Parameter pad disturbs training step?

[Antonio Paes]

Hi guys, 

I'm working with a CNN for image reconstruction, and I want my images no reduce your size after convolutions, so I use the  pad parameter set value equal to rounded half size of filter. 

For validate the generation of the files I do two experiments:

1 - First I don't use pad parameter and the network converge without problems.

2 - Using pad parameter the network don't converge and my loss stay between 15 and 16, which is very high.

My output function is Euclidean loss.

Anyone has any idea why this happened?

Thanks!