A problem regarding training on ImageNet dataset

I just use NAS to search a architecture on ImageNet, but I have a searching result like 50% acc, testing result always 1%. I don’t why this happens. I checked the dataset, there is no mismatch for different classes. I also train the transofrmed architecture searched on CIFAR10 and 100 successfully. But still got 1% acc on testset of ImageNet. It’s very weird. Is it a overfitting problem? Can this overfitting lead to 1% acc ?! I’m pleased if anyone can help me. Thank you so much.