MintNet: Building Invertible Neural Networks with Masked Convolutions
Authors: Yang Song, Chenlin Meng, Stefano Ermon
NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | To demonstrate their flexibility, we show that our invertible neural networks are competitive with Res Nets on MNIST and CIFAR-10 classification. When trained as generative models, our invertible networks achieve competitive likelihoods on MNIST, CIFAR-10 and Image Net 32 32, with bits per dimension of 0.98, 3.32 and 4.06 respectively. 5 Experiments In this section, we evaluate our Mint Net architectures on both image classification and density estimation. We focus on three common image datasets, namely MNIST, CIFAR-10 and Image Net 32 32. |
| Researcher Affiliation | Academia | Yang Song Stanford University yangsong@cs.stanford.edu Chenlin Meng Stanford University chenlin@cs.stanford.edu Stefano Ermon Stanford University ermon@cs.stanford.edu |
| Pseudocode | Yes | Algorithm 1 Fixed-point iteration method for computing f 1(z). |
| Open Source Code | No | The paper does not provide an explicit statement or link to open-source code for the described methodology. |
| Open Datasets | Yes | We focus on three common image datasets, namely MNIST, CIFAR-10 and Image Net 32 32. |
| Dataset Splits | No | The paper mentions 'test accuracy' and 'training accuracy' for evaluation, but does not provide specific details on training, validation, or test dataset splits (e.g., percentages, sample counts, or explicit references to predefined splits). |
| Hardware Specification | No | When training on dataset such as CIFAR-10, Mint Net required 2 GPUs for approximately 5 days, while FFJORD [9] used 6 GPUs for approximately 5 days, and Glow [16] used 8 GPUs for approximately 7 days. However, specific GPU models or other hardware details are not provided. |
| Software Dependencies | No | The paper does not provide specific software dependency details, such as library names with version numbers, needed to replicate the experiment. |
| Experiment Setup | Yes | We used grid search to select the step size α in Algorithm 1 and chose α = 3.5, 1.1, 1.15 respectively for MNIST, CIFAR-10 and Image Net 32 32. We also empirically verify that Algorithm 1 can provide accurate solutions within a small number of iterations. We provide more details about settings and model architectures in Appendix D. |