python - 导入透明图像会导致 RuntimeError：张量 a (4) 的大小必须与非单维 0 处的张量 b (3) 的大小相匹配

Question

我正在努力学习人工智能。我有带有 ALPHA 通道（透明度）图像的 GAN（生成对抗网络）代码。所有图像都有 alpha 通道。为了证明我写了image_validator.py如下的小程序

from PIL import Image
import glob


def main():
    image_list = []
    img_number = 0
    for filename in glob.glob('data/*/*.*'):
        try:
            im = Image.open(filename)
            # print(filename)
            if str(im.mode) != "RGBA":
                print("alpha " + str(im.mode))
                img_number = img_number+1
                print(str(img_number))
        except Exception as e:
            print("Error : "+filename)


if __name__ == "__main__":
    main()

上面的程序不打印任何内容，这意味着所有图像都有 ALPHA 通道。为了测试上面的程序，我添加了没有 ALPHA 通道的单个图像。所以我可以确认所有图像都有 ALPHA CHANNEL。

我generator.py的就像下面

import torch.nn as nn
class G(nn.Module):
    feature_maps = 512
    kernel_size = 4
    stride = 2
    padding = 1
    bias = False

    def __init__(self, input_vector):
        super(G, self).__init__()
        self.main = nn.Sequential(
            nn.ConvTranspose2d(input_vector, self.feature_maps, self.kernel_size, 1, 0, bias=self.bias),
            nn.BatchNorm2d(self.feature_maps), nn.ReLU(True),
            nn.ConvTranspose2d(self.feature_maps, int(self.feature_maps // 2), self.kernel_size, self.stride, self.padding,
                               bias=self.bias),
            nn.BatchNorm2d(int(self.feature_maps // 2)), nn.ReLU(True),
            nn.ConvTranspose2d(int(self.feature_maps // 2), int((self.feature_maps // 2) // 2), self.kernel_size, self.stride,
                               self.padding,
                               bias=self.bias),
            nn.BatchNorm2d(int((self.feature_maps // 2) // 2)), nn.ReLU(True),
            nn.ConvTranspose2d((int((self.feature_maps // 2) // 2)), int(((self.feature_maps // 2) // 2) // 2), self.kernel_size,
                               self.stride, self.padding,
                               bias=self.bias),
            nn.BatchNorm2d(int((self.feature_maps // 2) // 2) // 2), nn.ReLU(True),
            nn.ConvTranspose2d(int(((self.feature_maps // 2) // 2) // 2), 4, self.kernel_size, self.stride, self.padding,
                               bias=self.bias),
            nn.Tanh()
        )

    def forward(self, input):
        output = self.main(input)
        return output

我discriminator.py的就像下面

    import torch.nn as nn
class D(nn.Module):
    feature_maps = 64
    kernel_size = 4
    stride = 2
    padding = 1
    bias = False
    inplace = True

    def __init__(self):
        super(D, self).__init__()
        self.main = nn.Sequential(
            nn.Conv2d(4, self.feature_maps, self.kernel_size, self.stride, self.padding, bias=self.bias),
            nn.LeakyReLU(0.2, inplace=self.inplace),
            nn.Conv2d(self.feature_maps, self.feature_maps * 2, self.kernel_size, self.stride, self.padding,
                      bias=self.bias),
            nn.BatchNorm2d(self.feature_maps * 2), nn.LeakyReLU(0.2, inplace=self.inplace),
            nn.Conv2d(self.feature_maps * 2, self.feature_maps * (2 * 2), self.kernel_size, self.stride, self.padding,
                      bias=self.bias),
            nn.BatchNorm2d(self.feature_maps * (2 * 2)), nn.LeakyReLU(0.2, inplace=self.inplace),
            nn.Conv2d(self.feature_maps * (2 * 2), self.feature_maps * (2 * 2 * 2), self.kernel_size, self.stride,
                      self.padding, bias=self.bias),
            nn.BatchNorm2d(self.feature_maps * (2 * 2 * 2)), nn.LeakyReLU(0.2, inplace=self.inplace),
            nn.Conv2d(self.feature_maps * (2 * 2 * 2), 1, self.kernel_size, 1, 0, bias=self.bias),
            nn.Sigmoid()
        )

    def forward(self, input):
        output = self.main(input)
        return output.view(-1)

我的主程序gan.py如下

    # Importing the libraries
from __future__ import print_function
import torch.nn as nn
import torch.optim as optim
import torch.utils.data
import torchvision.datasets as dset
import torchvision.transforms as transforms
import torchvision.utils as vutils
from torch.autograd import Variable
from generator import G
from discriminator import D
import os
from PIL import Image

batchSize = 64  # We set the size of the batch.
imageSize = 64  # We set the size of the generated images (64x64).
input_vector = 100
nb_epochs = 500
# Creating the transformations
transform = transforms.Compose([transforms.Resize((imageSize, imageSize)), transforms.ToTensor(),
                                transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5,
                                                                       0.5)), ])  # We create a list of transformations (scaling, tensor conversion, normalization) to apply to the input images.


