注意
转到末尾 以下载完整的示例代码。
将 TensorDict 用于数据集¶
在本教程中，我们将演示如何使用 TensorDict 高效且透明地在训练管道中加载和管理数据。本教程主要基于 PyTorch 快速入门教程，并进行了修改以展示 TensorDict 的用法。
import torch
import torch.nn as nn
from tensordict import MemoryMappedTensor, TensorDict
from torch.utils.data import DataLoader
from torchvision import datasets
from torchvision.transforms import ToTensor
device = "cuda" if torch.cuda.is_available() else "cpu"
print(f"Using device: {device}")
Using device: cpu
torchvision.datasets 模块包含许多方便的预备数据集。在本教程中，我们将使用相对简单的 FashionMNIST 数据集。每张图像都是一件衣服，目标是分类图像中的衣服类型（例如，“包”、“运动鞋”等）。
training_data = datasets.FashionMNIST(
root="data",
train=True,
download=True,
transform=ToTensor(),
)
test_data = datasets.FashionMNIST(
root="data",
train=False,
download=True,
transform=ToTensor(),
)
我们将创建两个 tensordict，分别用于训练数据和测试数据。我们创建内存映射张量来保存数据。这将使我们能够高效地从磁盘加载转换后的数据批次，而不是重复加载和转换单个图像。
首先，我们创建 MemoryMappedTensor 容器。
training_data_td = TensorDict(
{
"images": MemoryMappedTensor.empty(
(len(training_data), *training_data[0][0].squeeze().shape),
dtype=torch.float32,
),
"targets": MemoryMappedTensor.empty((len(training_data),), dtype=torch.int64),
},
batch_size=[len(training_data)],
device=device,
)
test_data_td = TensorDict(
{
"images": MemoryMappedTensor.empty(
(len(test_data), *test_data[0][0].squeeze().shape), dtype=torch.float32
),
"targets": MemoryMappedTensor.empty((len(test_data),), dtype=torch.int64),
},
batch_size=[len(test_data)],
device=device,
)
然后我们可以迭代数据来填充内存映射张量。这需要一些时间，但预先执行转换可以节省后续训练中的重复工作。
数据加载器¶
我们将从 torchvision 提供的 Dataset 以及我们的内存映射 TensorDict 创建 DataLoaders。
由于 TensorDict 实现了 __len__ 和 __getitem__（以及 __getitems__），我们可以像 map-style Dataset 一样使用它，并直接从中创建 DataLoader。请注意，由于 TensorDict 已经可以处理批量索引，因此无需进行 collation（整理），因此我们将 identity 函数作为 collate_fn 传递。
batch_size = 64
train_dataloader = DataLoader(training_data, batch_size=batch_size) # noqa: TOR401
test_dataloader = DataLoader(test_data, batch_size=batch_size) # noqa: TOR401
train_dataloader_td = DataLoader( # noqa: TOR401
training_data_td, batch_size=batch_size, collate_fn=lambda x: x
)
test_dataloader_td = DataLoader( # noqa: TOR401
test_data_td, batch_size=batch_size, collate_fn=lambda x: x
)
模型¶
我们使用与 快速入门教程 中相同的模型。
class Net(nn.Module):
def __init__(self):
super().__init__()
self.flatten = nn.Flatten()
self.linear_relu_stack = nn.Sequential(
nn.Linear(28 * 28, 512),
nn.ReLU(),
nn.Linear(512, 512),
nn.ReLU(),
nn.Linear(512, 10),
)
def forward(self, x):
x = self.flatten(x)
logits = self.linear_relu_stack(x)
return logits
model = Net().to(device)
model_td = Net().to(device)
model, model_td
(Net(
(flatten): Flatten(start_dim=1, end_dim=-1)
(linear_relu_stack): Sequential(
(0): Linear(in_features=784, out_features=512, bias=True)
(1): ReLU()
(2): Linear(in_features=512, out_features=512, bias=True)
(3): ReLU()
(4): Linear(in_features=512, out_features=10, bias=True)
)
), Net(
(flatten): Flatten(start_dim=1, end_dim=-1)
(linear_relu_stack): Sequential(
(0): Linear(in_features=784, out_features=512, bias=True)
(1): ReLU()
(2): Linear(in_features=512, out_features=512, bias=True)
(3): ReLU()
(4): Linear(in_features=512, out_features=10, bias=True)
)
))
优化参数¶
我们将使用随机梯度下降和交叉熵损失来优化模型的参数。
loss_fn = nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(model.parameters(), lr=1e-3)
optimizer_td = torch.optim.SGD(model_td.parameters(), lr=1e-3)
