Model training is not always successful

[nikidimi]

Sometimes training produces a model that fails on some of the corners.
However, if train a model successfully, it will make it through both tracks without problems.

It seems to be related to the initial weights before training. If we load "good" initial weights, every training is successful.

[nikidimi]

Using init='normal' improves the situation - it is more likely to work than before, but still not 100%

Setting the convolution layer initial weights to -0.04, 0.08, 0.12, 0.02 works every time.