昇思25天学习打卡营第7天|Pix2Pix实现图像转换

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Pix2Pix图像转换

1、Pix2Pix 概述

Pix2Pix 是一个专门为图像到图像的转换任务设计的网络，可以实现语义/标签到真实图片、灰度图到彩色图、航空图到地图、白天到黑夜、线稿图到实物图的转换。Pix2Pix是将条件GAN（CGAN）应用于有监督（需要成对的输入素描图像Sketch和真实图像GT，来训练网络）的图像到图像翻译的经典之作，和所有的GANs一样，模型同样包括：生成器和判别器两个部分。

CGAN：CGAN(条件GAN) 的目标是生成与给定条件匹配的数据样本。这些条件可以是标签、部分实例标注数据或任何其他形式的多模态辅助信息。CGAN 通过将条件并入网络的生成器和判别器中来指导数据生成过程。

相比普通的生成对抗损失：
$$L_{GAN}(G,D)={\mathbb{E}}_y[log(D(y))]+{\mathbb{E}}_{(x,z)}[log(1-D(x,z))]$$

• $x$：代表观测图像的数据。
• $z$：代表随机噪声的数据。
• $y=G(x,z)$：生成器网络，给出由观测图像$x$与随机噪声$z$生成的“假”图片，其中$x$来自于训练数据而非生成器。
• $D(x,G(x,z))$：判别器网络，给出图像判定为真实图像的概率，其中$x$来自于训练数据，$G(x,z)$来自于生成器。

CGAN多了来自于观测图像的条件$x$（因此Pix2Pix训练时采用有监督的方式，需要标注好的语义数据，如下图中的
Map2Aerial数据集、Anime Sketch Colorization Pair 素描生成动漫数据集），

CGAN的目标可以表示为：

$$L_{CGAN}(G,D)={\mathbb{E}}_{(x,y)}[log(D(x,y))]+{\mathbb{E}}_{(x,z)}[log(1-D(x,G(x,z)))]$$

Pix2Pix 还包括 L1 损失，帮助生成器产生结构上接近真实图像的结果，这一点在图像翻译任务中尤为重要：
$$L_{L1}(G)={\mathbb{E}}_{(x,y,z)}[||y-G(x,z)|{|}_1]$$

进而得到最终目标：

$$arg\underset{G}{\min }\underset{D}{\max }{L}_{CGAN}(G,D)+\lambda L_{L1}(G)$$

图像转换问题本质上其实就是像素到像素的映射问题，Pix2Pix使用完全一样的网络结构和目标函数，仅更换不同的训练数据集就能分别实现以上的任务。

2、U-Net架构

U-Net架构：Pix2Pix 使用 U-Net 架构作为其生成器，在传统的编解码网络结构基础上加入了跳跃连接的方式，可以更好地捕捉图像的细节和上下文信息，适合于图像到图像的翻译任务。相比于普通的编解码结构（Encoder-Decoder），U-Net在编码器和解码器之间引入了跳跃连接，极大地改善了梯度流：

