Improving Generative Moment Matching Networks with Distribution Partition
Authors: Yong Ren, Yucen Luo, Jun Zhu9403-9410
AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The experimental results show that GMMN-DP can generate complex samples on datasets such as Celeb A and CIFAR-10, where the vanilla GMMN fails. We present experimental results of the generative task on the commonly used datasets MNIST (Le Cun et al. 1998), Celeb A (Liu et al. 2015) and CIFAR-10. |
| Researcher Affiliation | Academia | Yong Ren,1 Yucen Luo, 1 Jun Zhu 1 1 Department of Computer Science and Technology, Institute for AI, BNRist Center, Tsinghua University reny11@foxmail.com, luoyc15@mails.tsinghua.edu.cn, dcszj@mail.tsinghua.edu.cn |
| Pseudocode | Yes | The algorithm is summarized in Alg. 1. Algorithm 1: Stochastic training for GMMN-DP |
| Open Source Code | Yes | The code can be found HERE2. 2https://github.com/McGrady00H/Improving-MMD-with-Distribution-Partition |
| Open Datasets | Yes | We present experimental results of the generative task on the commonly used datasets MNIST (Le Cun et al. 1998), Celeb A (Liu et al. 2015) and CIFAR-10. |
| Dataset Splits | No | The paper describes mini-batch training but does not specify explicit training, validation, and test dataset splits with percentages or counts for reproducibility. |
| Hardware Specification | Yes | In our experimental settings with a single RTX 2080ti GPU, the average time per iteration on CIFAR10 with B = 64 and model size 4, 300MB is 0.25s. |
| Software Dependencies | No | The paper mentions software components but does not provide specific version numbers for any of them. |
| Experiment Setup | Yes | Hyper-parameters: We use a mixture of 7 RBF kernels K(x, x ) = P7 i=1 Kσi(x, x ) with σi to be {1, 4, 8, 16, 24, 32, 64} for the sample space and K(y, y ) to be RBF kernel with σ = 1. The model is optimized using Adam with learning rate 0.001 and β = (0.9, 0.999). The batch size B is set to be 64 for all the datasets. The regularization parameter λ for b CX|Y is set to be 0.01. The dimension for the additional randomness z is set to be 2. |