Self-supervised denoising with the Generalized R2R loss.

This example shows you how to train a denoiser network in a fully self-supervised way, using noisy images only via the Generalized Recorrupted2Recorrupted (GR2R) loss Monroy et al.[1], which exploits knowledge about the noise distribution. You can change the noise distribution by selecting from predefined noise models such as Gaussian, Poisson, and Gamma noise.

from pathlib import Path

import torch
from torch.utils.data import DataLoader
from torchvision import transforms, datasets

import deepinv as dinv
from deepinv.utils import get_data_home
from deepinv.models.utils import get_weights_url

Setup paths for data loading and results.

BASE_DIR = Path(".")
DATA_DIR = BASE_DIR / "measurements"
CKPT_DIR = BASE_DIR / "ckpts"
ORIGINAL_DATA_DIR = get_data_home()

# Set the global random seed from pytorch to ensure reproducibility of the example.
torch.manual_seed(0)

device = dinv.utils.get_freer_gpu() if torch.cuda.is_available() else "cpu"
print(device)

cpu

Load base image datasets

In this example, we use the MNIST dataset as the base image dataset.

operation = "denoising"
train_dataset_name = "MNIST"

transform = transforms.Compose([transforms.ToTensor()])

train_dataset = datasets.MNIST(
    root=ORIGINAL_DATA_DIR, train=True, transform=transform, download=True
)
test_dataset = datasets.MNIST(
    root=ORIGINAL_DATA_DIR, train=False, transform=transform, download=True
)

Downloading http://yann.lecun.com/exdb/mnist/train-images-idx3-ubyte.gz
Failed to download (trying next):
HTTP Error 404: Not Found

Downloading https://ossci-datasets.s3.amazonaws.com/mnist/train-images-idx3-ubyte.gz
Downloading https://ossci-datasets.s3.amazonaws.com/mnist/train-images-idx3-ubyte.gz to datasets/MNIST/raw/train-images-idx3-ubyte.gz

  0%|          | 0/9912422 [00:00<?, ?it/s]
 94%|█████████▍| 9306112/9912422 [00:00<00:00, 91994161.58it/s]
100%|██████████| 9912422/9912422 [00:00<00:00, 94298655.11it/s]
Extracting datasets/MNIST/raw/train-images-idx3-ubyte.gz to datasets/MNIST/raw

Downloading http://yann.lecun.com/exdb/mnist/train-labels-idx1-ubyte.gz
Failed to download (trying next):
HTTP Error 404: Not Found

Downloading https://ossci-datasets.s3.amazonaws.com/mnist/train-labels-idx1-ubyte.gz
Downloading https://ossci-datasets.s3.amazonaws.com/mnist/train-labels-idx1-ubyte.gz to datasets/MNIST/raw/train-labels-idx1-ubyte.gz

  0%|          | 0/28881 [00:00<?, ?it/s]
100%|██████████| 28881/28881 [00:00<00:00, 5159101.10it/s]
Extracting datasets/MNIST/raw/train-labels-idx1-ubyte.gz to datasets/MNIST/raw

Downloading http://yann.lecun.com/exdb/mnist/t10k-images-idx3-ubyte.gz
Failed to download (trying next):
HTTP Error 404: Not Found

Downloading https://ossci-datasets.s3.amazonaws.com/mnist/t10k-images-idx3-ubyte.gz
Downloading https://ossci-datasets.s3.amazonaws.com/mnist/t10k-images-idx3-ubyte.gz to datasets/MNIST/raw/t10k-images-idx3-ubyte.gz

  0%|          | 0/1648877 [00:00<?, ?it/s]
100%|██████████| 1648877/1648877 [00:00<00:00, 39926631.05it/s]
Extracting datasets/MNIST/raw/t10k-images-idx3-ubyte.gz to datasets/MNIST/raw

Downloading http://yann.lecun.com/exdb/mnist/t10k-labels-idx1-ubyte.gz
Failed to download (trying next):
HTTP Error 404: Not Found

Downloading https://ossci-datasets.s3.amazonaws.com/mnist/t10k-labels-idx1-ubyte.gz
Downloading https://ossci-datasets.s3.amazonaws.com/mnist/t10k-labels-idx1-ubyte.gz to datasets/MNIST/raw/t10k-labels-idx1-ubyte.gz

  0%|          | 0/4542 [00:00<?, ?it/s]
100%|██████████| 4542/4542 [00:00<00:00, 20506489.52it/s]
Extracting datasets/MNIST/raw/t10k-labels-idx1-ubyte.gz to datasets/MNIST/raw

Generate a dataset of noisy images

Generate a dataset of noisy images corrupted by Poisson noise. The predefined noise models in the physics module include Gaussian, Poisson, and Gamma noise. Here, we use Poisson noise as an example, but you can also use Gaussian or Gamma noise.

Note

We use a subset of the whole training set to reduce the computational load of the example. We recommend to use the whole set by setting n_images_max=None to get the best results.

