ModernTCN: A Modern Pure Convolution Structure for General Time Series Analysis

Authors: Luo donghao, wang xue

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate Modern TCN on five mainstream analysis tasks, including long-term and short-term forecasting, imputation, classification and anomaly detection to verify the generality of Modern TCN.
Researcher Affiliation Academia Donghao Luo, Xue Wang Department of Precision Instrument, Tsinghua University, Beijing 100084, China
Pseudocode No No figures or sections explicitly labeled 'Pseudocode' or 'Algorithm' were found.
Open Source Code Yes Code is available at this repository: https://github.com/luodhhh/Modern TCN.
Open Datasets Yes We evaluate the long-term forecasting performance on 9 popular real-world datasets, including Weather, Traffic, Electricity, Exchange, ILI and 4 ETT datasets (ETTh1, ETTh2, ETTm1, ETTm2). And for imputation tasks, we choose Weather, Electricity and 4 ETT datasets (ETTh1, ETTh2, ETTm1, ETTm2) for benchmarking. These datasets have been extensively utilized for benchmarking and cover many aspects of life. The dataset size (total timesteps), variable number and sampling frequency of each dataset are summarized in Table 6 . We follow standard protocol (Zhou et al., 2021) and split all datasets into training, validation and test set in chronological order by the ratio of 6:2:2 for the ETT dataset and 7:1:2 for the other datasets.
Dataset Splits Yes We follow standard protocol (Zhou et al., 2021) and split all datasets into training, validation and test set in chronological order by the ratio of 6:2:2 for the ETT dataset and 7:1:2 for the other datasets.
Hardware Specification Yes All the deep learning networks are implemented in Py Torch(Paszke et al., 2019) and conducted on NVIDIA A100 40GB GPU.
Software Dependencies No The paper mentions 'Py Torch(Paszke et al., 2019)' and 'ADAM (Kingma & Ba, 2014)', but does not provide specific version numbers for these or other software dependencies.
Experiment Setup Yes Our method is trained with the L2 loss, using the ADAM (Kingma & Ba, 2014) optimizer with an initial learning rate of 10^-4. The default training process is 100 epochs with proper early stopping. The mean square error (MSE) and mean absolute error (MAE) are used as metrics. All the experiments are repeated 5 times with different seeds and the means of the metrics are reported as the final results.