DTWNet: a Dynamic Time Warping Network
Authors: Xingyu Cai, Tingyang Xu, Jinfeng Yi, Junzhou Huang, Sanguthevar Rajasekaran
NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this experimental section, we compare the proposed scheme with existing approaches. We refer to the end-to-end DTW kernel as Full DTW, and the streaming version as SPRING DTW. We implement our approach in Py Torch [16]. Figure 4c and Figure 4d show the test accuracy and test loss for 400 iterations. The UCR repository [3] is used in this experiment. We evaluate our method against Soft DTW [4], DBA [17] and SSG [4]. We report the average of 5 runs for each experiment. The experimental results are summarized in Table 1. |
| Researcher Affiliation | Collaboration | Xingyu Cai University of Connecticut Tingyang Xu Tencent AI Lab Jinfeng Yi JD.com AI Lab Junzhou Huang Tencent AI Lab Sanguthevar Rajasekaran University of Connecticut |
| Pseudocode | Yes | A complete example of DTWNet on a classification task is illustrated in Algorithm 1. Algorithm 1 DTWNet training for a classification task. |
| Open Source Code | No | The paper does not provide concrete access to source code for the methodology described, nor does it contain explicit statements about code availability or links to a repository. |
| Open Datasets | Yes | The UCR repository [3] is used in this experiment. We use the Haptics dataset from the UCR repository to demonstrate the decomposition. Yanping Chen, Eamonn Keogh, Bing Hu, Nurjahan Begum, Anthony Bagnall, Abdullah Mueen, and Gustavo Batista. The ucr time series classification archive, July 2015. www.cs.ucr.edu/ ~eamonn/time_series_data/. |
| Dataset Splits | Yes | There are a total of 100 sequences in the training set, 50 in each category. Another 100 sequences form the testing set, 50 for each type as well. We use the entire training set to train the model to obtain the barycenter bi for each category, and then calculate the DTW loss as: Ldtw = 1 Nclass j=0 dtw(si,j, bi) where Nclass is the number of categories, Ni is the number of sequences in class i, and si,j is sequence j in class i. The DTW distance is computed using ℓ2 norm. Clearly, the less the loss, the better is the performance. We also evaluate on the testing set by using si,j from the testing set. |
| Hardware Specification | No | The paper does not provide any specific hardware details such as GPU or CPU models, processor types, or memory used for running the experiments. |
| Software Dependencies | No | The paper states 'We implement our approach in Py Torch [16]', but it does not specify a version number for PyTorch or any other software dependency. |
| Experiment Setup | Yes | The kernel lengths are set to 10. α = 0.1 for SPRING DTW. We append 3 linear layers to generate the prediction. Figure 4c and Figure 4d show the test accuracy and test loss for 400 iterations. Note that we first run Soft DTW with 4 different hyperparameter settings γ = 1, 0.1, 0.01 and 0.001 as in [4]. |