This is related to issues#439,#2280 and#2743

I am unable to reproduce my results from run to run with either tensorflow or theano backend (running on CPU) and I do callnp.random.seed(11) before importing keras.

>>> theano.__version__'0.9.0.dev-d18ce33b02273026ea7986fe9f9878546db6e8cf'>>> tensorflow.__version__'1.1.0'>>> np.__version__'1.12.1'>>> keras.__version__'2.0.4'