Kiwinicki/sat2map-generator

Loading the Generator Model

To load and initialize the Generator (based on CycleGAN with better cycles) model from the repository, follow these steps:

1. Install Required Packages

Ensure you have the necessary Python packages installed:

pip install torch==2.5.1 torchvision==0.20.1 huggingface_hub

2. Download Model Files

Retrieve the generator.pth and model.py files from the Hugging Face repository using the huggingface_hub library:

from huggingface_hub import hf_hub_download

repo_id = "Kiwinicki/sat2map-generator"
model_path = hf_hub_download(repo_id=repo_id, filename="generator.pth")
generator_code_path = hf_hub_download(repo_id=repo_id, filename="model.py")

3. Load the Model

Import the Generator class and load the model weights from the .pth file:

import torch
from model import Generator, GeneratorConfig


# Load the generator model
cfg = GeneratorConfig()
generator = Generator(cfg)
generator.load_state_dict(torch.load('generator.pth'))
generator.eval()

# Test the model
x = torch.randn([1, cfg.channels, 256, 256])
out = generator(x)
print(f"Output shape: {out.shape}")

4. Model Configuration

The model uses the following default configuration:

• channels: 3 (RGB images)
• num_features: 64 (base number of features)
• num_residuals: 12 (number of residual blocks)
• depth: 4 (network depth)

The generator is now ready for inference on satellite-to-map translation tasks.

Model trained by Andrii Norets from "Czarna Magia".