Glow: Generative Flow with Invertible 1x1 Convolutions

Authors: Durk P. Kingma, Prafulla Dhariwal

NeurIPS 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this paper we propose Glow, a simple type of generative flow using an invertible 1 1 convolution. Using our method we demonstrate a significant improvement in log-likelihood on standard benchmarks. Perhaps most strikingly, we demonstrate that a flow-based generative model optimized towards the plain log-likelihood objective is capable of efficient realistic-looking synthesis and manipulation of large images. ... 5 Quantitative Experiments ... 6 Qualitative Experiments
Researcher Affiliation Collaboration Diederik P. Kingma* , Prafulla Dhariwal *Open AI Google AI
Pseudocode No The paper describes its proposed generative flow and its components in detail using text and mathematical equations, but it does not include any structured pseudocode or algorithm blocks.
Open Source Code Yes The code for our model is available at https://github.com/openai/glow.
Open Datasets Yes We compare the average negative log-likelihood (bits per dimension) on the CIFAR-10 (Krizhevsky, 2009) dataset... In particular, we compare on CIFAR-10, Image Net (Russakovsky et al., 2015) and LSUN (Yu et al., 2015) datasets. ... We choose the Celeb A-HQ dataset (Karras et al., 2017)
Dataset Splits Yes Since the original Celeb A HQ dataset didn’t have a validation set, we separated it into a training set of 27000 images and a validation set of 3000 images.
Hardware Specification Yes More specifically, generating a 256 256 image at batch size 1 takes about 130ms on a single NVIDIA GTX 1080 Ti, and about 550ms on a NVIDIA Tesla K80.
Software Dependencies No The paper does not provide specific version numbers for software components such as programming languages, libraries (e.g., PyTorch), or frameworks used for the experiments.
Experiment Setup Yes In our experiments, we let each NN() have three convolutional layers, where the two hidden layers have Re LU activation functions and 512 channels. The first and last convolutions are 3 3, while the center convolution is 1 1... All models were trained with K = 32 and L = 3. ... K = 32 and L = 6. ... minibatch size 1 per PU, and use gradient checkpointing (Salimans and Bulatov, 2017).