Unsupervised Data Augmentation (UDA)

A PyTorch implementation for Unsupervised Data Augmentation.

Disclaimer

• This is not an official implementation. The official TensorFlow implementation is at this Github link.
• Plan to implement CIFAR10 and ImageNet experiments.

Updates

• 2019.06.28: CIFAR-10 with 4,000 labeled set achieves top-1 accuracy 93.69% without TSA. (on paper, 94.33% without TSA)

Performance

CIFAR-10

Exp	Top-1 acc(%) in paper	Top-1 acc(%)
Baseline	79.74	83.94
UDA (without TSA)	94.33	93.69
UDA	94.90	-

ImageNet (10% labeled)

Exp	Top-1 (paper)	Top-5 (paper)	Top-1	Top-5
RN50	55.09	77.26 (80.43 in S4L)	54.184	79.116
RN18	-	-	50.594	76.138
UDA(RN50)	68.66	88.52	-	-
S4L(RN50)	-	91.23 (ResNet50v2 4x)	-	-

TODO List

• CIFAR-10 baseline & UDA validation
• ImageNet ResNet50 baseline validation
• ImageNet ResNet50 UDA validation

MISC

• CIFAR10 baseline on paper is from Realistic Evaluation of Deep Semi-Supervised Learning Algorithms, and it may be sub-optimal OR use different data split from the UDA paper. A naive baseline with weight decay 5e-4 and 100K iteration with cosine annealing LR can achieve higher performance as shown in the table.
• CIFAR10 labeled set is from AutoAugment policy search subset.
• CIFAR10 AutoAugment policy includes full set (95 policies), rather than 25 policies.
• ImageNet labeled set is randomly selected 10% for each class.
• ImageNet baseline settings are from S4L: Self-Supervised Semi-Supervised Learning.