Questions about the precision in validation

[Zwette]

I trained the ResNet architecture (cifar_shakeshake26 in Pytorch version) on cifar-10 dataset with 1000 unlabeled images and 44000 labeled images (the resting 5000 images are used for validation) for about 180 epochs, setting the bach-size 256, labeled batch-size 62.
But I observed that the validation precision (top 1) would first rise from 43% up to 50% and then fall to only 13% (began to fall after about 10 epochs) along the training process. I was so puzzled about this phenomenon. Besides, the precision in training always rise and never fall, why the validation precision would fall??

[lukk47]

@Easquel Have you solved it?
I come across with the same issue. My validation accuracy stays around 60% to the end of training while the training accuracy can go up to 100%.

[lukk47]

@tarvaina May you help the pytorch on cifar10?
I ran the code with pytorch0.3.1. The validation accuracy went up to 50% and then went down to 10~20%.

[Wangzheaos]

训练了ResNet体

should be 1000 labeled images and 44000 labeled images