https://github.com/bv7dev/vae-lab

Experimenting with simple VAEs for image generation
https://github.com/bv7dev/vae-lab

Last synced: about 1 year ago
JSON representation

Experimenting with simple VAEs for image generation

• Host: GitHub
• URL: https://github.com/bv7dev/vae-lab
• Owner: bv7dev
• License: mit
• Created: 2024-12-15T11:53:48.000Z (over 1 year ago)
• Default Branch: main
• Last Pushed: 2025-01-06T12:12:32.000Z (over 1 year ago)
• Last Synced: 2025-02-13T18:46:12.571Z (over 1 year ago)
• Language: Jupyter Notebook
• Size: 23.4 MB
• Stars: 0
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

          
# vae-lab

Experiments with simple Variational-Autoencoder (VAE) in PyTorch

1. create a Python venv

2. install PyTorch

3. explore scripts & notebook

```python

# Explore the latent space

with torch.no_grad():

    dd = d[2] # desired direction

    iter = torch.arange(10, device=device)

    x = t + 50*iter[:, None]*dd[None, :]

    path = vae.decode(x)

    util.plot_image_grid(1, 10, path.cpu())

```

![continuous interpolation](public/continuous_interpolation.png)

```python

# Decode random latent vectors

with torch.no_grad():

    z = torch.randn(10, LATENT_SIZE).to(device)

    recon = vae.decode(z*3)

    util.plot_image_grid(2, 5, recon.cpu())

```

![random latent sampling example](public/random_latent_sampling.png)

#### ideas

- try a round gaussian mask for original and reconstruction to give the center

  of the image, the face, most weight for mse loss (fade into black, transparency?)

![target of some experiment i dont know anymore :D](target.png)