2022.11.18
제네시스랩 신재영
NeurIPS 2018
Contents
Glow: Generative Flow with Invertible 1×1 Convolutions
Introduction
Types of
generative models
Motivation
01
Background
Change of
variable theorem
Normalizing Flows
RealNVP
02
Architecture
03
Experiments
04
Overview
Components
Qualitative Experiment
Quantitative Experiment
Types of Generative Models
Introduction
Motivation
Introduction
Introduction
Types of Generative Models
Background
Change of Variable Theorem
Multi variable version:
Jacobian Matrix
Background
Change of Variable Theorem
Determinant
Background
Normalizing Flows
According to the inverse func theorem
According to a property of Jacobians of invertible func
Background
Normalizing Flows
Background
Normalizing Flows
Background
RealNVP
Condition 1: “It is easily invertible.”
Condition 2: “Its Jacobian determinant is easy to compute .”
Some dimensions remain unchanged
To make sure all the inputs have a chance to be altered, the model reverses the ordering in each layer
det(J) =
Background
RealNVP
Between layers of the RealNVP flow, the ordering of channels is reversed so that all the data dimensions have a chance to be altered.
A 1×1 convolution with equal number of input and output channels is a generalization of any permutation of the channel ordering.
Architecture
Overview
Components
Architecture
Invertible 1x1 convolution
Architecture
Actnorm
Architecture
Architecture
Overall Architecture
squeeze
Qualitative Experiment
Experiments
Qualitative Experiment
Experiments
Quantitative Experiment
Experiments
The affine coupling models converge faster than the additive coupling models.
The increase in wallclock time for the invertible 1 × 1 convolution model was only ≈ 7%,
thus the operation is computationally efficient as well.
THANK YOU FOR LISTENING
Introduction
Flow based generative models
Motivation
Introduction
Introduction
Flow-based generative models have so far gained little attention in the research community compared to GANs (Goodfellow et al., 2014) and VAEs (Kingma and Welling, 2013). Some of the merits of flow-based generative models include:
• Exact latent-variable inference and log-likelihood evaluation
• Efficient inference and efficient synthesis
• Useful latent space for downstream tasks
• Significant potential for memory savings
Background
Change of Variable Theorem
Between layers of the RealNVP flow, the ordering of channels is reversed so that all the data dimensions have a chance to be altered.
A 1×1 convolution with equal number of input and output channels is a generalization of any permutation of the channel ordering.
Architecture
Overview