定义UNet Skip Connection Block

import mindspore
import mindspore.nn as nn
import mindspore.ops as opsclass UNetSkipConnectionBlock(nn.Cell):def __init__(self, outer_nc, inner_nc, in_planes=None, dropout=False,submodule=None, outermost=False, innermost=False, alpha=0.2, norm_mode='batch'):super(UNetSkipConnectionBlock, self).__init__()down_norm = nn.BatchNorm2d(inner_nc)up_norm = nn.BatchNorm2d(outer_nc)use_bias = Falseif norm_mode == 'instance':down_norm = nn.BatchNorm2d(inner_nc, affine=False)up_norm = nn.BatchNorm2d(outer_nc, affine=False)use_bias = Trueif in_planes is None:in_planes = outer_ncdown_conv = nn.Conv2d(in_planes, inner_nc, kernel_size=4,stride=2, padding=1, has_bias=use_bias, pad_mode='pad')down_relu = nn.LeakyReLU(alpha)up_relu = nn.ReLU()if outermost:up_conv = nn.Conv2dTranspose(inner_nc * 2, outer_nc,kernel_size=4, stride=2,padding=1, pad_mode='pad')down = [down_conv]up = [up_relu, up_conv, nn.Tanh()]model = down + [submodule] + upelif innermost:up_conv = nn.Conv2dTranspose(inner_nc, outer_nc,kernel_size=4, stride=2,padding=1, has_bias=use_bias, pad_mode='pad')down = [down_relu, down_conv]up = [up_relu, up_conv, up_norm]model = down + upelse:up_conv = nn.Conv2dTranspose(inner_nc * 2, outer_nc,kernel_size=4, stride=2,padding=1, has_bias=use_bias, pad_mode='pad')down = [down_relu, down_conv, down_norm]up = [up_relu, up_conv, up_norm]model = down + [submodule] + upif dropout:model.append(nn.Dropout(p=0.5))self.model = nn.SequentialCell(model)self.skip_connections = not outermostdef construct(self, x):out = self.model(x)if self.skip_connections:out = ops.concat((out, x), axis=1)return out

2、生成器部分

原始CGAN的输入是条件x和噪声z两种信息，这里的生成器只使用了条件信息，因此不能生成多样性的结果。因此Pix2Pix在训练和测试时都使用了dropout，这样可以生成多样性的结果。

通过MindSpore实现基于U-Net的生成器：

class UNetGenerator(nn.Cell):def __init__(self, in_planes, out_planes, ngf=64, n_layers=8, norm_mode='bn', dropout=False):super(UNetGenerator, self).__init__()unet_block = UNetSkipConnectionBlock(ngf * 8, ngf * 8, in_planes=None, submodule=None,norm_mode=norm_mode, innermost=True)for _ in range(n_layers - 5):unet_block = UNetSkipConnectionBlock(ngf * 8, ngf * 8, in_planes=None, submodule=unet_block,norm_mode=norm_mode, dropout=dropout)unet_block = UNetSkipConnectionBlock(ngf * 4, ngf * 8, in_planes=None, submodule=unet_block,norm_mode=norm_mode)unet_block = UNetSkipConnectionBlock(ngf * 2, ngf * 4, in_planes=None, submodule=unet_block,norm_mode=norm_mode)unet_block = UNetSkipConnectionBlock(ngf, ngf * 2, in_planes=None, submodule=unet_block,norm_mode=norm_mode)self.model = UNetSkipConnectionBlock(out_planes, ngf, in_planes=in_planes, submodule=unet_block,outermost=True, norm_mode=norm_mode)def construct(self, x):return self.model(x)

3、基于PatchGAN的判别器

判别器使用的PatchGAN结构，可看做卷积。
生成的矩阵中的每个点代表原图的一小块区域（patch）。通过矩阵中的各个值来判断原图中对应每个Patch的真假。

import mindspore.nn as nnclass ConvNormRelu(nn.Cell):def __init__(self,in_planes,out_planes,kernel_size=4,stride=2,alpha=0.2,norm_mode='batch',pad_mode='CONSTANT',use_relu=True,padding=None):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 not padding:padding = (kernel_size - 1) // 2if pad_mode == 'CONSTANT':conv = nn.Conv2d(in_planes, out_planes, kernel_size, stride, pad_mode='pad',has_bias=has_bias, padding=padding)layers = [conv, norm]else:paddings = ((0, 0), (0, 0), (padding, padding), (padding, padding))pad = nn.Pad(paddings=paddings, mode=pad_mode)conv = nn.Conv2d(in_planes, out_planes, kernel_size, stride, pad_mode='pad', has_bias=has_bias)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 outputclass Discriminator(nn.Cell):def __init__(self, in_planes=3, ndf=64, n_layers=3, alpha=0.2, norm_mode='batch'):super(Discriminator, self).__init__()kernel_size = 4layers = [nn.Conv2d(in_planes, ndf, kernel_size, 2, pad_mode='pad', padding=1),nn.LeakyReLU(alpha)]nf_mult = ndffor i in range(1, n_layers):nf_mult_prev = nf_multnf_mult = min(2 ** i, 8) * ndflayers.append(ConvNormRelu(nf_mult_prev, nf_mult, kernel_size, 2, alpha, norm_mode, padding=1))nf_mult_prev = nf_multnf_mult = min(2 ** n_layers, 8) * ndflayers.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))self.features = nn.SequentialCell(layers)def construct(self, x, y):x_y = ops.concat((x, y), axis=1)output = self.features(x_y)return output

