remove style loss div by c*c in losses.py

[yihongXU]

Hi,
In utils.py gram matrix is already normalized by HxWxC :

        b, h, w, c = layer.get_shape().as_list() \
        num_elements = h*w*c
        features_matrix = tf.reshape(layer, tf.stack([b, -1, c]))
        gram_matrix = tf.matmul(features_matrix, features_matrix, transpose_a=True)
        gram_matrix = gram_matrix / tf.cast(num_elements, tf.float32)  

So I think that it might be wrong to normalize style loss by c1xc2 (c1 = c2 = c), since the stlye loss is already normalized by H^2 x W^2 x C^2 after L2 loss (cf. the formulation of style loss in Johnson's paper.)