When I trained the model on the flyingthins3d dataset, I printed out the value of loss and found that it was NaN. Why is this?