2.2 网络模型
Caffe训练网络的网络结果如下所示：
2.3相关术语及变量
N为LSTM同时处理的独立流的个数，在该实验中为输入LSTM相互独立的视频的个数，以该实验测试网络为例，本文取T=3。
T为LSTM网络层处理的时间步总数，在该实验中为输入LSTM的任意一独立视频的视频帧个数，以该实验测试网络为例，本文取T=16。
因此fc-reshape层输出维度为 T×N×4096T. 4096为AlexNet中全连接层的维度，即CNN特征的维度。
reshape-cm的输出维度为 T×N，即每一个帧有一个是否连续帧的标志。(第一帧为0，后面的连续帧为1)
reshape-label的维度同样为 T×N
name: "LeNet"
layer {
name: "mnist"
type: "Data"
top: "data"
top: "label"
include {
phase: TRAIN
}
transform_param {
scale: 0.00390625
}
data_param {
source: "lmdb/mnist_train_lmdb"
batch_size: 1
backend: LMDB
}
}
layer {
name: "mnist"
type: "Data"
top: "data"
top: "label"
include {
phase: TEST
}
transform_param {
scale: 0.00390625
}
data_param {
source: "lmdb/mnist_test_lmdb"
batch_size: 1
backend: LMDB
}
}
# 将 data 转换为 [T,Batch, H]
layer {
name: "reshape"
type: "Reshape"
bottom: "data"
top: "reshape1"
reshape_param {
shape { dim:28 dim:1 dim:28}
}
}
layer {
name: "cont"
type: "HDF5Data"
top: "cont"
include {
phase: TRAIN
}
hdf5_data_param {
source: "train_conv_h5.txt"
batch_size: 1
}
}
layer {
name: "cont"
type: "HDF5Data"
top: "cont"
include {
phase: TEST
}
hdf5_data_param {
source: "test_conv_h5.txt"
batch_size: 1
}
}
#将 cont 转换为 [T, batch]
layer {
name: "reshape"
type: "Reshape"
bottom: "cont"
top: "cont1"
reshape_param {
shape { dim:28 dim:1}
}
}
#num_output 代表每个 lstm的输出 为 10 个数值 ，只使用第28个时间步的值
layer {
name: "lstm1"
type: "LSTM"
bottom: "reshape1"
bottom: "cont1"
top: "lstm1"
recurrent_param {
num_output: 10
weight_filler {
type: "uniform"
min: -0.08
max: 0.08
}
bias_filler {
type: "constant"
value: 0
}
}
}
#输出为28个时间步的 output [t, batch, 10] 只使用第28个时间步的值 output2
layer {
name: "slice"
type: "Slice"
bottom: "lstm1"
top: "output1"
top: "output2"
slice_param {
axis: 0
slice_point: 27
}
}
#output1 一直输出影响观察，这里随便加的一层，没有作用
layer {
name: "reduction"
type: "Reduction"
bottom: "output1"
top: "output11"
reduction_param {
axis: 0
}
}
layer {
name: "reshape"
type: "Reshape"
bottom: "output2"
top: "output22"
reshape_param {
shape { dim:1, dim:-1}
}
}
layer {
name: "accuracy"
type: "Accuracy"
bottom: "output22"
bottom: "label"
top: "accuracy"
include {
phase: TEST
}
}
layer {
name: "loss"
type: "SoftmaxWithLoss"
bottom: "output22"
bottom: "label"
top: "loss"
}
cont 数据生成
Mat data_train1(64, 28, CV_8UC1), data_test1(64, 28, CV_8UC1);
for (size_t i = 0; i < 64; i++)
{
for (size_t j = 0; j < 28; j++)
{
if (j == 0)
{
data_train1.at<uchar>(i, j) = 0;
data_test1.at<uchar>(i, j) = 0;
}
else {
data_train1.at<uchar>(i, j) = 1;
data_test1.at<uchar>(i, j) = 1;
}
}
}
mat2hdf5(data_train1, "D:\\MNIST\\data\\train_cont.h5", "cont");
mat2hdf5(data_test1, "D:\\MNIST\\data\\test_cont.h5", "cont");
system("PAUSE");
void mat2hdf5(Mat &data, const char * filepath, string dataset1)
{
int data_cols = data.cols;
int data_rows = data.rows;
hid_t file_id;
herr_t status;
file_id = H5Fcreate(filepath, H5F_ACC_TRUNC, H5P_DEFAULT, H5P_DEFAULT);
int rank_data = 2;
hsize_t dims_data[2];
dims_data[0] = data_rows;
dims_data[1] = data_cols;
hid_t data_id = H5Screate_simple(rank_data, dims_data, NULL);
