Vanilla Unfolded algorithm for super-resolution

This is a simple example to show how to use vanilla unfolded Plug-and-Play. The DnCNN denoiser and the algorithm parameters (stepsize, regularization parameters) are trained jointly. For simplicity, we show how to train the algorithm on a small dataset. For optimal results, use a larger dataset.

import deepinv as dinv
import torch
from deepinv.models.utils import get_weights_url
from torch.utils.data import DataLoader
from deepinv.optim.data_fidelity import L2
from deepinv.optim.prior import PnP
from deepinv.optim import DRS
from torchvision import transforms
from deepinv.utils import get_data_home
from deepinv.datasets import BSDS500

Setup paths for data loading and results.

BASE_DIR = get_data_home()
DATA_DIR = BASE_DIR / "measurements"
RESULTS_DIR = BASE_DIR / "results"
CKPT_DIR = BASE_DIR / "ckpts"

# 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"

Selected GPU 0 with 5036.125 MiB free memory

Load base image datasets and degradation operators.

In this example, we use the CBSD500 dataset for training and the Set3C dataset for testing.

img_size = 64 if torch.cuda.is_available() else 32
n_channels = 3  # 3 for color images, 1 for gray-scale images
operation = "super-resolution"

Generate a dataset of low resolution images and load it.

We use the Downsampling class from the physics module to generate a dataset of low resolution images.

# For simplicity, we use a small dataset for training.
# To be replaced for optimal results. For example, you can use the larger DIV2K or LSDIR datasets (also provided in the library).

# Specify the  train and test transforms to be applied to the input images.
test_transform = transforms.Compose(
    [transforms.CenterCrop(img_size), transforms.ToTensor()]
)
train_transform = transforms.Compose(
    [transforms.RandomCrop(img_size), transforms.ToTensor()]
)
# Define the base train and test datasets of clean images.
train_base_dataset = BSDS500(
    BASE_DIR, download=True, train=True, transform=train_transform
)
test_base_dataset = BSDS500(
    BASE_DIR, download=False, train=False, transform=test_transform
)

# Use parallel dataloader if using a GPU to speed up training, otherwise, as all computes are on CPU, use synchronous
# dataloading.
num_workers = 4 if torch.cuda.is_available() else 0

# Degradation parameters
factor = 2
noise_level_img = 0.03

# Generate the gaussian blur downsampling operator.
physics = dinv.physics.Downsampling(
    filter="gaussian",
    img_size=(n_channels, img_size, img_size),
    factor=factor,
    device=device,
    noise_model=dinv.physics.GaussianNoise(sigma=noise_level_img),
)
my_dataset_name = "demo_unfolded_sr"
n_images_max = (
    None if torch.cuda.is_available() else 10
)  # max number of images used for training (use all if you have a GPU)
measurement_dir = DATA_DIR / "BSDS500" / operation
generated_datasets_path = dinv.datasets.generate_dataset(
    train_dataset=train_base_dataset,
    test_dataset=test_base_dataset,
    physics=physics,
    device=device,
    save_dir=measurement_dir,
    train_datapoints=n_images_max,
    num_workers=num_workers,
    dataset_filename=str(my_dataset_name),
)

