/CSI

CSI Project,DL net for paper,use python and tensorflow

Primary LanguagePython

CSI

Deep DeepCount

LSTM--->CNN-->FC-->Softmax

LSTM

Input dimension:time_360

one layer

Units number:N=64

Output dimension time_N_1

CNN

Two CNN blocks each block contains filter and max pooling components

first filter

  • cnn

input dimision:200_64

so the time is 200 ?

6 filters

kernel size:5_5

stride:1_1

  • max pool:

size:2_2 stride:2_2

output:98_30_6

second filter

  • cnn

10 filters

kernel size:5_3

stride:3*3

  • max pool

kernel size:1

stride:1

del this max pool?

out put:32_10_10

FC

Three layers

Input:32_10_10(flat to 3200*1)

  • 1000
  • 200
  • 5

output:5_1

Softmax

5 units

Data Processing

Algorithm

  • Amplitude Noise Removal

使用加权平均算法对振幅进行降噪,m设置为100

  • Phase Sanitization

首先对原始phase数据(180--->6*30) unwrap,然后计算出every subcarrier的均值y,利用y和x:[0:Sub-1]进行线性拟合(linear fit),最终算出calibrated phase value 并返回.

Code

  • DataProcess

使用weight moving等算法对原始数据进行处理,得到净数据

  • Normalize(已经在上一步进行了归一化处理)

对净数据进行归一化处理

经过以上两步处理得到fixed、open、semi三个文件夹下的数据文件夹,每个数据文件夹下的数据都是N×360的且已经做过归一化处理

Comparative Experiment

在两个方面做两组对比实验:

  • Only Amplitude(Without P)
  • Only Phase(Without A)
  • Without Amplitude noise removal but with Phase noise removal(Without A)
  • With Amplitude noise removal but without Phase noise removal(Without P)
  • Amplitude with noise removal and Phase with noise removal(With P*A)
  • Amplitude without noise removal and Phase without noise removal(Raw Data)

基于此,数据集应有如下几种:

  • Amplitude with noise removal

  • Phase with noise removal

  • Amplitude without noise removal (原始数据就是,但是需要将小数据集拼接成一个数据集)

  • Phase without noise removal (原始数据就是,但是需要将小数据集拼接成一个数据集)

使用以上四个数据集,组合成以下数据集进行训练:

  • Amplitude without noise removal&Phase with noise removal

  • Amplitude with noise removal&Phase without noise removal

  • Amplitude with noise removal&Phase with noise removal

  • Amplitude without noise removal&Phase without noise removal

另外使用另一个网络对如下数据集进行训练:

  • Amplitude with noise removal

  • Phase with noise removal

新网络的改进策略是讲原来的360统一换成180,切片长度和Units number不变,这样LSTM的输出维度就不变,这样CNN部分就不用修改。