face parsing is not precise -> remove face mask/ add total variation loss
attention
in order to speed up the training process, we fix the (content,style) pair. remove this constraint should improve the model's performance on validation data and test data.
we simply apply the face parser in opencv, the original paper use different face parsing tool.