# defined physics
predefined_noise_models = dict(
    gaussian=dinv.physics.GaussianNoise(sigma=0.1),
    poisson=dinv.physics.PoissonNoise(gain=0.5),
    gamma=dinv.physics.GammaNoise(l=10.0),
)

noise_name = "poisson"
noise_model = predefined_noise_models[noise_name]
physics = dinv.physics.Denoising(noise_model)
operation = f"{operation}_{noise_name}"

# Use parallel dataloader if using a GPU to speed up training,
# otherwise, as all computes are on CPU, use synchronous data loading.
num_workers = 0 if torch.cuda.is_available() else 0

n_images_max = (
    100 if torch.cuda.is_available() else 5
)  # number of images used for training

measurement_dir = DATA_DIR / train_dataset_name / operation
deepinv_datasets_path = dinv.datasets.generate_dataset(
    train_dataset=train_dataset,
    test_dataset=test_dataset,
    physics=physics,
    device=device,
    save_dir=measurement_dir,
    train_datapoints=n_images_max,
    test_datapoints=n_images_max,
    num_workers=num_workers,
    dataset_filename="demo_r2r",
)

train_dataset = dinv.datasets.HDF5Dataset(path=deepinv_datasets_path, train=True)
test_dataset = dinv.datasets.HDF5Dataset(path=deepinv_datasets_path, train=False)

Dataset has been saved at measurements/MNIST/denoising_poisson/demo_r2r0.h5

Set up the denoiser network

We use a simple U-Net architecture with 2 scales as the denoiser network.

model = dinv.models.ArtifactRemoval(
    dinv.models.UNet(in_channels=1, out_channels=1, scales=2, residual=False).to(device)
)

Set up the training parameters

We set deepinv.loss.R2RLoss as the training loss.

Note

There are GR2R losses for various noise distributions, which can be specified by the noise model.

epochs = 1  # choose training epochs
learning_rate = 1e-4
batch_size = 64 if torch.cuda.is_available() else 1

# choose self-supervised training loss
loss = dinv.loss.R2RLoss(noise_model=None)
model = loss.adapt_model(model)  # important step!

# choose optimizer and scheduler
optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate, weight_decay=1e-4)
scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=int(epochs * 0.8) + 1)

# # start with a pretrained model to reduce training time

if noise_name == "poisson":
    file_name = "ckp_10_demo_r2r_poisson.pth"
    url = get_weights_url(model_name="demo", file_name=file_name)
    ckpt = torch.hub.load_state_dict_from_url(
        url, map_location=lambda storage, loc: storage, file_name=file_name
    )
    model.load_state_dict(ckpt["state_dict"])

Downloading: "https://huggingface.co/deepinv/demo/resolve/main/ckp_10_demo_r2r_poisson.pth?download=true" to /home/runner/.cache/torch/hub/checkpoints/ckp_10_demo_r2r_poisson.pth

  0%|          | 0.00/5.14M [00:00<?, ?B/s]
100%|██████████| 5.14M/5.14M [00:00<00:00, 117MB/s]

Train the network

To simulate a realistic self-supervised learning scenario, we do not use any supervised metrics for training, such as PSNR or SSIM, which require clean ground truth images.

Tip

We can use the same self-supervised loss for evaluation, as it does not require clean images, to monitor the training process (e.g. for early stopping). This is done automatically when metrics=None and early_stop>0 in the trainer.

verbose = True  # print training information

train_dataloader = DataLoader(
    train_dataset, batch_size=batch_size, num_workers=num_workers, shuffle=False
)

test_dataloader = DataLoader(
    test_dataset, batch_size=batch_size, num_workers=num_workers, shuffle=False
)

# Initialize the trainer
trainer = dinv.Trainer(
    model=model,
    physics=physics,
    epochs=epochs,
    scheduler=scheduler,
    losses=loss,
    optimizer=optimizer,
    device=device,
    metrics=None,  # no supervised metrics
    train_dataloader=train_dataloader,
    eval_dataloader=test_dataloader,
    early_stop=2,  # early stop using the self-supervised loss on the test set
    compute_eval_losses=True,  # use self-supervised loss for evaluation
    early_stop_on_losses=True,  # stop using self-supervised eval loss
    plot_images=True,
    save_path=str(CKPT_DIR / operation),
    verbose=verbose,
    show_progress_bar=False,  # disable progress bar for better vis in sphinx gallery.
)

# Train the network
model = trainer.train()

• 
• 

The model has 444737 trainable parameters
Train epoch 0: TotalLoss=0.414
Eval epoch 0: TotalLoss=0.495
Best model saved at epoch 1

Test the network

We now assume that we have access to a small test set of clean images to evaluate the performance of the trained network. and we compute the PSNR between the denoised images and the clean ground truth images.

trainer.test(test_dataloader, metrics=dinv.metric.PSNR())

Eval epoch 0: TotalLoss=0.425, PSNR=20.752, PSNR no learning=12.709
Test results:
PSNR no learning: 12.709 +- 2.129
PSNR: 20.752 +- 1.770

{'PSNR no learning': 12.70888900756836, 'PSNR no learning_std': 2.1292703663268076, 'PSNR': 20.75245475769043, 'PSNR_std': 1.7697566969332181}

References:

[1]

Brayan Monroy, Jorge Bacca, and Julián Tachella. Generalized recorrupted-to-recorrupted: self-supervised learning beyond gaussian noise. In Proceedings of the Computer Vision and Pattern Recognition Conference, 28155–28164. 2025.