Self-Attention ConvLSTM for Spatiotemporal Prediction
Authors: Zhihui Lin, Maomao Li, Zhuobin Zheng, Yangyang Cheng, Chun Yuan11531-11538
AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In experiments, we apply the SA-Conv LSTM to perform frame prediction on the Moving MNIST and KTH datasets and traffic flow prediction on the Texi BJ dataset. |
| Researcher Affiliation | Collaboration | 1Department of Computer Science and Technologies, Tsinghua University, Beijing, China 2Graduate School at Shenzhen, Tsinghua University, Shenzhen, China 3Peng Cheng Laboratory, Shenzhen, China |
| Pseudocode | No | The paper describes mathematical formulations and processes but does not provide a clearly labeled pseudocode or algorithm block. |
| Open Source Code | No | The paper does not provide an explicit statement about releasing source code or a link to a code repository for the described methodology. |
| Open Datasets | Yes | Moving MNIST is a commonly used dataset contains a variety of sequences generated by the method mentioned in (Srivastava et al. 2015)... Taxi BJ is collected from the chaotic real-world environment... KTH (Schuldt et al. 2004) contains 6 categories of human actions... |
| Dataset Splits | No | The paper mentions 'The scheduled sampling strategy (Bengio et al. 2015) and Layer Norm (Ba, Kiros, and Hinton 2016) are also adopted in the training process.' and describes input/prediction frames but does not provide explicit training, validation, and test dataset split percentages or counts for reproduction. |
| Hardware Specification | Yes | We also evaluated the efficiency of each model based on a GTX 1080TI GPU and the Tensor Flow framework. |
| Software Dependencies | No | The paper mentions 'Tensor Flow framework' but does not specify its version number or any other software dependencies with specific version numbers. |
| Experiment Setup | Yes | a 4-layer architecture with 64 hidden states in each layer for every model. The scheduled sampling strategy (Bengio et al. 2015) and Layer Norm (Ba, Kiros, and Hinton 2016) are also adopted in the training process. Each model is trained with an ADAM optimizer and a beginning learning rate of 0.001. During training, the mini-batch is set to 8, and the training process is stopped after 80,000 iterations. We use L2 loss for the Moving MNIST and the Taxi BJ datasets, while L1 + L2 loss for the KTH dataset. |