Strange Result Validation for Custom Datasets

[hanwenran1]

Hi! Thanks for sharing your work :)
Shouldn't the accuracy rate of training on PAWS with labeled dataset A and unlabeled dataset B be higher than that of full supervised learning with labeled dataset A alone? I can achieve an accuracy of 90% on Resnet using a self built labeled dataset. Why did I add unlabeled data into this dataset and only achieve an accuracy of over 50% using Paws? It's strange.