DICE Loss Function becomes unlearnable

Question

I am doing a binary segmentation problem where I use the custom DICE Loss function designed as below.

import keras
import keras.backend as K

def DiceLoss(targets, inputs, smooth=1e-6):

   inputs = K.flatten(inputs)
   targets = K.flatten(targets)

   intersection = K.sum(targets*inputs)
   dice = (2.*intersection + smooth) / (K.sum(targets) + K.sum(inputs) + smooth)
   return 1 - dice

Here the U-net learns when the loss has (2.*intersection + smooth) in the numerator but it doesn't learn when (2*intersection + smooth) is the numerator. The fact that 2 is a float or a int decides the learnability of the loss function. The loss cannot turn into integer value due to the 2 being int or float, since the tensors used in the calculation and K.sum also returns float values. So I am pretty perplexed why the change of 2 to 2. makes things right. Is there any relation with the computational graph built for back-propagation with this?