AFN

A pytorch implementation of Adversarially Adaptive Normalization for Single Domain Generalization. An official implementation of AFN: Adaptive Fusion Normalization via Encoder-Decoder Framework [https://arxiv.org/abs/2308.03321].

Experiment Results

Domain Generalization

Single domain generalization accuracies on Digits.

This picture shows the digits experiment in ASRNorm. It uses adversarial data augmentation, and only chooses MNIST as the training set, and the other three as the testing set. Furthermore, it increases the channels of MNIST and USPS to make sure these datasets compatible.

• Choosing one as the training set, and the rest as the testing set. (e.g. MNIST->SVHN, USPS, MNIST-M, SVHN->MNIST, USPS, MNIST-M). Iterate about 4 rounds until each dataset is treated as a training set, and get the average accuracies. (30 epochs, learning rate 0.1, SGD) Don`t use any adversarial augmentation like ASRNorm[https://arxiv.org/pdf/2106.01899.pdf]. Reduce the channels of M-MNIST and SVHN to ensure all datasets are compatible.

Method	MNIST	SVHN	USPS	MNIST-M	Avg.
BN	74.0	30.3	73.2	38.7	54.1
BIN	71.4	30.6	70.6	42.5	53.8
ASRNorm	75.6	34.0	70.9	45.5	56.5
AFN	77.6	33.8	73.4	44.8	57.4

• Choosing MNIST as the training set, and the rest as the testing set. Don`t use any adversarial augmentation like ASRNorm[https://arxiv.org/pdf/2106.01899.pdf]. Increase the channels of MNIST and USPS to ensure all datasets are compatible.

Method	SVHN	USPS	MNIST-M	Avg.
BN	27.8	76.9	52.7	52.5
ASRNorm	34.1	78.5	64.3	59.0
AFN	33.8	81.2	63.8	44.8

---

Single domain generalization accuracies on Cifar10-C.

• CIFAR-10 is used as the training domain, while CIFAR-10-C with different corruption types and corruption levels are used as the testing domains.

Method	Level 1	Level 2	Level 3	Level 4	Level 5	Avg.
BN	87.8	81.5	73.2	75.5	56.1	74.8
ASRNorm	89.4	86.1	82.9	78.6	72.9	82.0
AFN	89.3	86.6	83.7	79.9	77.0	83.3

---

Single domain generalisation accuracies on PACS.

• One domain is used as the training set and the other domains are used as the testing set.

Method	Art painting	Cartoon	Sketch	Photo	Avg.
ERM	70.9	76.5	53.1	42.2	60.7
RSC	73.4	75.9	56.2	41.6	61.8
RSC+ASRNorm	76.7	79.3	61.6	54.6	68.1
RSC+AFN	77.3	78.2	61.3	62.1	69.7

---

Multi-domains generalisation accuracies on PACS.

• Multi-domain generalisation accuracies on PACS. One domain is used as the test set and the other domains are used as the training sets. During training, we remove the domain labels.

Method	Art painting	Cartoon	Sketch	Photo	Avg.
ERM	82.7	78.7	78.6	95.1	83.8
RSC	82.7	79.8	80.3	95.6	84.6
RSC+ASRNorm	84.8	81.8	82.6	96.1	86.3
RSC+AFN	84.4	82.4	82.9	96.1	86.5

---

Image Classification

CIFAR-10

Backbone	BN	ASRNorm	AFN
ResNet18	0.9495	0.9400	0.9515
ResNet32	0.9419	0.9308	0.9432
ResNet56	0.9450	0.9388	0.9472
WRN-16-2	0.9404	0.9353	0.9447
VGG13	0.9179	0.9238	0.9343
VGG16	0.9435	0.9244	0.9448
VGG19	0.9442	0.9334	0.9451

---

CIFAR-100

Backbone	BN	ASRNorm	AFN
ResNet18	0.7327	0.6418	0.7457
ResNet32x4	0.7521	0.7054	0.7569
ResNet50	0.7509	0.6830	0.7558
WRN-40-10	0.7875	0.7538	0.7956
VGG13	0.6995	0.5150	0.7032
VGG16	0.6922	NAN	0.6972

---

M-MNIST

Backbone	BN	ASRNorm	AFN
ResNet18	0.9880	0.9870	0.9887
ResNet32	0.9887	0.9847	0.9893
WRN16-2	0.9859	0.9855	0.9891
VGG13	0.9830	0.9831	0.9875
VGG16	0.9838	NAN	0.9849

---

SVHN

Backbone	BN	ASRNorm	AFN
ResNet18	0.9427	0.9441	0.9462
ResNet32	0.9511	0.9472	0.9538
WRN-16-2	0.9512	0.9489	0.9523
VGG13	0.9411	0.9409	0.9427
VGG16	0.9433	NAN	0.9461

---