4、Pix2Pix的生成器和判别器初始化

实例化Pix2Pix生成器和判别器：

import mindspore.nn as nn
from mindspore.common import initializer as initg_in_planes = 3
g_out_planes = 3
g_ngf = 64
g_layers = 8
d_in_planes = 6
d_ndf = 64
d_layers = 3
alpha = 0.2
init_gain = 0.02
init_type = 'normal'net_generator = UNetGenerator(in_planes=g_in_planes, out_planes=g_out_planes,ngf=g_ngf, n_layers=g_layers)
for _, cell in net_generator.cells_and_names():if isinstance(cell, (nn.Conv2d, nn.Conv2dTranspose)):if init_type == 'normal':cell.weight.set_data(init.initializer(init.Normal(init_gain), cell.weight.shape))elif init_type == 'xavier':cell.weight.set_data(init.initializer(init.XavierUniform(init_gain), cell.weight.shape))elif init_type == 'constant':cell.weight.set_data(init.initializer(0.001, cell.weight.shape))else:raise NotImplementedError('initialization method [%s] is not implemented' % init_type)elif isinstance(cell, nn.BatchNorm2d):cell.gamma.set_data(init.initializer('ones', cell.gamma.shape))cell.beta.set_data(init.initializer('zeros', cell.beta.shape))net_discriminator = Discriminator(in_planes=d_in_planes, ndf=d_ndf,alpha=alpha, n_layers=d_layers)
for _, cell in net_discriminator.cells_and_names():if isinstance(cell, (nn.Conv2d, nn.Conv2dTranspose)):if init_type == 'normal':cell.weight.set_data(init.initializer(init.Normal(init_gain), cell.weight.shape))elif init_type == 'xavier':cell.weight.set_data(init.initializer(init.XavierUniform(init_gain), cell.weight.shape))elif init_type == 'constant':cell.weight.set_data(init.initializer(0.001, cell.weight.shape))else:raise NotImplementedError('initialization method [%s] is not implemented' % init_type)elif isinstance(cell, nn.BatchNorm2d):cell.gamma.set_data(init.initializer('ones', cell.gamma.shape))cell.beta.set_data(init.initializer('zeros', cell.beta.shape))class Pix2Pix(nn.Cell):"""Pix2Pix模型网络"""def __init__(self, discriminator, generator):super(Pix2Pix, self).__init__(auto_prefix=True)self.net_discriminator = discriminatorself.net_generator = generatordef construct(self, reala):fakeb = self.net_generator(reala)return fakeb

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；
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；
训练生成器是希望能产生更好的虚假图像；
在这两个部分中，分别获取训练过程中的损失，并在每个周期结束时进行统计。

