I'm training ResNet 50 for 1 class problem but getting predictions all around zeros, what is wrong?

[Ilya Zhenin]

I took ResNet50 architecture replacing loss layer with SigmoidCrossEntropy because I have just zero and one classes. 

I've tried to train it, through iterations loss gets reduced  from 0.79 +- to 0.30+-, but all predictions that I get from network are close to zero or ones, network just labeles everything as one class but, I guess, batch normalization wont allow it to predict absolute zeros so it's 0.07+-0.03 for any image.

I've managed to train on this dataset VGG-16 getting fine results. Could it be that ResNet50 somehow not fitted for one class problem, or is just to deep for cosiderable 'simple' 1 class task? Or something else?