def pil_loader_rgba(path: str) -> Image.Image:
    with open(path, 'rb') as f:
        img = Image.open(f)
        return img.convert('RGBA')


# Loading the dataset
dataset = dset.ImageFolder(root='./data', transform=transform, loader=pil_loader_rgba)
dataloader = torch.utils.data.DataLoader(dataset, batch_size=batchSize, shuffle=True,
                                         num_workers=2)  # We use dataLoader to get the images of the training set batch by batch.


# Defining the weights_init function that takes as input a neural network m and that will initialize all its weights.
def weights_init(m):
    classname = m.__class__.__name__
    if classname.find('Conv') != -1:
        m.weight.data.normal_(0.0, 0.02)
    elif classname.find('BatchNorm') != -1:
        m.weight.data.normal_(1.0, 0.02)
        m.bias.data.fill_(0)


def is_cuda_available():
    return torch.cuda.is_available()


def is_gpu_available():
    if is_cuda_available():
        if int(torch.cuda.device_count()) > 0:
            return True
        return False
    return False


# Create results directory
def create_dir(name):
    if not os.path.exists(name):
        os.makedirs(name)


# Creating the generator
netG = G(input_vector)
netG.apply(weights_init)

# Creating the discriminator
netD = D()
netD.apply(weights_init)

if is_gpu_available():
    netG.cuda()
    netD.cuda()

# Training the DCGANs

criterion = nn.BCELoss()
optimizerD = optim.Adam(netD.parameters(), lr=0.0002, betas=(0.5, 0.999))
optimizerG = optim.Adam(netG.parameters(), lr=0.0002, betas=(0.5, 0.999))

generator_model = 'generator_model'
discriminator_model = 'discriminator_model'


def save_model(epoch, model, optimizer, error, filepath, noise=None):
    if os.path.exists(filepath):
        os.remove(filepath)
    torch.save({
        'epoch': epoch,
        'model_state_dict': model.state_dict(),
        'optimizer_state_dict': optimizer.state_dict(),
        'loss': error,
        'noise': noise
    }, filepath)


def load_checkpoint(filepath):
    if os.path.exists(filepath):
        return torch.load(filepath)
    return None


def main():
    print("Device name : " + torch.cuda.get_device_name(0))
    for epoch in range(nb_epochs):

        for i, data in enumerate(dataloader, 0):
            checkpointG = load_checkpoint(generator_model)
            checkpointD = load_checkpoint(discriminator_model)
            if checkpointG:
                print("checkpointG")
                netG.load_state_dict(checkpointG['model_state_dict'])
                optimizerG.load_state_dict(checkpointG['optimizer_state_dict'])
            if checkpointD:
                netD.load_state_dict(checkpointD['model_state_dict'])
                optimizerD.load_state_dict(checkpointD['optimizer_state_dict'])

            # 1st Step: Updating the weights of the neural network of the discriminator

            netD.zero_grad()

            # Training the discriminator with a real image of the dataset
            real, _ = data
            if is_gpu_available():
                print("True")
                input = Variable(real.cuda()).cuda()
                target = Variable(torch.ones(input.size()[0]).cuda()).cuda()
            else:
                input = Variable(real)
                target = Variable(torch.ones(input.size()[0]))
            output = netD(input)
            errD_real = criterion(output, target)

            # Training the discriminator with a fake image generated by the generator
            if is_gpu_available():
                noise = Variable(torch.randn(input.size()[0], input_vector, 1, 1)).cuda()
                target = Variable(torch.zeros(input.size()[0])).cuda()
            else:
                noise = Variable(torch.randn(input.size()[0], input_vector, 1, 1))
                target = Variable(torch.zeros(input.size()[0]))
            fake = netG(noise)
            output = netD(fake.detach())
            errD_fake = criterion(output, target)

            # Backpropagating the total error
            errD = errD_real + errD_fake
            errD.backward()
            optimizerD.step()

            # 2nd Step: Updating the weights of the neural network of the generator
            netG.zero_grad()
            if is_gpu_available():
                target = Variable(torch.ones(input.size()[0])).cuda()
            else:
                target = Variable(torch.ones(input.size()[0]))
            output = netD(fake)
            errG = criterion(output, target)
            errG.backward()
            optimizerG.step()

