Class Anchor Clustering

Class Anchor Clustering (CAC) can be seen as a multi-class generalization of Deep One-Class Learning, where there are several centers \(\{\mu_1, \mu_2, ..., \mu_y\}\) in the output space of the model, one for each class. During training, the representation \(f_{\theta}(x)\) from class \(y\) is drawn towards the corresponding center \(\mu_y\).

Here, we train the model for 10 epochs on the CIFAR10 dataset, using a backbone pre-trained on the \(32 \times 32\) resized version of the ImageNet as a foundation.

import torch
from torch.optim import Adam
from torch.optim.lr_scheduler import CosineAnnealingLR
from torch.utils.data import DataLoader
from torchmetrics import Accuracy
from torchvision.datasets import CIFAR10
from tqdm import tqdm

from pytorch_ood.dataset.img import Textures
from pytorch_ood.loss import CACLoss
from pytorch_ood.model import WideResNet
from pytorch_ood.utils import OODMetrics, ToUnknown, fix_random_seed, is_known

fix_random_seed(123)

n_epochs = 10
device = "cuda:0"

trans = WideResNet.transform_for("imagenet32-nocifar")

# setup IN training data
dataset_in_train = CIFAR10(root="data", train=True, download=True, transform=trans)

# setup IN test data
dataset_in_test = CIFAR10(root="data", train=False, transform=trans)

# setup OOD test data, use ToUnknown() to mark labels as OOD
dataset_out_test = Textures(
    root="data", download=True, transform=trans, target_transform=ToUnknown()
)

# create data loaders
train_loader = DataLoader(dataset_in_train, batch_size=64, shuffle=True, num_workers=16)
test_loader = DataLoader(dataset_in_test + dataset_out_test, batch_size=64, num_workers=16)

Create DNN, pretrained on the imagenet excluding cifar10 classes. We have to replace the final layer to match the number of classes.

model = WideResNet(num_classes=1000, pretrained="imagenet32-nocifar")
model.fc = torch.nn.Linear(model.fc.in_features, 10)
model.to(device)

opti = Adam(model.parameters())
criterion = CACLoss(n_classes=10, magnitude=5, alpha=2).to(device)
scheduler = CosineAnnealingLR(opti, T_max=n_epochs * len(train_loader))

Define a function that evaluates the model

def test():
    metrics = OODMetrics()
    acc = Accuracy(num_classes=10, task="multiclass")

    model.eval()

    with torch.no_grad(), tqdm(test_loader, desc="Testing") as bar:
        for x, y in bar:
            # calculate embeddings
            z = model(x.to(device))
            # calculate the distance of each embedding to each center
            distances = criterion.distance(z).cpu()
            # the CAC Loss proposes its own method for score calculation.
            # We could, however, also use the minimum distance.
            metrics.update(CACLoss.score(distances), y)
            known = is_known(y)
            if known.any():
                acc.update(distances[known].min(dim=1).indices, y[known])

    print(metrics.compute())
    print(f"Accuracy: {acc.compute().item():.2%}")
    model.train()

Start training

for epoch in range(n_epochs):
    loss_ema = 0

    with tqdm(train_loader, desc=f"Epoch {epoch}") as bar:
        for x, y in bar:
            # calculate embeddings
            z = model(x.to(device))
            # calculate the distance of each embedding to each center
            distances = criterion.distance(z)
            # calculate CAC loss, based on distances to centers
            loss = criterion(distances, y.cuda())
            opti.zero_grad()
            loss.backward()
            opti.step()
            scheduler.step()

            loss_ema = loss.item() if not loss_ema else 0.99 * loss_ema + 0.01 * loss.item()
            bar.set_postfix_str(f"loss: {loss_ema:.3f} lr: {scheduler.get_last_lr()[0]:.6f}")

        test()

{‘AUROC’: 0.8958120346069336, ‘AUPR-IN’: 0.8456918001174927, ‘AUPR-OUT’: 0.8196930885314941, ‘FPR95TPR’: 0.5187000036239624}

Accuracy: 93.80%