SCRN中间层的特征图可视化

# todo 特征图可视化
            # 0： 展示特征的形状
            print('=e21', x_e21.shape)
            # == e21            torch.Size([1, 32, 88, 88])
            # 1: 去除多余层[1, 32, 88, 88] => [32, 88, 88]
            x_e21 = x_e21.squeeze()
            # 2: 取出32通道的第一个通道
            x_e21 = x_e21[0]
            # 3: 归一化：用sigmoid函数将(-1,1)范围变为 => (0,1)
            x_e21 = np.squeeze(torch.sigmoid(x_e21).cpu().data.numpy())
            # 4: 将归一化的结果扩展为RGB的255 (0,1) => (0,255)
            x_e21 = 255 * x_e21
            # 5: 保存图片
            misc.imsave(save_path + name + '_xxxe21.png', x_e21)

遵照上面原理
下面可视化e21 e22 s21 s22

 # 1: 去除多余层[1, 32, 88, 88] => [32, 88, 88]
            x_e21 = x_e21.squeeze()
            x_e22 = x_e22.squeeze()
            x_s21 = x_s21.squeeze()
            x_s22 = x_s22.squeeze()

            # 2: 取出32通道的第一个通道
            x_e21 = x_e21[0]
            x_e22 = x_e22[0]
            x_s21 = x_s21[0]
            x_s22 = x_s22[0]

            # 3: 归一化：用sigmoid函数将(-1,1)范围变为 => (0,1)
            x_e21 = np.squeeze(torch.sigmoid(x_e21).cpu().data.numpy())
            x_e22 = np.squeeze(torch.sigmoid(x_e22).cpu().data.numpy())
            x_s21 = np.squeeze(torch.sigmoid(x_s21).cpu().data.numpy())
            x_s22 = np.squeeze(torch.sigmoid(x_s22).cpu().data.numpy())
            # 4: 将归一化的结果扩展为RGB的255 (0,1) => (0,255)
            x_e21 = 255 * x_e21
            x_e22 = 255 * x_e22
            x_s21 = 255 * x_s21
            x_s22 = 255 * x_s22

            # 5: 保存图片
            misc.imsave(save_path + name + '=e21.png', x_e21)
            misc.imsave(save_path + name + '=e22.png', x_e22)
            misc.imsave(save_path + name + '=s21.png', x_s21)
            misc.imsave(save_path + name + '=s22.png', x_s22)

8特征实用例子
 

# todo 特征图可视化
            # 0： 展示特征的形状
            print('=e21', x_e21.shape)

            # == e21            torch.Size([1, 32, 88, 88])

            # 1: 去除多余层[1, 32, 88, 88] => [32, 88, 88]
            x_e21 = x_e21.squeeze()
            x_e22 = x_e22.squeeze()
            x_e23 = x_e23.squeeze()
            x_e24 = x_e24.squeeze()
            x_s21 = x_s21.squeeze()
            x_s22 = x_s22.squeeze()
            x_s23 = x_s23.squeeze()
            x_s24 = x_s24.squeeze()

            # 2: 取出32通道的第一个通道
            x_e21 = x_e21[0]
            x_e22 = x_e22[0]
            x_e23 = x_e23[0]
            x_e24 = x_e24[0]
            x_s21 = x_s21[0]
            x_s22 = x_s22[0]
            x_s23 = x_s23[0]
            x_s24 = x_s24[0]

            # 3: 归一化：用sigmoid函数将(-1,1)范围变为 => (0,1)
            x_e21 = np.squeeze(torch.sigmoid(x_e21).cpu().data.numpy())
            x_e22 = np.squeeze(torch.sigmoid(x_e22).cpu().data.numpy())
            x_e23 = np.squeeze(torch.sigmoid(x_e23).cpu().data.numpy())
            x_e24 = np.squeeze(torch.sigmoid(x_e24).cpu().data.numpy())
            x_s21 = np.squeeze(torch.sigmoid(x_s21).cpu().data.numpy())
            x_s22 = np.squeeze(torch.sigmoid(x_s22).cpu().data.numpy())
            x_s23 = np.squeeze(torch.sigmoid(x_s23).cpu().data.numpy())
            x_s24 = np.squeeze(torch.sigmoid(x_s24).cpu().data.numpy())
            # 4: 将归一化的结果扩展为RGB的255 (0,1) => (0,255)
            x_e21 = 255 * x_e21
            x_e22 = 255 * x_e22
            x_e23 = 255 * x_e23
            x_e24 = 255 * x_e24
            x_s21 = 255 * x_s21
            x_s22 = 255 * x_s22
            x_s23 = 255 * x_s23
            x_s24 = 255 * x_s24

            # 5: 保存图片
            misc.imsave(save_path + name + '=e21.png', x_e21)
            misc.imsave(save_path + name + '=e22.png', x_e22)
            misc.imsave(save_path + name + '=e23.png', x_e23)
            misc.imsave(save_path + name + '=e24.png', x_e24)
            misc.imsave(save_path + name + '=s21.png', x_s21)
            misc.imsave(save_path + name + '=s22.png', x_s22)
            misc.imsave(save_path + name + '=s23.png', x_s23)
            misc.imsave(save_path + name + '=s24.png', x_s24)

1 原版SCRN

2 (这里中间层监督出了问题es对掉)

(中间层监督es正常)