            # 3rd Step: Printing the losses and saving the real images and the generated images of the minibatch every 100 steps

            print('[%d/%d][%d/%d] Loss_D: %.4f Loss_G: %.4f' % (
            epoch, nb_epochs, i, len(dataloader), errD.data, errG.data))
            save_model(epoch, netG, optimizerG, errG, generator_model, noise)
            save_model(epoch, netD, optimizerD, errD, discriminator_model, noise)

            if i % 100 == 0:
                create_dir('results')
                vutils.save_image(real, '%s/real_samples.png' % "./results", normalize=True)
                fake = netG(noise)
                vutils.save_image(fake.data, '%s/fake_samples_epoch_%03d.png' % ("./results", epoch), normalize=True)


if __name__ == "__main__":
    main()

但是当我运行我的程序时，我收到了这个错误

Traceback (most recent call last):
  File ".\gans.py", line 178, in <module>
    main()
  File ".\gans.py", line 109, in main
    for i, data in enumerate(dataloader, 0):
  File "C:\Users\Akila\anaconda3\lib\site-packages\torch\utils\data\dataloader.py", line 521, in __next__
    data = self._next_data()
  File "C:\Users\Akila\anaconda3\lib\site-packages\torch\utils\data\dataloader.py", line 1203, in _next_data
    return self._process_data(data)
  File "C:\Users\Akila\anaconda3\lib\site-packages\torch\utils\data\dataloader.py", line 1229, in _process_data
    data.reraise()
  File "C:\Users\Akila\anaconda3\lib\site-packages\torch\_utils.py", line 425, in reraise
    raise self.exc_type(msg)
RuntimeError: Caught RuntimeError in DataLoader worker process 0.
Original Traceback (most recent call last):
  File "C:\Users\Akila\anaconda3\lib\site-packages\torch\utils\data\_utils\worker.py", line 287, in _worker_loop
    data = fetcher.fetch(index)
  File "C:\Users\Akila\anaconda3\lib\site-packages\torch\utils\data\_utils\fetch.py", line 44, in fetch
    data = [self.dataset[idx] for idx in possibly_batched_index]
  File "C:\Users\Akila\anaconda3\lib\site-packages\torch\utils\data\_utils\fetch.py", line 44, in <listcomp>
    data = [self.dataset[idx] for idx in possibly_batched_index]
  File "C:\Users\Akila\anaconda3\lib\site-packages\torchvision\datasets\folder.py", line 234, in __getitem__
    sample = self.transform(sample)
  File "C:\Users\Akila\anaconda3\lib\site-packages\torchvision\transforms\transforms.py", line 60, in __call__
    img = t(img)
  File "C:\Users\Akila\anaconda3\lib\site-packages\torch\nn\modules\module.py", line 1051, in _call_impl
    return forward_call(*input, **kwargs)
  File "C:\Users\Akila\anaconda3\lib\site-packages\torchvision\transforms\transforms.py", line 221, in forward
    return F.normalize(tensor, self.mean, self.std, self.inplace)
  File "C:\Users\Akila\anaconda3\lib\site-packages\torchvision\transforms\functional.py", line 335, in normalize
    tensor.sub_(mean).div_(std)
RuntimeError: The size of tensor a (4) must match the size of tensor b (3) at non-singleton dimension 0

我进行了调试，发现问题出在这一行

for i, data in enumerate(dataloader, 0):

如果我改变这一行 -->return img.convert('RGBA') 到这个 -->return img.convert('RGB')

程序运行良好。但我可以保证所有图像都有 alpha 通道。因为我的image_validator.py程序什么也没打印

我什至尝试使用具有 alpha 通道的 SINGLE 图像运行我的主程序仍然给出相同的错误。

我究竟做错了什么 ？如何保持图像的透明度？我不想失去我的透明度。

标签： pythonpython-3.xpytorchgenerative-adversarial-networkpytorch-dataloader

Answer 1

关于错误信息

RuntimeError : 张量 a(4)的大小必须与b (3)非单维的张量大小匹配0

会导致暗示此调用存在问题：sample = self.transform(sample)

实际上，问题在于您使用的T.Normalize转换只需要三个通道（您只为三个通道指定了平均值和标准，而不是四个）。

transform = transforms.Compose([
   transforms.Resize((imageSize, imageSize)), 
   transforms.ToTensor(),
   transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5,0.5))])

相反，您应该为两个参数提供一个四元素元组。例如（这是一个示例，这可能会运行但不一定有意义......请参阅下面的解释）：

    transforms.Normalize((0.5, 0.5, 0.5, 0.5), (0.5, 0.5,0.5, 0.5))])

---

除此之外，我应该问：你知道你为什么同时使用and.5的参数meanstd吗？如果没有，您可能没有正确使用它。请阅读此答案及其应用程序。