def train(dataloader, model, loss_fn, optimizer):
size = len(dataloader.dataset)
model.train()
for batch, (X, y) in enumerate(dataloader):
X, y = X.to(device), y.to(device)
pred = model(X)
loss = loss_fn(pred, y)
optimizer.zero_grad()
loss.backward()
optimizer.step()
if batch % 100 == 0:
loss, current = loss.item(), batch * len(X)
print(f"loss: {loss:>7f} [{current:>5d}/{size:>5d}]")
我们基于 TensorDict 的 DataLoader 的训练循环非常相似，我们只需调整如何解包数据，使其采用 TensorDict 提供的更显式的基于键的检索方式。.contiguous() 方法加载存储在 memmap 张量中的数据。
def train_td(dataloader, model, loss_fn, optimizer):
size = len(dataloader.dataset)
model.train()
for batch, data in enumerate(dataloader):
X, y = data["images"].contiguous(), data["targets"].contiguous()
pred = model(X)
loss = loss_fn(pred, y)
optimizer.zero_grad()
loss.backward()
optimizer.step()
if batch % 100 == 0:
loss, current = loss.item(), batch * len(X)
print(f"loss: {loss:>7f} [{current:>5d}/{size:>5d}]")
def test(dataloader, model, loss_fn):
size = len(dataloader.dataset)
num_batches = len(dataloader)
model.eval()
test_loss, correct = 0, 0
with torch.no_grad():
for X, y in dataloader:
X, y = X.to(device), y.to(device)
pred = model(X)
test_loss += loss_fn(pred, y).item()
correct += (pred.argmax(1) == y).type(torch.float).sum().item()
test_loss /= num_batches
correct /= size
print(
f"Test Error: \n Accuracy: {(100 * correct):>0.1f}%, Avg loss: {test_loss:>8f} \n"
)
def test_td(dataloader, model, loss_fn):
size = len(dataloader.dataset)
num_batches = len(dataloader)
model.eval()
test_loss, correct = 0, 0
with torch.no_grad():
for batch in dataloader:
X, y = batch["images"].contiguous(), batch["targets"].contiguous()
pred = model(X)
test_loss += loss_fn(pred, y).item()
correct += (pred.argmax(1) == y).type(torch.float).sum().item()
test_loss /= num_batches
correct /= size
print(
f"Test Error: \n Accuracy: {(100 * correct):>0.1f}%, Avg loss: {test_loss:>8f} \n"
)
for d in train_dataloader_td:
print(d)
break
import time
t0 = time.time()
epochs = 5
for t in range(epochs):
print(f"Epoch {t + 1}\n-------------------------")
train_td(train_dataloader_td, model_td, loss_fn, optimizer_td)
test_td(test_dataloader_td, model_td, loss_fn)
print(f"TensorDict training done! time: {time.time() - t0: 4.4f} s")
t0 = time.time()
epochs = 5
for t in range(epochs):
print(f"Epoch {t + 1}\n-------------------------")
train(train_dataloader, model, loss_fn, optimizer)
test(test_dataloader, model, loss_fn)
print(f"Training done! time: {time.time() - t0: 4.4f} s")
TensorDict(
fields={
images: Tensor(shape=torch.Size([64, 28, 28]), device=cpu, dtype=torch.float32, is_shared=False),
targets: Tensor(shape=torch.Size([64]), device=cpu, dtype=torch.int64, is_shared=False)},
batch_size=torch.Size([64]),
device=cpu,
is_shared=False)
Epoch 1
-------------------------
loss: 2.288576 [ 0/60000]
loss: 2.279963 [ 6400/60000]
loss: 2.268667 [12800/60000]
loss: 2.270308 [19200/60000]
loss: 2.233671 [25600/60000]
loss: 2.215914 [32000/60000]
loss: 2.214695 [38400/60000]
loss: 2.178674 [44800/60000]
loss: 2.172899 [51200/60000]
loss: 2.149208 [57600/60000]
Test Error:
Accuracy: 48.3%, Avg loss: 2.142256
Epoch 2
-------------------------
loss: 2.145386 [ 0/60000]
loss: 2.135478 [ 6400/60000]
loss: 2.081977 [12800/60000]
