Mindspore框架：CycleGAN模型实现图像风格迁移算法
- Mindspore框架CycleGAN模型实现图像风格迁移|（一）CycleGAN神经网络模型构建
- Mindspore框架CycleGAN模型实现图像风格迁移|（二）实例数据集（苹果2橘子）
- Mindspore框架CycleGAN模型实现图像风格迁移|（三）损失函数计算
- Mindspore框架CycleGAN模型实现图像风格迁移|（四）CycleGAN模型训练
- Mindspore框架CycleGAN模型实现图像风格迁移|（五）CycleGAN模型推理与资源下载
CycleGAN神经网络模型构建
1. 关于CycleGAN神经网络模型
CycleGAN(Cycle Generative Adversarial Network) 即循环对抗生成网络。
该模型实现了一种在没有配对示例的情况下学习将图像从源域 X 转换到目标域 Y 的方法。实现图像风格迁移。
经典风格迁移模型
Pix2Pix，要求训练数据必须是成对的，而现实生活中，要找到两个域（画风）成对出现的图片是相当困难的。CycleGAN是一种新的无监督的图像迁移网络。更有可行性。
CycleGAN 网络本质上是由两个镜像对称的 GAN 网络组成，其结构如下：
X是一种风格，比如苹果；Y 是另一种风格，比如橘子。
𝐺为将苹果风格X生成橘子风格Y的生成器网络；𝐹为将橘子生成的苹果风格的生成器网络。
𝐷𝑋和 𝐷𝑌为其相应判别器网络。
模型最终能够输出两个模型的权重，分别将两种图像的风格进行彼此迁移，生成新的图像。
2.CycleGAN构建
输入大小为256×256以上的，需要采用9个残差块相连，超参数 n_layers=9 参数控制残差块数。
构建ConvNormReLU，ResidualBlock子网络，搭建ResNetGenerator实现G、F生成器。
- 生成器网络构建
import mindspore.nn as nn
import mindspore.ops as ops
from mindspore.common.initializer import Normal
weight_init = Normal(sigma=0.02)
class ConvNormReLU(nn.Cell):
def __init__(self, input_channel, out_planes, kernel_size=4, stride=2, alpha=0.2, norm_mode='instance',
pad_mode='CONSTANT', use_relu=True, padding=None, transpose=False):
super(ConvNormReLU, self).__init__()
norm = nn.BatchNorm2d(out_planes)
if norm_mode == 'instance':
norm = nn.BatchNorm2d(out_planes, affine=False)
has_bias = (norm_mode == 'instance')
if padding is None:
padding = (kernel_size - 1) // 2
if pad_mode == 'CONSTANT':
if transpose:
conv = nn.Conv2dTranspose(input_channel, out_planes, kernel_size, stride, pad_mode='same',
has_bias=has_bias, weight_init=weight_init)
else:
conv = nn.Conv2d(input_channel, out_planes, kernel_size, stride, pad_mode='pad',
has_bias=has_bias, padding=padding, weight_init=weight_init)
layers = [conv, norm]
else:
paddings = ((0, 0), (0, 0), (padding, padding), (padding, padding))
pad = nn.Pad(paddings=paddings, mode=pad_mode)
if transpose:
conv = nn.Conv2dTranspose(input_channel, out_planes, kernel_size, stride, pad_mode='pad',
has_bias=has_bias, weight_init=weight_init)
else:
conv = nn.Conv2d(input_channel, out_planes, kernel_size, stride, pad_mode='pad',
has_bias=has_bias, weight_init=weight_init)
layers = [pad, conv, norm]
if use_relu:
relu = nn.ReLU()
if alpha > 0:
relu = nn.LeakyReLU(alpha)
layers.append(relu)
self.features = nn.SequentialCell(layers)
def construct(self, x):
output = self.features(x)
return output
class ResidualBlock(nn.Cell):
def __init__(self, dim, norm_mode='instance', dropout=False, pad_mode="CONSTANT"):
super(ResidualBlock, self).__init__()
self.conv1 = ConvNormReLU(dim, dim, 3, 1, 0, norm_mode, pad_mode)
self.conv2 = ConvNormReLU(dim, dim, 3, 1, 0, norm_mode, pad_mode, use_relu=False)
self.dropout = dropout
if dropout:
self.dropout = nn.Dropout(p=0.5)
def construct(self, x):
out = self.conv1(x)
if self.dropout:
out = self.dropout(out)
