PSA-GAN: Progressive Self Attention GANs for Synthetic Time Series
Authors: Paul Jeha, Michael Bohlke-Schneider, Pedro Mercado, Shubham Kapoor, Rajbir Singh Nirwan, Valentin Flunkert, Jan Gasthaus, Tim Januschowski
ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In Section 4, we evaluate our proposed GAN model using the proposed Context-FID score and through several downstream forecasting tasks. We also directly evaluate our model as a forecasting algorithm and perform an ablation study. |
| Researcher Affiliation | Collaboration | Paul Jeha , Technical University of Denmark pauje@dtu.dk Michael Bohlke-Schneider AWS AI Labs bohlkem@amazon.com Pedro Mercado AWS AI Labs pedroml@amazon.com Shubham Kapoor AWS AI Labs kapooshu@amazon.com Rajbir Singh Nirwan AWS AI Labs nirwar@amazon.com Valentin Flunkert AWS AI Labs flunkert@amazon.com Jan Gasthaus AWS AI Labs gasthaus@amazon.com Tim Januschowski Zalando SE tim.januschowski@zalando.de |
| Pseudocode | No | The paper describes the model architecture and training procedure in textual form and through diagrams (e.g., Figure 1, Section 3), but it does not include formal pseudocode or algorithm blocks. |
| Open Source Code | Yes | The code we used in the paper is available under: https://github.com/mbohlkeschneider/psa-gan and we will additionally disseminate PSA-GAN via Gluon TS: https://github.com/awslabs/gluon-ts. |
| Open Datasets | Yes | We use the following public, standard benchmark datasets in the time series domain: M4, hourly time series competition data (414 time series) (Makridakis et al., 2020); Solar, hourly solar energy collection data in Alabama State (137 stations) (Lai et al., 2018); Electricity, hourly electricity consumption data (370 customers) (Dheeru & Karra Taniskidou, 2017); Traffic: hourly occupancy rate of lanes in San Francisco (963 lanes) (Dheeru & Karra Taniskidou, 2017). |
| Dataset Splits | No | Unless stated otherwise, we split all data into a training/test set with a fixed date and use all data before that date for training. For testing, we use a rolling window evaluation with a window size of 32 and seven windows. |
| Hardware Specification | Yes | PSA-GAN (and its variants) has been trained on ml.p2.xlarge Amazon instances for 21 hours. |
| Software Dependencies | No | All GAN models in this paper have been implemented using Py Torch from Paszke et al. (2019). We use the Deep AR (Salinas et al., 2020) implementation in Gluon TS (Alexandrov et al., 2020) for the forecasting experiments. The paper mentions software such as PyTorch and Gluon TS but does not specify their version numbers. |
| Experiment Setup | Yes | PSA-GAN has been trained over 6500 epochs, where a new block is added every 1000 epochs and is faded over 500 epochs. At each epoch, PSA-GAN trains on 100 batches of size 512. It optimises its parameters using the Adam (Kingma & Ba, 2014) with a learning rate of 0.0005 and betas of (0.9, 0.999) for both the generator and the discriminator. We use the default hyperparameters of Deep AR with the following exceptions: epochs=100, num batches per epoch=100, dropout rate=0.01, scaling=False, prediction length=32, context length=64, use feat static cat=True. We use the Adam optimizer with a learning rate of 1e 3 and weight decay of 1e 8 . Additionally, we clip the gradient to 10.0 and reduce the learning rate by a factor of 2 for every 10 consecutive updates without improvement, up to a minimum learning rate of 5e 5. |