In this repo, we introduce it from four parts : first to introduce the dataset and how to download dataset, second to discuss the environments you need, third to disscuss the sturcture of this paper and how to train, and finally to introduce our experiments.
In our experiments, we use the UCR time series achive to evaluate our model, which it includes 128 vary datasets and its details are shown at http://www.timeseriesclassification.com . Here you can download dataset from the official websites or my baidu Cloud, which its key is 8u8u. There is a sample picture in the standard dataset:
The codes of this repo are tested on Python 3.6 and Tensorflow 1.13. First you must install some API your computers need. You can do as follows.
pip3 install --user -r request.txt or sudo pip3 install -r request.txt
python2: you can change pip3 into pip .
First we introduce the structure of our codes:
--- Execute Model:
-----main.py : to Execute the primary examples of our proposed approach
---------TSCaps Dir: data.py : to normalize the raw datasets
---------TSCaps Dir: config.py : configure the preference of our model
---------TSCaps Dir: TSPCap.py : the details of our proposed model
--- Our Strored Model:
------Examples Dir : there are more than 43 files. If you want to check the accuracy of model, you first download the stored parameters and add it into the respective files.
There are details of each dataset which was trained by our model.
--- Comparison Classic Model:
------ In the Compared Models Dir: You will find the InceptionNet , resnet, lstm+resnet, MLP ,TimeSeriesCNNs etc.
----------The model details are shown in ./Compared Models/classicnet.py
python3 main.py
#if your envoirnments are python2, you must change some codes of training example.
Second we discuss our proposed approach architectures:
Our Proposed Approach includes four parts, which its details are shown as follows :
Use our decode layer to rebuild original features. There are some samples in the ECG5000 dataset of UCR Time Series Achive.
In this paper, we chose 43 datasets from the UCR time series achives.
Examples dir: there are some detailed training files. You can Download the keypointModel and add its path into the respectively file. The results are same of above picture.
As the some datasets in the UCR are very fragile, we therefore use the lr = 0.0001 to learn initially. When we find some local minimum of the training model, we must choose the lr*0.1.
When we do as above methods, we therefore need much more time to train and more training epoch.
Compared with other existing algorithms, you can
The Draw.py files can draw above two pictures. Thanks for supporting this tool.
You can add your dataset into the example.csv or reasult.csv if you need this picture.
Now Under Review in the IEEE Transaction on nerual networks and learning systems.
@article{IEEE Transaction on nerual networks and learning systems,
title={TSCaps:A Novel Multi-Capsule Neural Architecture for Time Series Classification},
author={Zhiwen Xiao, Xin Xu, Haoxi Zhang,and Edward Szczerbicki},
journal={Uner Review at IEEE Transaction on nerual networks and learning systems},
year={2020}
}