out = self.conv2(out)
return x + out
class ResNetGenerator(nn.Cell):
def __init__(self, input_channel=3, output_channel=64, n_layers=9, alpha=0.2, norm_mode='instance', dropout=False,
pad_mode="CONSTANT"):
super(ResNetGenerator, self).__init__()
self.conv_in = ConvNormReLU(input_channel, output_channel, 7, 1, alpha, norm_mode, pad_mode=pad_mode)
self.down_1 = ConvNormReLU(output_channel, output_channel * 2, 3, 2, alpha, norm_mode)
self.down_2 = ConvNormReLU(output_channel * 2, output_channel * 4, 3, 2, alpha, norm_mode)
layers = [ResidualBlock(output_channel * 4, norm_mode, dropout=dropout, pad_mode=pad_mode)] * n_layers
self.residuals = nn.SequentialCell(layers)
self.up_2 = ConvNormReLU(output_channel * 4, output_channel * 2, 3, 2, alpha, norm_mode, transpose=True)
self.up_1 = ConvNormReLU(output_channel * 2, output_channel, 3, 2, alpha, norm_mode, transpose=True)
if pad_mode == "CONSTANT":
self.conv_out = nn.Conv2d(output_channel, 3, kernel_size=7, stride=1, pad_mode='pad',
padding=3, weight_init=weight_init)
else:
pad = nn.Pad(paddings=((0, 0), (0, 0), (3, 3), (3, 3)), mode=pad_mode)
conv = nn.Conv2d(output_channel, 3, kernel_size=7, stride=1, pad_mode='pad', weight_init=weight_init)
self.conv_out = nn.SequentialCell([pad, conv])
def construct(self, x):
x = self.conv_in(x)
x = self.down_1(x)
x = self.down_2(x)
x = self.residuals(x)
x = self.up_2(x)
x = self.up_1(x)
output = self.conv_out(x)
return ops.tanh(output)
# 实例化生成器
net_rg_a = ResNetGenerator()
net_rg_a.update_parameters_name('net_rg_a.')
net_rg_b = ResNetGenerator()
net_rg_b.update_parameters_name('net_rg_b.')
- 构建判别器网络
判别器其实是一个二分类网络模型，输出判定该图像为真实图的概率。网络模型使用的是 Patch 大小为 70x70 的 PatchGANs 模型。通过一系列的 Conv2d 、 BatchNorm2d 和 LeakyReLU 层对其进行处理，最后通过 Sigmoid 激活函数得到最终概率。
# 定义判别器
class Discriminator(nn.Cell):
def __init__(self, input_channel=3, output_channel=64, n_layers=3, alpha=0.2, norm_mode='instance'):
super(Discriminator, self).__init__()
kernel_size = 4
layers = [nn.Conv2d(input_channel, output_channel, kernel_size, 2, pad_mode='pad', padding=1, weight_init=weight_init),
nn.LeakyReLU(alpha)]
nf_mult = output_channel
for i in range(1, n_layers):
nf_mult_prev = nf_mult
nf_mult = min(2 ** i, 8) * output_channel
layers.append(ConvNormReLU(nf_mult_prev, nf_mult, kernel_size, 2, alpha, norm_mode, padding=1))
nf_mult_prev = nf_mult
nf_mult = min(2 ** n_layers, 8) * output_channel
layers.append(ConvNormReLU(nf_mult_prev, nf_mult, kernel_size, 1, alpha, norm_mode, padding=1))
layers.append(nn.Conv2d(nf_mult, 1, kernel_size, 1, pad_mode='pad', padding=1, weight_init=weight_init))
self.features = nn.SequentialCell(layers)
def construct(self, x):
output = self.features(x)
return output
# 判别器初始化
net_d_a = Discriminator()
net_d_a.update_parameters_name('net_d_a.')
net_d_b = Discriminator()
net_d_b.update_parameters_name('net_d_b.')
