VFlow: More Expressive Generative Flows with Variational Data Augmentation

Authors: Jianfei Chen, Cheng Lu, Biqi Chenli, Jun Zhu, Tian Tian

ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental For image density modeling on the CIFAR-10 dataset, VFlow achieves a new state-of-the-art 2.98 bits per dimension. We first evaluate VFlow on a toy DX = 2 Checkerboard dataset... We study the impact of the dimensionality of the flow DX + DZ {2, 4, 6, 8, 10}... We evaluate VFlow on CIFAR-10 and Image Net for density estimation of images.
Researcher Affiliation Collaboration 1Department of Computer Science and Technology, Institute for AI, BNRist Center, Tsinghua University 2Real AI.
Pseudocode No The paper does not contain any clearly labeled pseudocode or algorithm blocks.
Open Source Code Yes Our code is open-sourced at https://github.com/thu-ml/vflow.
Open Datasets Yes For image density modeling on the CIFAR-10 dataset... We evaluate VFlow on CIFAR-10 and Image Net (Russakovsky et al., 2015) for density estimation of images.
Dataset Splits Yes For this set of experiments, we randomly hold out 10,000 samples from the training set for validation.
Hardware Specification Yes All the experiments are run on 16 RTX 2080Ti GPUs.
Software Dependencies No The paper mentions using an "Adam optimizer (Kingma & Ba, 2015)" but does not provide specific version numbers for software libraries, frameworks, or programming languages used (e.g., Python, PyTorch, TensorFlow, CUDA versions).
Experiment Setup Yes The model is trained with an Adam optimizer (Kingma & Ba, 2015) with a batch size 64 for 2,000 epochs. Following (Ho et al., 2019), the learning rate linearly warms up to 0.0012 during the first 2,000 training steps, and exponentially decays at a rate of 0.99999 per step starting from the 50,000-th step until it reaches 0.0003.