On the Quantitative Analysis of Decoder-Based Generative Models
Authors: Yuhuai Wu, Yuri Burda, Ruslan Salakhutdinov, Roger Grosse
ICLR 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | All of our experiments were performed on the MNIST dataset of images of handwritten digits (Le Cun et al., 1998). We evaluated the trained models using AIS and KDE on 1000 test examples of MNIST; results are shown in Table 2. |
| Researcher Affiliation | Collaboration | Yuhuai Wu Department of Computer Science University of Toronto ywu@cs.toronto.edu Yuri Burda Open AI yburda@openai.com Ruslan Salakhutdinov School of Computer Science Carnegie Mellon University rsalakhu@cs.cmu.edu Roger Grosse Department of Computer Science University of Toronto rgrosse@cs.cmu.edu |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | The evaluation code is provided at https://github.com/tonywu95/eval_gen. |
| Open Datasets | Yes | All of our experiments were performed on the MNIST dataset of images of handwritten digits (Le Cun et al., 1998). |
| Dataset Splits | Yes | We use the standard split of MNIST into 60,000 training and 10,000 test examples, and used 50,000 images from the training set for training, and remaining 10,000 images for validation. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for running experiments (e.g., GPU/CPU models, memory). |
| Software Dependencies | No | The paper mentions software like Lasagne and Theano with citations but does not provide specific version numbers for these or other software dependencies used in the experiments. |
| Experiment Setup | Yes | For GAN-10, we used a discriminator with the architecture 784-512-256-1, where each layer used dropout with parameter 0.5. For GAN-50, we used a discriminator with architecture 784-4096-4096-4096-4096-1. All hidden layers used dropout with parameter 0.8. All hidden layers in both networks used the tanh activation function, and the output layers used the logistic function. For training GAN/VAE, we use our own implementation. We use Adam for optimization, and perform grid search of learning rate from {0.001, 0.0001, 0.00001}. For training GMMN, we take the implementation from https://github.com/yujiali/gmmn.git. Following the implementation, we use SGD with momentum for optimization, and perform grid search of learning rate from {0.1, 0.5, 1, 2}, with momentum 0.9. For all experiments in this section, we used 10,000 intermediate distributions for AIS, 1 million simulated samples for KDE, and 200,000 importance samples for the IWAE bound. |