Hugging Face's logo

thewhatifproject
/
gaIA

Unconditional Image Generation
PyTorch
Italian
gan
nvidia
stylegan
stylegan3
Model card Files Community
gaIA / README.md
thewhatifproject's picture
thewhatifproject
Update README.md
b5413db verified
preview code
|
raw
history blame contribute delete
5.43 kB
metadata
library_name: pytorch
license: apache-2.0
tags:
  - nvidia
  - gan
  - stylegan
  - stylegan3
  - nvidia
pipeline_tag: unconditional-image-generation
language:
  - it
extra_gated_prompt: >-
  You agree to not use the model to conduct experiments that cause harm to human
  subjects. You agree to cite this model for every usage using its DOI.
extra_gated_fields:
  Company: text
  Country: country
  I want to use this model for:
    type: select
    options:
      - Research
      - Education
      - Art & Exhibitions
      - label: Other
        value: other
  I agree to use this model for non-commercial use ONLY: checkbox
extra_gated_heading: Acknowledge license and conditions to accept the repository
extra_gated_description: Our team may take 1-2 days to process your request
extra_gated_button_content: I accept

gaIA: Italian Landscape GAN Model

gaIA is the first Italian GAN model trained on satellite images of a selection of Italy's main glaciers, forests, lakes, rivers, and coasts that are most affected by climate change. It is usable for scientific and artistic purposes.

Intro

Dataset

  • Images: 12k
  • Image Format: 1024x1024
  • Source: Copernicus Sentinel 2A
  • Reference Years: 2017 – June 2024

Dataset

  • 29 Covered Areas:
    • Glaciers: Adamello, Gran Paradiso, Marmolada, Presena, Forni, Belvedere
    • Lakes: Bracciano, Garda, Maggiore, Trasimeno, Iseo, Como
    • Rivers: Tiber, Adige, Arno, etc.
    • Islands/Coasts: Chia, Marina di Pisa, Venezia, Stromboli, Rosolina Mare, Costiera Amalfitana
    • Parks: Abruzzo, Casentinesi, Pollino, Sila, Gargano, Aspromonte

Dataset Location

Training

  • Framework: StyleGAN3-T
  • GPUs: 1 - NVIDIA A100 80GB
  • Batch: 32
  • Gamma: 32
  • Kimg: 5152.0
  • Augmentations: 38,040
  • Time: ~220 hours

Training

Requirements

Please refer to Official NVIDIA Paper Requirements

How to Start 

import torch
from PIL import Image
import numpy as np
import pickle

# Set the device to GPU
device = torch.device('cuda')

# Load the model
with open('/thewhatifproject/gaIA_v1.pkl', 'rb') as f:
    G = pickle.load(f)['G_ema'].cuda()  # torch.nn.Module

# Set the model to evaluation mode
G.eval()

# Set the seed for reproducibility
seed = 28

# Generate latent codes using the specified seed
z = torch.from_numpy(np.random.RandomState(seed).randn(1, G.z_dim)).to(device)

# Generate the image using the generator
with torch.no_grad():
    img = G(z, None, truncation_psi=1, noise_mode='const')

# Process the image for saving
# - Change dimensions order from NCHW to NHWC
# - Scale from range [-1, +1] to [0, 255]
# - Clamp values to ensure they are within [0, 255]
# - Convert to uint8
img = (img.permute(0, 2, 3, 1) * 127.5 + 128).clamp(0, 255).to(torch.uint8)

# Save the image using PIL
Image.fromarray(img[0].cpu().numpy(), 'RGB').save('generated_image.png')

print("Image saved as 'generated_image.png'")

The above code requires torch_utils and dnnlib to be accessible via PYTHONPATH. It does not need source code for the networks themselves — their class definitions are loaded from the pickle via torch_utils.persistence.

The pickle contains three networks. G and D are instantaneous snapshots taken during training, and G_ema represents a moving average of the generator weights over several training steps. The networks are regular instances of torch.nn.Module, with all of their parameters and buffers placed on the CPU at import and gradient computation disabled by default.

The generator consists of two submodules, G.mapping and G.synthesis, that can be executed separately.

See NVIDIA Repo for additional information.

A dedicated Repo for gaIA inference with ready-to-use scripts is on the way! Stay tuned!

Inference Samples

Generation

Uses

Scientific

  • Transfer Learning
  • Synthetic data generation
  • Future scenario simulations *
  • Comparative analysis *

*It is necessary to integrate external predictive climate models to generate future scenarios sumulation

Artistic

  • Art installations & exhibitions
  • Public awareness campaigns
  • Multimedia performances

License

This project and repository contains two licenses:

  1. Apache 2.0 License: Applies to the model and any modifications or additions made by The "What If" Project.
  2. NVIDIA Source Code License for StyleGAN3: Applies to the original StyleGAN3 software used for training the model.

Please see the LICENSE files in the repository for more details.

How to Contribute

Join us in using our model to make a differente! For more information and updates, visit gaIA spotlight.

Contact

For any questions or support, contact us through our website and follow us on Instagram.