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

0it [00:00, ?it/s]
2.69MB [00:00, 28.0MB/s]
5.56MB [00:00, 29.1MB/s]
8.38MB [00:00, 28.7MB/s]
11.1MB [00:00, 28.5MB/s]
13.9MB [00:00, 28.6MB/s]
16.6MB [00:00, 27.9MB/s]
19.3MB [00:00, 28.0MB/s]
22.4MB [00:00, 29.3MB/s]
25.6MB [00:00, 30.4MB/s]
28.8MB [00:01, 31.4MB/s]
31.9MB [00:01, 31.6MB/s]
35.0MB [00:01, 31.8MB/s]
38.6MB [00:01, 33.2MB/s]
41.9MB [00:01, 33.6MB/s]
45.1MB [00:01, 33.4MB/s]
48.4MB [00:01, 33.6MB/s]
51.7MB [00:01, 33.3MB/s]
54.9MB [00:01, 33.3MB/s]
58.1MB [00:01, 28.0MB/s]
60.9MB [00:02, 24.5MB/s]
63.4MB [00:02, 22.4MB/s]
65.7MB [00:02, 21.1MB/s]
67.8MB [00:02, 20.1MB/s]
69.8MB [00:02, 19.5MB/s]
71.8MB [00:02, 18.9MB/s]
73.6MB [00:02, 18.5MB/s]
75.4MB [00:03, 18.2MB/s]
77.2MB [00:03, 18.1MB/s]
79.1MB [00:03, 18.5MB/s]
80.9MB [00:03, 18.1MB/s]
82.8MB [00:03, 18.3MB/s]
84.6MB [00:03, 18.0MB/s]
86.4MB [00:03, 17.7MB/s]
88.4MB [00:03, 18.3MB/s]
90.1MB [00:03, 18.1MB/s]
91.9MB [00:03, 17.5MB/s]
93.6MB [00:04, 17.5MB/s]
95.3MB [00:04, 17.5MB/s]
97.1MB [00:04, 17.7MB/s]
98.8MB [00:04, 17.2MB/s]
101MB [00:04, 18.1MB/s]
102MB [00:04, 17.9MB/s]
104MB [00:04, 17.7MB/s]
106MB [00:04, 17.8MB/s]
108MB [00:04, 18.0MB/s]
110MB [00:05, 17.9MB/s]
111MB [00:05, 17.9MB/s]
113MB [00:05, 18.2MB/s]
115MB [00:05, 18.0MB/s]
117MB [00:05, 17.6MB/s]
119MB [00:05, 18.0MB/s]
120MB [00:05, 17.7MB/s]
122MB [00:05, 18.1MB/s]
124MB [00:05, 18.0MB/s]
126MB [00:05, 17.6MB/s]
128MB [00:06, 17.7MB/s]
129MB [00:06, 18.0MB/s]
131MB [00:06, 18.2MB/s]
133MB [00:06, 9.08MB/s]
138MB [00:06, 17.0MB/s]
143MB [00:07, 18.3MB/s]
149MB [00:07, 25.0MB/s]
152MB [00:07, 27.2MB/s]
156MB [00:07, 17.3MB/s]
158MB [00:08, 12.2MB/s]
160MB [00:08, 20.2MB/s]

Extracting:   0%|          | 0/2492 [00:00<?, ?it/s]
Extracting:   4%|▍         | 102/2492 [00:00<00:02, 1018.01it/s]
Extracting:   8%|▊         | 204/2492 [00:00<00:02, 977.53it/s]
Extracting:  12%|█▏        | 305/2492 [00:00<00:02, 987.93it/s]
Extracting:  16%|█▌        | 404/2492 [00:00<00:02, 974.50it/s]
Extracting:  20%|██        | 503/2492 [00:00<00:02, 977.24it/s]
Extracting:  25%|██▍       | 615/2492 [00:00<00:01, 1023.53it/s]
Extracting:  29%|██▉       | 724/2492 [00:00<00:01, 1043.36it/s]
Extracting:  33%|███▎      | 833/2492 [00:00<00:01, 1056.82it/s]
Extracting:  38%|███▊      | 941/2492 [00:00<00:01, 1063.15it/s]
Extracting:  42%|████▏     | 1048/2492 [00:01<00:01, 799.27it/s]
Extracting:  46%|████▌     | 1138/2492 [00:01<00:02, 594.66it/s]
Extracting:  49%|████▊     | 1211/2492 [00:01<00:02, 504.31it/s]
Extracting:  51%|█████     | 1272/2492 [00:01<00:02, 472.90it/s]
Extracting:  53%|█████▎    | 1327/2492 [00:01<00:02, 436.47it/s]
Extracting:  55%|█████▌    | 1376/2492 [00:02<00:02, 407.99it/s]
Extracting:  57%|█████▋    | 1420/2492 [00:02<00:02, 393.04it/s]
Extracting:  59%|█████▊    | 1462/2492 [00:02<00:02, 382.19it/s]
Extracting:  60%|██████    | 1502/2492 [00:02<00:02, 370.83it/s]
Extracting:  65%|██████▍   | 1608/2492 [00:02<00:01, 533.78it/s]
Extracting:  67%|██████▋   | 1666/2492 [00:02<00:01, 527.71it/s]
Extracting:  69%|██████▉   | 1722/2492 [00:02<00:01, 387.59it/s]
Extracting: 100%|██████████| 2492/2492 [00:02<00:00, 839.22it/s]
Dataset has been saved at datasets/measurements/BSDS500/super-resolution/demo_unfolded_sr0.h5

Define the unfolded PnP algorithm.

The chosen algorithm is here DRS (Douglas-Rachford Splitting). Note that if the prior (resp. a parameter) is initialized with a list of length max_iter, then a distinct model (resp. parameter) is trained for each iteration. For fixed trained model prior (resp. parameter) across iterations, initialize with a single element.

