文章目录
- 昇思MindSpore应用实践
- 基于MindSpore的CycleGAN图像风格迁移互换
- 1、CycleGAN 概述
- 2、生成器部分
- 3、判别器部分
- 4、优化器和损失函数
- 5、模型训练
- 6、模型推理
- Reference
昇思MindSpore应用实践
本系列文章主要用于记录昇思25天学习打卡营的学习心得。
基于MindSpore的CycleGAN图像风格迁移互换
1、CycleGAN 概述
CycleGAN(Cycle Generative Adversarial Network) 即循环对抗生成网络,作者同样是 Jun-Yan Zhu,两篇文章都分别获得了高达2w+的citation.
CycleGAN 网络本质上是由两个镜像对称的 GAN 网络组成,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。
该模型一个重要应用领域是域迁移(Domain Adaptation),可以通俗地理解为图像风格迁移。其实在 CycleGAN 之前,就已经有了域迁移模型,比如 Pix2Pix ,但是 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)中成对出现的图片是相当困难的,而 CycleGAN 只需要两种域的数据,且不需要像Pix2Pix2那样有严格的对应关系,是一种新的无监督的图像迁移网络。
为了方便理解,这里以苹果和橘子为例介绍。上图中 X X X 可以理解为苹果, Y Y Y 为橘子; G G G 为将苹果生成橘子风格的生成器, F F F 为将橘子生成的苹果风格的生成器, D X D_{X} DX 和 D Y D_{Y} DY 为其相应判别器,具体生成器和判别器的结构可见下文代码。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。
该模型一个很重要的部分就是损失函数,在所有损失里面循环一致损失(Cycle Consistency Loss)
是最重要的。
循环损失的计算过程如下图所示:
图中苹果图片 x x x 经过橘子风格生成器 G G G 得到伪橘子 Y ^ \hat{Y} Y^,然后将伪橘子 Y ^ \hat{Y} Y^ 结果送进苹果风格生成器 F F F 又产生苹果风格的结果 x ^ \hat{x} x^,最后将生成的苹果风格结果 x ^ \hat{x} x^ 与原苹果图片 x x x 一起计算出循环一致损失,反之亦然。循环损失捕捉了这样的直觉,即如果我们从一个域转换到另一个域,然后再转换回原始域时,输出图像应尽可能接近原始输入图像。这种损失的引入帮助CycleGAN学习到在没有成对训练样本的情况下(即无监督迁移学习)也能实现有效的域转换,通过确保信息的双向循环来平衡风格转换与内容保持,从而使风格转换在追求风格表达的同时,不牺牲图像的原始属性和识别性。
2、生成器部分
本案例生成器的模型结构参考的 ResNet 模型的结构,参考原论文,对于128×128大小的输入图片采用6个残差块相连,图片大小为256×256以上的需要采用9个残差块相连,所以本文网络有9个残差块相连,超参数 n_layers
参数控制残差块数。
生成器的结构如下所示:
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.')
3、判别器部分
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.')
4、优化器和损失函数
对生成器 G G G 及其判别器 D Y D_{Y} DY ,目标损失函数定义为:
L G A N ( G , D Y , X , Y ) = E y − p d a t a ( y ) [ l o g D Y ( y ) ] + E x − p d a t a ( x ) [ l o g ( 1 − D Y ( G ( x ) ) ) ] L_{GAN}(G,D_Y,X,Y)=E_{y-p_{data}(y)}[logD_Y(y)]+E_{x-p_{data}(x)}[log(1-D_Y(G(x)))] LGAN(G,DY,X,Y)=Ey−pdata(y)[logDY(y)]+Ex−pdata(x)[log(1−DY(G(x)))]
其中 G G G 试图生成看起来与 Y Y Y 中的图像相似的图像 G ( x ) G(x) G(x) ,而 D Y D_{Y} DY 的目标是区分翻译样本 G ( x ) G(x) G(x) 和真实样本 y y y ,生成器的目标是最小化翻译样本与真实样本之间的损失函数值以此来对抗判别器,判别器的目的是最大化生成的虚假翻译样本与真实样本之间的损失函数值,以此训练鉴别能力。即 m i n G m a x D Y L G A N ( G , D Y , X , Y ) min_{G} max_{D_{Y}}L_{GAN}(G,D_{Y} ,X,Y ) minGmaxDYLGAN(G,DY,X,Y) ;
单独的对抗损失不能保证所学函数可以将单个输入映射到期望的输出,为了进一步减少可能的映射函数的空间,学习到的映射函数应该是周期一致的,例如对于 X X X 的每个图像 x x x ,图像转换周期应能够将 x x x 带回原始图像,可以称之为正向循环一致性,即 x → G ( x ) → F ( G ( x ) ) ≈ x x→G(x)→F(G(x))\approx x x→G(x)→F(G(x))≈x 。对于 Y Y Y ,类似的 x → G ( x ) → F ( G ( x ) ) ≈ x x→G(x)→F(G(x))\approx x x→G(x)→F(G(x))≈x 。可以理解采用了一个循环一致性损失来激励这种行为。
循环一致损失函数定义如下:
L c y c ( G , F ) = E x − p d a t a ( x ) [ ∥ F ( G ( x ) ) − x ∥ 1 ] + E y − p d a t a ( y ) [ ∥ G ( F ( y ) ) − y ∥ 1 ] L_{cyc}(G,F)=E_{x-p_{data}(x)}[\Vert F(G(x))-x\Vert_{1}]+E_{y-p_{data}(y)}[\Vert G(F(y))-y\Vert_{1}] Lcyc(G,F)=Ex−pdata(x)[∥F(G(x))−x∥1]+Ey−pdata(y)[∥G(F(y))−y∥1]
循环一致损失能够保证重建图像 F ( G ( x ) ) F(G(x)) F(G(x)) 与输入图像 x x x 紧密匹配。
# 构建生成器,判别器优化器
optimizer_rg_a = nn.Adam(net_rg_a.trainable_params(), learning_rate=0.0002, beta1=0.5)
