The NAN problem seems to be due to theano backend. I changed it to tensorflow it worked fine.
When I made vgg16 submission, I got 0.16 loss.
Now, with resnet50 (according to @jeremy lesson2 notebook from DL1 V2), my submission received only 4.4))))
Seems some error on saving probs stage, but have not figured out yet…
Thanks in advance for any advice!