https://github.com/ieee8023/latentshift

A method to generate counterfactuals
https://github.com/ieee8023/latentshift

Last synced: 5 months ago
JSON representation

A method to generate counterfactuals

• Host: GitHub
• URL: https://github.com/ieee8023/latentshift
• Owner: ieee8023
• Created: 2022-09-05T23:46:06.000Z (over 2 years ago)
• Default Branch: main
• Last Pushed: 2024-04-25T06:30:48.000Z (12 months ago)
• Last Synced: 2024-11-07T03:08:00.903Z (6 months ago)
• Language: Jupyter Notebook
• Homepage:
• Size: 14.9 MB
• Stars: 11
• Watchers: 1
• Forks: 1
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • Citation: CITATION

Awesome Lists containing this project

README

        
# Latent Shift - A Simple Autoencoder Approach to Counterfactual Generation

[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/ieee8023/latentshift/blob/main/example.ipynb)

# The idea

Read the paper about Latent Shift: https://arxiv.org/abs/2102.09475

Watch a video: https://www.youtube.com/watch?v=1fxSDP8DheI

Read the paper about Counterfactual Alignment: https://arxiv.org/abs/2312.02186

The main diagram:

![latentshift.gif](docs/latentshift.gif)

## Animations/GIFs

| Smiling | Arched Eyebrows|

| ----------- | ----------- |

|  |   | 

|Mouth Slightly Open  | Young  |

| ----------- | ----------- |

|  |   | 

# Generating a transition sequence

For a predicting of `smiling`

![gen_sequence.png](docs/gen_sequence.png)

# Multiple different targets



  

# Comparison to traditional methods 

For a predicting of `pointy_nose`

![comparison.png](docs/comparison.png)

# Getting Started

```bash

$pip install latentshift

````

```python3

import latentshift

# Load classifier and autoencoder

model = latentshift.classifiers.FaceAttribute(download=True)

ae = latentshift.autoencoders.VQGAN(weights="faceshq", download=True)

# Load image

input = torch.randn(1, 3, 1024, 1024)

# Defining Latent Shift module

attr = captum.attr.LatentShift(model, ae)

# Computes counterfactual for class 3.

output = attr.attribute(input, target=3)

```