将 pytorch 数据加载器加载到 GPU

Question

Is there a way to load a pytorch DataLoader ( torch.utils.data.Dataloader ) entirely into my GPU?

Now, I load every batch separately into my GPU.

CTX = torch.device('cuda')

train_loader = torch.utils.data.DataLoader(
    train_dataset,
    batch_size=BATCH_SIZE,
    shuffle=True,
    num_workers=0,
)

net = Net().to(CTX)
criterion = nn.CrossEntropyLoss()
optimizer = optim.SGD(net.parameters(), lr=LEARNING_RATE)

for epoch in range(EPOCHS):
    for inputs, labels in test_loader:
        inputs = inputs.to(CTX)        # this is where the data is loaded into GPU
        labels = labels.to(CTX)        

        optimizer.zero_grad()

        outputs = net(inputs)
        loss = criterion(outputs, labels)
        loss.backward()
        optimizer.step()

    print(f'training accuracy: {net.validate(train_loader, device=CTX)}/{len(train_dataset)}')
    print(f'validation accuracy: {net.validate(test_loader, device=CTX)}/{len(test_dataset)}')

where the Net.validate() function is given by

def validate(self, val_loader, device=torch.device('cpu')):
    correct = 0
    for inputs, labels in val_loader:
        inputs = inputs.to(device)
        labels = labels.to(device)
        outputs = torch.argmax(self(inputs), dim=1)
        correct += int(torch.sum(outputs==labels))
    return correct

I would like to improve the speed by loading the entire dataset trainloader into my GPU, instead of loading every batch separately. So, I would like to do something like

train_loader.to(CTX)

Is there an equivalent function for this? Because torch.utils.data.DataLoader does not have this attribute .to() .

I work with an NVIDIA GeForce RTX 2060 with CUDA Toolkit 10.2 installed.

Answer 1

you can put your data of dataset in advance

train_dataset.train_data.to(CTX)  #train_dataset.train_data is a Tensor(input data)
train_dataset.train_labels.to(CTX)

for example of minst

import torch
from torch.utils.data import DataLoader
from torchvision import datasets
from torchvision import transforms
batch_size = 64
transform = transforms.Compose([
    transforms.ToTensor(),
    transforms.Normalize((0.1307,), (0.3081,))
])
train_data = datasets.MNIST(
    root='./dataset/minst/',
    train=True,
    download=False,
    transform=transform
)
train_loader = DataLoader(
    dataset=train_data,
    shuffle=True,
    batch_size=batch_size
)
train_data.train_data.to(torch.device("cuda:0"))  # put data into GPU entirely
train_data.train_labels.to(torch.device("cuda:0"))

I got this solution by using debugger...

Answer 2

In the "Wrapping Dataloader" part in this tutorial ( https://pytorch.org/tutorials/beginner/nn_tutorial.html ), data are loaded into GPU entirely. The wrapper dataloader code is as follows:

def preprocess(x, y):
    return x.view(-1, 1, 28, 28).to(dev), y.to(dev)

train_dl, valid_dl = get_data(train_ds, valid_ds, bs)
train_dl = WrappedDataLoader(train_dl, preprocess)
valid_dl = WrappedDataLoader(valid_dl, preprocess)