# Unrolled optimization algorithm parameters
max_iter = 5  # number of unfolded layers

# Select the data fidelity term
data_fidelity = L2()

# Set up the trainable denoising prior
# Here the prior model is common for all iterations
prior = PnP(denoiser=dinv.models.DnCNN(depth=20, pretrained="download").to(device))

# The parameters are initialized with a list of length max_iter, so that a distinct parameter is trained for each iteration.
stepsize = [1.0] * max_iter  # stepsize of the algorithm
sigma_denoiser = [
    1.0
] * max_iter  # noise level parameter of the denoiser (not used by DnCNN)
beta = 1.0  # relaxation parameter of the Douglas-Rachford splitting
trainable_params = [
    "stepsize",
    "beta",
    "sigma_denoiser",
]  # define which parameters are trainable

# Logging parameters
verbose = True

# Define the unfolded trainable model.
model = DRS(
    stepsize=stepsize,
    sigma_denoiser=sigma_denoiser,
    beta=beta,
    trainable_params=trainable_params,
    data_fidelity=data_fidelity,
    max_iter=max_iter,
    prior=prior,
    unfold=True,
)

Define the training parameters.

We use the Adam optimizer and the StepLR scheduler.

# training parameters
epochs = 5 if torch.cuda.is_available() else 1
learning_rate = 5e-4
train_batch_size = 32 if torch.cuda.is_available() else 1
test_batch_size = 3

# choose optimizer and scheduler
optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate, weight_decay=1e-8)

# If working on CPU, start with a pretrained model to reduce training time
if not torch.cuda.is_available():
    file_name = "demo_vanilla_unfolded.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"])
    optimizer.load_state_dict(ckpt["optimizer"])

# choose supervised training loss
losses = [dinv.loss.SupLoss(metric=dinv.metric.MSE())]

train_dataloader = DataLoader(
    train_dataset, batch_size=train_batch_size, num_workers=num_workers, shuffle=True
)
test_dataloader = DataLoader(
    test_dataset, batch_size=test_batch_size, num_workers=num_workers, shuffle=False
)

Train the network

We train the network using the deepinv.Trainer class.

trainer = dinv.Trainer(
    model,
    physics=physics,
    train_dataloader=train_dataloader,
    eval_dataloader=test_dataloader,
    epochs=epochs,
    losses=losses,
    optimizer=optimizer,
    device=device,
    early_stop=True,  # set to None to disable early stopping
    save_path=str(CKPT_DIR / operation),
    verbose=verbose,
    show_progress_bar=False,  # disable progress bar for better vis in sphinx gallery.
)

model = trainer.train()

/local/jtachell/deepinv/deepinv/deepinv/training/trainer.py:1352: UserWarning: non_blocking_transfers=True but DataLoader.pin_memory=False; set pin_memory=True to overlap host-device copies with compute.
  self.setup_train()
The model has 668238 trainable parameters
/local/jtachell/deepinv/deepinv/deepinv/training/trainer.py:514: UserWarning: early_stop should be an integer or None. Setting early_stop=3. This behaviour will be deprecated in future versions.
  warnings.warn(
Train epoch 0: TotalLoss=0.009, PSNR=21.591
Eval epoch 0: PSNR=20.581
Best model saved at epoch 1
Train epoch 1: TotalLoss=0.007, PSNR=23.006
Eval epoch 1: PSNR=21.237
Best model saved at epoch 2
Train epoch 2: TotalLoss=0.007, PSNR=23.412
Eval epoch 2: PSNR=21.421
Best model saved at epoch 3
Train epoch 3: TotalLoss=0.006, PSNR=23.779
Eval epoch 3: PSNR=21.717
Best model saved at epoch 4
Train epoch 4: TotalLoss=0.006, PSNR=24.052
Eval epoch 4: PSNR=21.854
Best model saved at epoch 5

Test the network

trainer.test(test_dataloader)

test_sample, _ = next(iter(test_dataloader))
model.eval()
test_sample = test_sample.to(device)

# Get the measurements and the ground truth
y = physics(test_sample)
with torch.no_grad():
    rec = model(y, physics=physics)

backprojected = physics.A_adjoint(y)

dinv.utils.plot(
    [backprojected, rec, test_sample],
    titles=["Linear", "Reconstruction", "Ground truth"],
    suptitle="Reconstruction results",
)

/local/jtachell/deepinv/deepinv/deepinv/training/trainer.py:1544: UserWarning: non_blocking_transfers=True but DataLoader.pin_memory=False; set pin_memory=True to overlap host-device copies with compute.
  self.setup_train(train=False)
Eval epoch 0: PSNR=21.854, PSNR no learning=9.122
Test results:
PSNR no learning: 9.122 +- 2.903
PSNR: 21.854 +- 3.672
/local/jtachell/deepinv/deepinv/deepinv/utils/plotting.py:387: UserWarning: This figure was using a layout engine that is incompatible with subplots_adjust and/or tight_layout; not calling subplots_adjust.
  fig.subplots_adjust(top=0.75)