Densely connected normalizing flows
Authors: Matej Grcić, Ivan Grubišić, Siniša Šegvić
NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments show significant improvements due to the proposed contributions and reveal state-of-the-art density estimation under moderate computing budgets. |
| Researcher Affiliation | Academia | Matej Grci c, Ivan Grubiši c and Siniša Šegvi c Faculty of Electrical Engineering and Computing University of Zagreb |
| Pseudocode | No | The paper does not contain any pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code available at: https://github.com/matejgrcic/Dense Flow |
| Open Datasets | Yes | We study the accuracy of density estimation on CIFAR-10 [35], Image Net [36] resized to 32 32 and 64 64 pixels and Celeb A [37]. |
| Dataset Splits | Yes | One MC sample is enough for accurate log-likelihood estimation since the per-example standard deviation is already about 0.01 bpd and a validation dataset size N additionally divides it by N. When compared with corresponding validation datasets, we achieve 37.1 on CIFAR10 and 38.5 on Image Net32. |
| Hardware Specification | Yes | Table 2: CIFAR-10 VFlow [24] 38M RTX 2080Ti 16 500 2.98 ... Dense Flow-74-10 130M RTX 3090 1 250 2.98. Image Net32 VFlow [24] 38M Tesla V100 16 1440 3.83. |
| Software Dependencies | No | The paper does not provide specific version numbers for software dependencies. |
| Experiment Setup | Yes | The first block of Dense Flow-74-10 uses 6 units with 5 glow-like modules in each Dense Flow unit, the second block uses 4 units with 6 modules, while the third block uses a single unit with 20 modules. We use the growth rate of 10 in all units. Each intra-module coupling starts with a projection to 48 channels. Subsequently, it includes a dense block with 7 densely connected layers, and the Nyström self-attention module with a single head. All models are trained on CIFAR-10 for 300 epochs and then fine-tuned for 10 epochs. We use the same training hyperparameters for all models. Training details are available in Appendix C. |