Synthesize new images using the score-based generative models.

Currently, diffusion_sde supports release of Python 3.7 onwards.

To install the current release:

$ pip install -U diffusion_sde

Start by instantiating a dataset class with a path where the custom dataset is located

from diffusion_sde import datasets

# Specify the path of the custom dataset in the dataset class
ds = datasets(path_to_dataset)

Then, instantiate the diffSDE class to train the model and generate samples and pass the dataset using .set_loaders() method

from diffusion_sde import diffSDE

# Instantiate the diffSDE class
cls_diff = diffSDE()

# Set the dataloaders by passing the dataset instantiation as above
cls_diff.set_loaders(dataset=ds)

Begin the model training using the .train() method and select the desired number of epochs for training.

# Train the model
cls_diff.train(n_iters)

Generate the samples from the trained model with the .generate_samples() method and specify the desired number of steps for the sampler. We suggest setting the value of n_steps in the range of $\sim1500$-$2000$ steps to produce high-quality samples

# Generate samples from the trained model
cls_diff.generate_samples(n_steps)

Pretrained model can be loaded to generate new samples on Google Colab