昇思25天学习打卡营第17天|munger85

CycleGAN图像风格迁移互换

有了这个网络，可以无监督将源图片的风格应用于目的图片，不需要数据是对称的。

上图中 𝑋
可以理解为苹果， 𝑌
为橘子； 𝐺
为将苹果生成橘子风格的生成器， 𝐹
为将橘子生成的苹果风格的生成器， 𝐷𝑋
和 𝐷𝑌
为其相应判别器
from download import download

url = “https://mindspore-website.obs.cn-north-4.myhuaweicloud.com/notebook/models/application/CycleGAN_apple2orange.zip”

download(url, “.”, kind=“zip”, replace=True)
上面是苹果和橘子的图片，训练。

1019行数据，批大小为1

看看数据是什么样

最前面的图里有结构，所以根据那个结构，要先创建源和目的的模型

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

# 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.')

要训练，肯定要优化器和损失函数

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，不知道为什么。

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

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

上面就是前向传播过程

下面是反向传播，判别器也是在这个过程里起作用

from mindspore import value_and_grad

# 实例化求梯度的方法
grad_g_a = value_and_grad(generator_forward_grad, None, net_rg_a.trainable_params())
grad_g_b = value_and_grad(generator_forward_grad, None, net_rg_b.trainable_params())

grad_d_a = value_and_grad(discriminator_forward_a, None, net_d_a.trainable_params())
grad_d_b = value_and_grad(discriminator_forward_b, None, net_d_b.trainable_params())

# 计算生成器的梯度，反向传播更新参数
def train_step_g(img_a, img_b):
    net_d_a.set_grad(False)
    net_d_b.set_grad(False)

    fake_a, fake_b, lg, lga, lgb, lca, lcb, lia, lib = generator_forward(img_a, img_b)

    _, grads_g_a = grad_g_a(img_a, img_b)
    _, grads_g_b = grad_g_b(img_a, img_b)
    optimizer_rg_a(grads_g_a)
    optimizer_rg_b(grads_g_b)

    return fake_a, fake_b, lg, lga, lgb, lca, lcb, lia, lib

# 计算判别器的梯度，反向传播更新参数
def train_step_d(img_a, img_b, fake_a, fake_b):
    net_d_a.set_grad(True)
    net_d_b.set_grad(True)

    loss_d_a, grads_d_a = grad_d_a(img_a, fake_a)
    loss_d_b, grads_d_b = grad_d_b(img_b, fake_b)

    loss_d = (loss_d_a + loss_d_b) * 0.5

    optimizer_d_a(grads_d_a)
    optimizer_d_b(grads_d_b)

    return loss_d

这样训练就得到4个模型

生成器模型源和目的2个，然后是判别器的2个

推理

使用生成器模型去生成一个输入的图片
苹果就变成了橘子

这里有个疑问，为什么不用训练好的模型，而是用了mindrecord训练好的自带的。这不是搞错了吗？