loss: 2.100613 [19200/60000]
loss: 2.030724 [25600/60000]
loss: 1.981814 [32000/60000]
loss: 1.995256 [38400/60000]
loss: 1.916826 [44800/60000]
loss: 1.928639 [51200/60000]
loss: 1.847330 [57600/60000]
Test Error:
Accuracy: 54.9%, Avg loss: 1.853513
Epoch 3
-------------------------
loss: 1.888883 [ 0/60000]
loss: 1.856458 [ 6400/60000]
loss: 1.744976 [12800/60000]
loss: 1.781419 [19200/60000]
loss: 1.656465 [25600/60000]
loss: 1.631539 [32000/60000]
loss: 1.631200 [38400/60000]
loss: 1.546029 [44800/60000]
loss: 1.584942 [51200/60000]
loss: 1.466830 [57600/60000]
Test Error:
Accuracy: 61.1%, Avg loss: 1.489801
Epoch 4
-------------------------
loss: 1.562606 [ 0/60000]
loss: 1.525372 [ 6400/60000]
loss: 1.381523 [12800/60000]
loss: 1.445276 [19200/60000]
loss: 1.316936 [25600/60000]
loss: 1.337760 [32000/60000]
loss: 1.331586 [38400/60000]
loss: 1.266051 [44800/60000]
loss: 1.314883 [51200/60000]
loss: 1.210139 [57600/60000]
Test Error:
Accuracy: 64.0%, Avg loss: 1.230520
Epoch 5
-------------------------
loss: 1.308549 [ 0/60000]
loss: 1.290072 [ 6400/60000]
loss: 1.127975 [12800/60000]
loss: 1.229955 [19200/60000]
loss: 1.095840 [25600/60000]
loss: 1.140236 [32000/60000]
loss: 1.147135 [38400/60000]
loss: 1.088268 [44800/60000]
loss: 1.143719 [51200/60000]
loss: 1.056760 [57600/60000]
Test Error:
Accuracy: 65.5%, Avg loss: 1.069039
TensorDict training done! time: 8.5605 s
Epoch 1
-------------------------
loss: 2.299359 [ 0/60000]
loss: 2.285545 [ 6400/60000]
loss: 2.273665 [12800/60000]
loss: 2.269405 [19200/60000]
loss: 2.254834 [25600/60000]
loss: 2.229681 [32000/60000]
loss: 2.230662 [38400/60000]
loss: 2.202860 [44800/60000]
loss: 2.191047 [51200/60000]
loss: 2.169984 [57600/60000]
Test Error:
Accuracy: 49.4%, Avg loss: 2.163469
Epoch 2
-------------------------
loss: 2.173521 [ 0/60000]
loss: 2.155031 [ 6400/60000]
loss: 2.110353 [12800/60000]
loss: 2.122467 [19200/60000]
loss: 2.073355 [25600/60000]
loss: 2.024080 [32000/60000]
loss: 2.040011 [38400/60000]
loss: 1.973018 [44800/60000]
loss: 1.969435 [51200/60000]
loss: 1.899309 [57600/60000]
Test Error:
Accuracy: 56.6%, Avg loss: 1.901211
Epoch 3
-------------------------
loss: 1.941732 [ 0/60000]
loss: 1.895479 [ 6400/60000]
loss: 1.793313 [12800/60000]
loss: 1.823587 [19200/60000]
loss: 1.725540 [25600/60000]
loss: 1.682630 [32000/60000]
loss: 1.695444 [38400/60000]
loss: 1.609786 [44800/60000]
loss: 1.630640 [51200/60000]
loss: 1.519337 [57600/60000]
Test Error:
Accuracy: 61.0%, Avg loss: 1.542196
Epoch 4
-------------------------
loss: 1.619711 [ 0/60000]
loss: 1.563892 [ 6400/60000]
loss: 1.428948 [12800/60000]
loss: 1.484869 [19200/60000]
loss: 1.386341 [25600/60000]
loss: 1.376361 [32000/60000]
loss: 1.382083 [38400/60000]
loss: 1.319620 [44800/60000]
loss: 1.352009 [51200/60000]
loss: 1.241564 [57600/60000]
Test Error:
Accuracy: 63.6%, Avg loss: 1.274075
Epoch 5
-------------------------
loss: 1.360628 [ 0/60000]
loss: 1.321853 [ 6400/60000]
loss: 1.170190 [12800/60000]
loss: 1.258210 [19200/60000]
loss: 1.151774 [25600/60000]
loss: 1.169670 [32000/60000]
loss: 1.181415 [38400/60000]
loss: 1.131277 [44800/60000]
loss: 1.170511 [51200/60000]
loss: 1.072559 [57600/60000]
Test Error:
Accuracy: 64.7%, Avg loss: 1.101794
Training done! time: 34.6420 s
脚本总运行时间：(0 分钟 56.008 秒)