optimizer_rg_b = nn.Adam(net_rg_b.trainable_params(), learning_rate=0.0002, beta1=0.5)
optimizer_d_a = nn.Adam(net_d_a.trainable_params(), learning_rate=0.0002, beta1=0.5)
optimizer_d_b = nn.Adam(net_d_b.trainable_params(), learning_rate=0.0002, beta1=0.5)
# GAN网络损失函数,这里最后一层不使用sigmoid函数
loss_fn = nn.MSELoss(reduction='mean')
l1_loss = nn.L1Loss("mean")
def gan_loss(predict, target):
target = ops.ones_like(predict) * target
loss = loss_fn(predict, target)
return loss
CycleGAN模型的特点是为了减少模型振荡[1],这里遵循 Shrivastava 等人的策略[2],使用生成器生成图像的历史数据而不是生成器生成的最新图像数据来更新鉴别器。这里创建 image_pool
函数,保留了一个图像缓冲区,用于存储生成器生成前的50个图像。
import mindspore as ms
# 前向计算
def generator(img_a, img_b):
fake_a = net_rg_b(img_b)
fake_b = net_rg_a(img_a)
rec_a = net_rg_b(fake_b)
rec_b = net_rg_a(fake_a)
identity_a = net_rg_b(img_a)
identity_b = net_rg_a(img_b)
return fake_a, fake_b, rec_a, rec_b, identity_a, identity_b
lambda_a = 10.0
lambda_b = 10.0
lambda_idt = 0.5
def generator_forward(img_a, img_b):
true = Tensor(True, dtype.bool_)
fake_a, fake_b, rec_a, rec_b, identity_a, identity_b = generator(img_a, img_b)
loss_g_a = gan_loss(net_d_b(fake_b), true)
loss_g_b = gan_loss(net_d_a(fake_a), true)
loss_c_a = l1_loss(rec_a, img_a) * lambda_a
loss_c_b = l1_loss(rec_b, img_b) * lambda_b
loss_idt_a = l1_loss(identity_a, img_a) * lambda_a * lambda_idt
loss_idt_b = l1_loss(identity_b, img_b) * lambda_b * lambda_idt
loss_g = loss_g_a + loss_g_b + loss_c_a + loss_c_b + loss_idt_a + loss_idt_b
return fake_a, fake_b, loss_g, loss_g_a, loss_g_b, loss_c_a, loss_c_b, loss_idt_a, loss_idt_b
def generator_forward_grad(img_a, img_b):
_, _, loss_g, _, _, _, _, _, _ = generator_forward(img_a, img_b)
return loss_g
def discriminator_forward(img_a, img_b, fake_a, fake_b):
false = Tensor(False, dtype.bool_)
true = Tensor(True, dtype.bool_)
d_fake_a = net_d_a(fake_a)
d_img_a = net_d_a(img_a)
d_fake_b = net_d_b(fake_b)
d_img_b = net_d_b(img_b)
loss_d_a = gan_loss(d_fake_a, false) + gan_loss(d_img_a, true)
loss_d_b = gan_loss(d_fake_b, false) + gan_loss(d_img_b, true)
loss_d = (loss_d_a + loss_d_b) * 0.5
return loss_d
def discriminator_forward_a(img_a, fake_a):
false = Tensor(False, dtype.bool_)
true = Tensor(True, dtype.bool_)
d_fake_a = net_d_a(fake_a)
d_img_a = net_d_a(img_a)
loss_d_a = gan_loss(d_fake_a, false) + gan_loss(d_img_a, true)
return loss_d_a
def discriminator_forward_b(img_b, fake_b):
false = Tensor(False, dtype.bool_)
true = Tensor(True, dtype.bool_)
d_fake_b = net_d_b(fake_b)
d_img_b = net_d_b(img_b)
loss_d_b = gan_loss(d_fake_b, false) + gan_loss(d_img_b, true)
return loss_d_b
# 保留了一个图像缓冲区,用来存储之前创建的50个图像
pool_size = 50
def image_pool(images):
num_imgs = 0
image1 = []
if isinstance(images, Tensor):
images = images.asnumpy()
return_images = []
for image in images:
if num_imgs < pool_size:
num_imgs = num_imgs + 1
image1.append(image)
return_images.append(image)
else:
if random.uniform(0, 1) > 0.5:
random_id = random.randint(0, pool_size - 1)
tmp = image1[random_id].copy()