通过MindSpore进行训练：

import numpy as np
import os
import datetime
from mindspore import value_and_grad, Tensorepoch_num = 3
ckpt_dir = "results/ckpt"
dataset_size = 400
val_pic_size = 256
lr = 0.0002
n_epochs = 100
n_epochs_decay = 100def get_lr():lrs = [lr] * dataset_size * n_epochslr_epoch = 0for epoch in range(n_epochs_decay):lr_epoch = lr * (n_epochs_decay - epoch) / n_epochs_decaylrs += [lr_epoch] * dataset_sizelrs += [lr_epoch] * dataset_size * (epoch_num - n_epochs_decay - n_epochs)return Tensor(np.array(lrs).astype(np.float32))dataset = ds.MindDataset("./dataset/dataset_pix2pix/train.mindrecord", columns_list=["input_images", "target_images"], shuffle=True, num_parallel_workers=1)
steps_per_epoch = dataset.get_dataset_size()
loss_f = nn.BCEWithLogitsLoss()
l1_loss = nn.L1Loss()def forword_dis(reala, realb):lambda_dis = 0.5fakeb = net_generator(reala)pred0 = net_discriminator(reala, fakeb)pred1 = net_discriminator(reala, realb)loss_d = loss_f(pred1, ops.ones_like(pred1)) + loss_f(pred0, ops.zeros_like(pred0))loss_dis = loss_d * lambda_disreturn loss_disdef forword_gan(reala, realb):lambda_gan = 0.5lambda_l1 = 100fakeb = net_generator(reala)pred0 = net_discriminator(reala, fakeb)loss_1 = loss_f(pred0, ops.ones_like(pred0))loss_2 = l1_loss(fakeb, realb)loss_gan = loss_1 * lambda_gan + loss_2 * lambda_l1return loss_gand_opt = nn.Adam(net_discriminator.trainable_params(), learning_rate=get_lr(),beta1=0.5, beta2=0.999, loss_scale=1)
g_opt = nn.Adam(net_generator.trainable_params(), learning_rate=get_lr(),beta1=0.5, beta2=0.999, loss_scale=1)grad_d = value_and_grad(forword_dis, None, net_discriminator.trainable_params())
grad_g = value_and_grad(forword_gan, None, net_generator.trainable_params())def train_step(reala, realb):loss_dis, d_grads = grad_d(reala, realb)loss_gan, g_grads = grad_g(reala, realb)d_opt(d_grads)g_opt(g_grads)return loss_dis, loss_ganif not os.path.isdir(ckpt_dir):os.makedirs(ckpt_dir)g_losses = []
d_losses = []
data_loader = dataset.create_dict_iterator(output_numpy=True, num_epochs=epoch_num)for epoch in range(epoch_num):for i, data in enumerate(data_loader):start_time = datetime.datetime.now()input_image = Tensor(data["input_images"])target_image = Tensor(data["target_images"])dis_loss, gen_loss = train_step(input_image, target_image)end_time = datetime.datetime.now()delta = (end_time - start_time).microsecondsif i % 2 == 0:print("ms per step:{:.2f}  epoch:{}/{}  step:{}/{}  Dloss:{:.4f}  Gloss:{:.4f} ".format((delta / 1000), (epoch + 1), (epoch_num), i, steps_per_epoch, float(dis_loss), float(gen_loss)))d_losses.append(dis_loss.asnumpy())g_losses.append(gen_loss.asnumpy())if (epoch + 1) == epoch_num:mindspore.save_checkpoint(net_generator, ckpt_dir + "Generator.ckpt")

6、模型推理

导入模型训练保存的权重：

from mindspore import load_checkpoint, load_param_into_netparam_g = load_checkpoint(ckpt_dir + "Generator.ckpt")
load_param_into_net(net_generator, param_g)
dataset = ds.MindDataset("./dataset/dataset_pix2pix/train.mindrecord", columns_list=["input_images", "target_images"], shuffle=True)
data_iter = next(dataset.create_dict_iterator())
predict_show = net_generator(data_iter["input_images"])
plt.figure(figsize=(10, 3), dpi=140)
for i in range(10):plt.subplot(2, 10, i + 1)plt.imshow((data_iter["input_images"][i].asnumpy().transpose(1, 2, 0) + 1) / 2)plt.axis("off")plt.subplots_adjust(wspace=0.05, hspace=0.02)plt.subplot(2, 10, i + 11)plt.imshow((predict_show[i].asnumpy().transpose(1, 2, 0) + 1) / 2)plt.axis("off")plt.subplots_adjust(wspace=0.05, hspace=0.02)
plt.show()

图像翻译效果如下：

Reference

昇思官方文档-Pix2Pix实现图像转换
昇思大模型平台
AI 助你无码看片，生成对抗网络（GAN）大显身手