 1, 2, 1, "relu");
ls[81] = make_shortcut_layer(ls[74], 1, "linear");
ls[82] = make_maxpool_layer(3, 1, 1);
ls[83] = make_convolutional_layer(128, 1, 1, 0, 1, "relu");
Layer **inception4e = malloc(4*sizeof(Layer*));
ls[84] = make_inception_layer(inception4e, 4, 2);
inception4e[0] = ls[75];
inception4e[1] = ls[78];
inception4e[2] = ls[81];
inception4e[3] = ls[84];
ls[85] = make_maxpool_layer(2, 2, 0); // 83 25
ls[86] = make_convolutional_layer(256, 1, 1, 0, 1, "relu");
ls[87] = make_shortcut_layer(ls[86], 1, "linear");
ls[88] = make_convolutional_layer(160, 1, 1, 0, 1, "relu");
ls[89] = make_convolutional_layer(320, 3, 1, 1, 1, "relu");
ls[90] = make_shortcut_layer(ls[86], 1, "linear");
ls[91] = make_convolutional_layer(32, 1, 1, 0, 1, "relu");
ls[92] = make_convolutional_layer(128, 5, 1, 2, 1, "relu");
ls[93] = make_shortcut_layer(ls[86], 1, "linear");
ls[94] = make_maxpool_layer(3, 1, 1);
ls[95] = make_convolutional_layer(128, 1, 1, 0, 1, "relu");
Layer **inception5a = malloc(4*sizeof(Layer*));
ls[96] = make_inception_layer(inception5a, 4, 2);
inception5a[0] = ls[87];
inception5a[1] = ls[90];
inception5a[2] = ls[93];
inception5a[3] = ls[96];
ls[97] = make_convolutional_layer(384, 1, 1, 0, 1, "relu"); // 95
ls[98] = make_shortcut_layer(ls[97], 1, "linear");
ls[99] = make_convolutional_layer(192, 1, 1, 0, 1, "relu");
ls[100] = make_convolutional_layer(384, 3, 1, 1, 1, "relu");
ls[101] = make_shortcut_layer(ls[97], 1, "linear");
ls[102] = make_convolutional_layer(48, 1, 1, 0, 1, "relu");
ls[103] = make_convolutional_layer(128, 5, 1, 2, 1, "relu");
ls[104] = make_shortcut_layer(ls[97], 1, "linear");
ls[105] = make_maxpool_layer(3, 1, 1);
ls[106] = make_convolutional_layer(128, 1, 1, 0, 1, "relu");
Layer **inception5b = malloc(4*sizeof(Layer*));
ls[107] = make_inception_layer(inception5b, 4, 2);
inception5b[0] = ls[98];
inception5b[1] = ls[101];
inception5b[2] = ls[104];
inception5b[3] = ls[107];
ls[108] = make_avgpool_layer(7, 7, 0); // 106
ls[109] = make_dropout_layer(0.5);
ls[110] = make_connect_layer(100, 1, "linear");
ls[111] = make_crossentropy_layer(NULL, -1);
for (int i = 0; i < 112; ++i) {
append_layer2grpah(graph, ls[i]);
}
Session *sess = create_session(graph, 96, 96, 3, 100, type, path);
float *mean = calloc(3, sizeof(float));
float *std = calloc(3, sizeof(float));
mean[0] = 0.485;
mean[1] = 0.456;
mean[2] = 0.406;
std[0] = 0.229;
std[1] = 0.224;
std[2] = 0.225;
transform_normalize_sess(sess, mean, std);
transform_resize_sess(sess, 96, 96);
set_detect_params(sess);
init_session(sess, "./data/cifar100/train.txt", "./data/cifar100/train_label.txt");
detect_classification(sess);
}
在Lumos框架中demo目录下，您能找到googlenet.c文件，这就是我们已实现的googlenet模型
结果展示
该网络在经过150个epoch训练后，Top1分类精度在90%左右