image1[random_id] = image
return_images.append(tmp)
else:
return_images.append(image)
output = Tensor(return_images, ms.float32)
if output.ndim != 4:
raise ValueError("img should be 4d, but get shape {}".format(output.shape))
return output
5、模型训练
训练分为两个主要部分:训练判别器和训练生成器,采用了最小二乘损失代替负对数似然目标。
-
训练判别器:训练判别器的目的是最大程度地提高判别图像真伪的概率。需要训练判别器来最小化 E y − p d a t a ( y ) [ ( D ( y ) − 1 ) 2 ] E_{y-p_{data}(y)}[(D(y)-1)^2] Ey−pdata(y)[(D(y)−1)2] ;
-
训练生成器:如 CycleGAN 论文所述,我们希望通过最小化 E x − p d a t a ( x ) [ ( D ( G ( x ) − 1 ) 2 ] E_{x-p_{data}(x)}[(D(G(x)-1)^2] Ex−pdata(x)[(D(G(x)−1)2] 来训练生成器,以产生更好的虚假图像。
import os
import time
import random
import numpy as np
from PIL import Image
from mindspore import Tensor, save_checkpoint
from mindspore import dtype
# 由于时间原因,epochs设置为1,可根据需求进行调整,实验结果表明epoch=1也能实现基本的风格转换
epochs = 10
save_step_num = 80
save_checkpoint_epochs = 1
save_ckpt_dir = './train_ckpt_outputs/'
print('Start training!')
for epoch in range(epochs):
g_loss = []
d_loss = []
start_time_e = time.time()
for step, data in enumerate(dataset.create_dict_iterator()):
start_time_s = time.time()
img_a = data["image_A"]
img_b = data["image_B"]
res_g = train_step_g(img_a, img_b)
fake_a = res_g[0]
fake_b = res_g[1]
res_d = train_step_d(img_a, img_b, image_pool(fake_a), image_pool(fake_b))
loss_d = float(res_d.asnumpy())
step_time = time.time() - start_time_s
res = []
for item in res_g[2:]:
res.append(float(item.asnumpy()))
g_loss.append(res[0])
d_loss.append(loss_d)
if step % save_step_num == 0:
print(f"Epoch:[{int(epoch + 1):>3d}/{int(epochs):>3d}], "
f"step:[{int(step):>4d}/{int(datasize):>4d}], "
f"time:{step_time:>3f}s,\n"
f"loss_g:{res[0]:.2f}, loss_d:{loss_d:.2f}, "
f"loss_g_a: {res[1]:.2f}, loss_g_b: {res[2]:.2f}, "
f"loss_c_a: {res[3]:.2f}, loss_c_b: {res[4]:.2f}, "
f"loss_idt_a: {res[5]:.2f}, loss_idt_b: {res[6]:.2f}")
epoch_cost = time.time() - start_time_e
per_step_time = epoch_cost / datasize
mean_loss_d, mean_loss_g = sum(d_loss) / datasize, sum(g_loss) / datasize
print(f"Epoch:[{int(epoch + 1):>3d}/{int(epochs):>3d}], "
f"epoch time:{epoch_cost:.2f}s, per step time:{per_step_time:.2f}, "
f"mean_g_loss:{mean_loss_g:.2f}, mean_d_loss:{mean_loss_d :.2f}")
if epoch % save_checkpoint_epochs == 0:
os.makedirs(save_ckpt_dir, exist_ok=True)
save_checkpoint(net_rg_a, os.path.join(save_ckpt_dir, f"g_a_{epoch}.ckpt"))
save_checkpoint(net_rg_b, os.path.join(save_ckpt_dir, f"g_b_{epoch}.ckpt"))
save_checkpoint(net_d_a, os.path.join(save_ckpt_dir, f"d_a_{epoch}.ckpt"))
save_checkpoint(net_d_b, os.path.join(save_ckpt_dir, f"d_b_{epoch}.ckpt"))
print('End of training!')
6、模型推理
Reference
[1] I. Goodfellow. NIPS 2016 tutorial: Generative ad-versarial networks. arXiv preprint arXiv:1701.00160,2016. 2, 4, 5
[2] A. Shrivastava, T. Pfister, O. Tuzel, J. Susskind, W. Wang, R. Webb. Learning from simulated and unsupervised images through adversarial training. In CVPR, 2017. 3, 5, 6, 7
[3] 昇思大模型平台
[4] 昇思官方文档-CycleGAN图像风格迁移互换