Decreasing MXNet recognition model performance immediately after the start of training

Hello, thank you for grait framework with MXNet support!

I download casia-webface from your dataset-zoo.
Then i download mobilefacenet model and get 99% quality on LFW.
But when i start training on casia-webface, no matter what lr i use, model accuracy decreases to 97% or smaller. (I use lr=1e-15 and verbose=200. Training log says that training is called exactly with this lr).

I trayed this with another training dataset and situation is the same. Model performance is decreasing immediately.