silent bug, training with NaNs values

[AzulGarza]

We should include an exception when missing values are passed in the y column. The loss becomes nan during training but it is hard to notice that the problem is in the data.