hid_t dataset_id = H5Dcreate2(file_id, dataset1.c_str(), H5T_NATIVE_FLOAT, data_id, H5P_DEFAULT, H5P_DEFAULT, H5P_DEFAULT);
int i, j;
float* data_mem = new float[data_rows*data_cols];
float **array_data = new float*[data_rows];
for (j = 0; j < data_rows; j++) {
array_data[j] = data_mem + j* data_cols;
for (i = 0; i < data_cols; i++)
{
array_data[j][i] = data.at<uchar>(j, i);
}
}
status = H5Dwrite(dataset_id, H5T_NATIVE_FLOAT, H5S_ALL, H5S_ALL, H5P_DEFAULT, array_data[0]);
//关闭
status = H5Sclose(data_id);
status = H5Dclose(dataset_id);
status = H5Fclose(file_id);
delete[] array_data;
}
test_cont 与 train_cont 数据一样，64*28，其中64 是可变的。
由于 输入数据 要进行 维度转换 由[batch, 1, w, h] 这里把 w 看成时间部 T
应该转换成 [t, batch, h] ，在caffe 里面没有找到维度变换的方法，所以直接使用的reshape ,为了保证有效，batch 必须设置成1，否者与实际转换不符。
训练结果
I0715 17:11:19.591768 10532 solver.cpp:348] Iteration 790000, Testing net (#0)
I0715 17:11:19.600821 10532 solver.cpp:415] Test net output #0: accuracy = 0.83
I0715 17:11:19.600821 10532 solver.cpp:415] Test net output #1: loss = 1.18451 (* 1 = 1.18451 loss)
I0715 17:11:19.600821 10532 solver.cpp:415] Test net output #2: output11 = -35.0393
I0715 17:11:19.600821 10532 solver.cpp:220] Iteration 790000 (6993.01 iter/s, 0.715s/5000 iters), loss = 1.00843
I0715 17:11:19.600821 10532 solver.cpp:239] Train net output #0: loss = 1.00847 (* 1 = 1.00847 loss)
I0715 17:11:19.600821 10532 solver.cpp:239] Train net output #1: output11 = -43.1359
I0715 17:11:19.600821 10532 sgd_solver.cpp:105] Iteration 790000, lr = 0.000373837
I0715 17:11:20.325745 10532 solver.cpp:348] Iteration 795000, Testing net (#0)
I0715 17:11:20.335746 10532 solver.cpp:415] Test net output #0: accuracy = 0.77
I0715 17:11:20.335746 10532 solver.cpp:415] Test net output #1: loss = 1.24829 (* 1 = 1.24829 loss)
I0715 17:11:20.335746 10532 solver.cpp:415] Test net output #2: output11 = -37.8989
I0715 17:11:20.335746 10532 solver.cpp:220] Iteration 795000 (6811.99 iter/s, 0.734s/5000 iters), loss = 0.868423
I0715 17:11:20.335746 10532 solver.cpp:239] Train net output #0: loss = 0.868455 (* 1 = 0.868455 loss)
I0715 17:11:20.335746 10532 solver.cpp:239] Train net output #1: output11 = -54.416
I0715 17:11:20.335746 10532 sgd_solver.cpp:105] Iteration 795000, lr = 0.000372094
I0715 17:11:21.056661 10532 solver.cpp:468] Snapshotting to binary proto file save_path/_iter_800000.caffemodel
I0715 17:11:21.058668 10532 sgd_solver.cpp:273] Snapshotting solver state to binary proto file save_path/_iter_800000.solverstate
I0715 17:11:21.059669 10532 solver.cpp:328] Iteration 800000, loss = 1.01535
I0715 17:11:21.059669 10532 solver.cpp:348] Iteration 800000, Testing net (#0)
I0715 17:11:21.068692 10532 solver.cpp:415] Test net output #0: accuracy = 0.75
I0715 17:11:21.068692 10532 solver.cpp:415] Test net output #1: loss = 1.32086 (* 1 = 1.32086 loss)
I0715 17:11:21.068692 10532 solver.cpp:415] Test net output #2: output11 = -33.7201
I0715 17:11:21.068692 10532 solver.cpp:333] Optimization Done.
I0715 17:11:21.068692 10532 caffe.cpp:260] Optimization Done.