layers_3.py

# coding:utf-8
import numpy as np
import struct
import os
import scipy.io
import time

class ContentLossLayer(object):
    def __init__(self):
        print('\tContent loss layer.')
    def forward(self, input_layer, content_layer):
        # TODO： 计算风格迁移图像和目标内容图像的内容损失
        loss = np.sum((input_layer - content_layer) ** 2) / (2 * input_layer.shape[0] * input_layer.shape[1] * input_layer.shape[2] * input_layer.shape[3])
        return loss
    def backward(self, input_layer, content_layer):
        # TODO： 计算内容损失的反向传播
        bottom_diff = (input_layer - content_layer) / (input_layer.shape[0] * input_layer.shape[1] * input_layer.shape[2] * input_layer.shape[3])
        return bottom_diff

class StyleLossLayer(object):
    def __init__(self):
        print('\tStyle loss layer.')
    def forward(self, input_layer, style_layer):
        # TODO： 计算风格迁移图像和目标风格图像的Gram 矩阵
        self.style_layer_reshape = np.reshape(style_layer, [style_layer.shape[0], style_layer.shape[1], -1]) #除第一维（样本数）与第二维（通道数）的其他维度合并
        self.gram_style = np.zeros([style_layer.shape[0], style_layer.shape[1], style_layer.shape[1]]) #G
        self.input_layer_reshape = np.reshape(input_layer, [input_layer.shape[0], input_layer.shape[1], -1])
        self.gram_input = np.zeros([input_layer.shape[0], input_layer.shape[1], input_layer.shape[1]]) #A
        for idxn in range(input_layer.shape[0]):
            self.gram_input[idxn, :, :] = np.matmul(self.input_layer_reshape[idxn, :, :], self.input_layer_reshape[idxn, :, :].T)
            self.gram_style[idxn, :, :] = np.matmul(self.style_layer_reshape[idxn, :, :], self.style_layer_reshape[idxn, :, :].T)
        
        M = input_layer.shape[2] * input_layer.shape[3] #H*W
        N = input_layer.shape[1]  #C
        self.div = M * M * N * N
        # TODO： 计算风格迁移图像和目标风格图像的风格损失
        style_diff = np.sum((self.gram_style - self.gram_input) ** 2)
        loss = style_diff / (4 * input_layer.shape[0] * self.div)
        return loss
    def backward(self, input_layer, style_layer):
        
        bottom_diff = np.zeros([input_layer.shape[0], input_layer.shape[1], input_layer.shape[2]*input_layer.shape[3]])
        for idxn in range(input_layer.shape[0]):
            # TODO： 计算风格损失的反向传播
            bottom_diff[idxn, :, :] = \
                np.matmul((self.gram_input - self.gram_style)[idxn, :, :].T, style_layer.reshape(style_layer.shape[0], style_layer.shape[1], -1)[idxn, :, :]) / (style_layer.shape[0] * self.div)
        bottom_diff = np.reshape(bottom_diff, input_layer.shape)
        return bottom_diff