Learning Dynamic Generator Model by Alternating Back-Propagation through Time
Authors: Jianwen Xie, Ruiqi Gao, Zilong Zheng, Song-Chun Zhu, Ying Nian Wu5498-5507
AAAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In our experiments, we show that we can learn the dynamic generator models using the ABPTT algorithm for dynamic textures and action sequences. |
| Researcher Affiliation | Collaboration | 1Hikvision Research Institute, Santa Clara, USA 2University of California, Los Angeles, USA |
| Pseudocode | Yes | Algorithm 1 Learning and inference by alternating backpropagation through time (ABPTT) |
| Open Source Code | Yes | The code and more results can be found at http://www.stat.ucla.edu/ jxie/Dynamic_Generator/Dynamic_Generator.html |
| Open Datasets | Yes | The video clips for training are collected from Dyn Tex++ dataset of (Ghanem and Ahuja 2010) and the Internet. We learn the model using the Weizmann action dataset (Gorelick et al. 2007). We compare with Mo Co GAN and TGAN (Saito, Matsumoto, and Saito 2017) by training on 9 selected categories (e.g., Playing Cello, Playing Daf, Playing Dhol, Playing Flute, Playing Guitar, Playing Piano, Playing Sitar, Playing Tabla, and Playing Violin) of videos in the UCF101 (Soomro, Zamir, and Shah 2012) database. We test our model on burning fire dataset (Xie, Zhu, and Wu 2017), and MUG Facial Expression dataset (N. Aifanti and Delopoulos 2010). |
| Dataset Splits | No | The paper describes its training process and parameters but does not explicitly provide details about training/validation/test dataset splits or reference predefined splits with citations for reproducibility. |
| Hardware Specification | Yes | We gratefully acknowledge the support of NVIDIA Corporation with the donation of the Titan Xp GPU used for this research. |
| Software Dependencies | No | The paper mentions 'Python with TensorFlow (Abadi and et al. 2015)' but does not specify version numbers for these software components. |
| Experiment Setup | Yes | We use the Adam (Kingma and Ba 2015) for optimization with β1 = 0.5 and the learning rate is 0.002. We set the Langevin step size to be δ = 0.03 for all latent variables, and the standard deviation of residual error σ = 1. We run l = 15 steps of Langevin dynamics for inference of the latent noise vectors within each learning iteration. In this experiment, the length of chunk is set to be 30 image frames. |