如何解决PyTorch中奇怪的cuda错误？

Question

我从凯拉斯搬到了皮托克。我对搬到CUDA这件事很陌生。我花了好几个小时在网上冲浪，却什么都没找到？修复可能是一两行。如果有人知道如何解决这个问题，我会很感激的？

下面是我的代码，首先我将我的u-net模型定义为一个nn类。类似于以下代码的模块：

import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.data import Dataset, DataLoader
from torchvision import transforms, utils


class unet(nn.Module):
def __init__(self):
    super(unet, self).__init__()
    self.conv1 = nn.Conv3d(1, 32, 3, padding=1)
    self.conv1_1 = nn.Conv3d(32, 32, 3, padding=1)
    self.conv2 = nn.Conv3d(32, 64, 3, padding=1)
    self.conv2_2 = nn.Conv3d(64, 64, 3, padding=1)
    self.conv3 = nn.Conv3d(64, 128, 3, padding=1)
    self.conv3_3 = nn.Conv3d(128, 128, 3, padding=1)
    self.convT1 = nn.ConvTranspose3d(128, 64, 3, stride=(2,2,2), padding=1, output_padding=1)
    self.conv4 = nn.Conv3d(128, 64, 3, padding=1)
    self.conv4_4 = nn.Conv3d(64, 64, 3, padding=1)
    self.convT2 = nn.ConvTranspose3d(64, 32, 3,stride=(2,2,2), padding=1, output_padding=1)
    self.conv5 = nn.Conv3d(64, 32, 3, padding=1)
    self.conv5_5 = nn.Conv3d(32, 32, 3, padding=1)
    self.conv6 = nn.Conv3d(32, 1 ,3, padding=1)


def forward(self, inputs):
    conv1 = F.relu(self.conv1(inputs))
    conv1 = F.relu(self.conv1_1(conv1))
    pool1 = F.max_pool3d(conv1, 2)

    conv2 = F.relu(self.conv2(pool1))
    conv2 = F.relu(self.conv2_2(conv2))
    pool2 = F.max_pool3d(conv2, 2)        
    conv3 = F.relu(self.conv3(pool2))
    conv3 = F.relu(self.conv3_3(conv3))        
    conv3 = self.convT1(conv3)

    up1 = torch.cat((conv3, conv2), dim=1)
    conv4 = F.relu(self.conv4(up1))
    conv4 = F.relu(self.conv4_4(conv4))

    conv4 = self.convT2(conv4)
    up2 = torch.cat((conv4, conv1), dim=1)
    conv5 = F.relu(self.conv5(up2))
    conv5 = F.relu(self.conv5_5(conv5))

    conv6 = F.relu(self.conv6(conv5))

    return conv6

然后我像下面的代码一样运行我的unet。请注意，在定义模块时，我将其设置为cuda。我还将输入数据及其标签设置为cuda。

device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model = unet().to(device)
optimizer = torch.optim.Adam(model.parameters(), lr=1e-3)
loss_fn = nn.MSELoss()
datasets = torch.utils.data.TensorDataset(data_recon, data_truth)
train_loader = DataLoader(datasets, batch_size=2, shuffle=True)

def training_loop(n_epochs, optimizer, model, loss_fn, train_loader):
    for epoch in range(1, n_epochs + 1):
        loss_train = 0
        for imgs, labels in train_loader:
            imgs.to(device)
            labels.to(device)
            outputs = model(imgs)
            loss = loss_fn(outputs, labels)

            optimizer.zero_grad()
            loss.backward()
            optimizer.step()

            loss_train += loss.item()


        print('{} Epoch {}, Training loss    {}'.format(datetime.datetime.now(), epoch, float(loss_train)))

training_loop(50, optimizer, model, loss_fn, train_loader)

但是我得到了这个错误：

RuntimeError Traceback（最近的调用最后）在----

在训练循环中（n个时期、优化器、模型、损耗、训练装载机）5个IMG。至（装置）6个标签。至（装置）----

/opt/anaconda3/lib/python3.7/site-packages/torch/nn/modules/module.py在调用（自我，*输入，**kwargs）491结果=自我。_slow_forward（*输入，**kwargs）492其他：-

前进档（自身，输入）18 19 def前进档（自身，输入）：---

/opt/anaconda3/lib/python3.7/site-packages/torch/nn/modules/module.py在调用（自我，*输入，**kwargs）491结果=自我。_slow_forward（*输入，**kwargs）492其他：-

/opt/anaconda3/lib/python3。7/现场包/torch/nn/modules/conv.py正向（自我，输入）474自我。自我膨胀。组）475返回F.conv3d（输入、体重、偏差、步幅、--

运行时错误：后端CPU的预期对象，但参数#2的权重得到后端CUDA

我花了几个小时上网，却什么也没找到？修复可能是一两行。如果有人知道如何解决这个问题，我将不胜感激？

Answer 1

问题是这条线

imgs.to(device)
labels.to(device)

. to（设备）返回一个新的张量，并且不会更改imgs和标签。所以cuda错误是有效的。您可以简单地通过分配新的张量来修复它，如下所示：

imgs = imgs.to(device)
labels = labels.to(device)