.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
d2l.show_images(gluon.data.vision.CIFAR10(
train=True)[0:32][0], 4, 8, scale=0.8);
.. raw:: latex
\diilbookstyleoutputcell
.. parsed-literal::
:class: output
Downloading /opt/mxnet/datasets/cifar10/cifar-10-binary.tar.gz from https://apache-mxnet.s3-accelerate.dualstack.amazonaws.com/gluon/dataset/cifar10/cifar-10-binary.tar.gz...
.. figure:: output_image-augmentation_7d0887_123_1.svg
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
all_images = torchvision.datasets.CIFAR10(train=True, root="../data",
download=True)
d2l.show_images([all_images[i][0] for i in range(32)], 4, 8, scale=0.8);
.. raw:: latex
\diilbookstyleoutputcell
.. parsed-literal::
:class: output
Downloading https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz to ../data/cifar-10-python.tar.gz
.. raw:: latex
\diilbookstyleoutputcell
.. parsed-literal::
:class: output
0%| | 0/170498071 [00:00
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
all_images = paddlevision.datasets.Cifar10(mode='train' , download=True)
print(len(all_images))
d2l.show_images([all_images[i][0] for i in range(32)], 4, 8, scale=0.8);
.. raw:: latex
\diilbookstyleoutputcell
.. parsed-literal::
:class: output
Cache file /home/ci/.cache/paddle/dataset/cifar/cifar-10-python.tar.gz not found, downloading https://dataset.bj.bcebos.com/cifar/cifar-10-python.tar.gz
Begin to download
item 41626/41626 [============================>.] - ETA: 0s - 1ms/item
Download finished
50000
.. figure:: output_image-augmentation_7d0887_129_1.svg
.. raw:: html
.. raw:: html
为了在预测过程中得到确切的结果，我们通常对训练样本只进行图像增广，且在预测过程中不使用随机操作的图像增广。
在这里，我们只使用最简单的随机左右翻转。
此外，我们使用\ ``ToTensor``\ 实例将一批图像转换为深度学习框架所要求的格式，即形状为（批量大小，通道数，高度，宽度）的32位浮点数，取值范围为0～1。
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
train_augs = gluon.data.vision.transforms.Compose([
gluon.data.vision.transforms.RandomFlipLeftRight(),
gluon.data.vision.transforms.ToTensor()])
test_augs = gluon.data.vision.transforms.Compose([
gluon.data.vision.transforms.ToTensor()])
接下来，我们定义了一个辅助函数，以便于读取图像和应用图像增广。Gluon数据集提供的\ ``transform_first``\ 函数将图像增广应用于每个训练样本的第一个元素（由图像和标签组成），即应用在图像上。有关\ ``DataLoader``\ 的详细介绍，请参阅
:numref:`sec_fashion_mnist`\ 。
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
def load_cifar10(is_train, augs, batch_size):
return gluon.data.DataLoader(
gluon.data.vision.CIFAR10(train=is_train).transform_first(augs),
batch_size=batch_size, shuffle=is_train,
num_workers=d2l.get_dataloader_workers())
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
train_augs = torchvision.transforms.Compose([
torchvision.transforms.RandomHorizontalFlip(),
torchvision.transforms.ToTensor()])
test_augs = torchvision.transforms.Compose([
torchvision.transforms.ToTensor()])
接下来，我们定义一个辅助函数，以便于读取图像和应用图像增广。PyTorch数据集提供的\ ``transform``\ 参数应用图像增广来转化图像。有关\ ``DataLoader``\ 的详细介绍，请参阅
:numref:`sec_fashion_mnist`\ 。
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
def load_cifar10(is_train, augs, batch_size):
dataset = torchvision.datasets.CIFAR10(root="../data", train=is_train,
transform=augs, download=True)
dataloader = torch.utils.data.DataLoader(dataset, batch_size=batch_size,
shuffle=is_train, num_workers=d2l.get_dataloader_workers())
return dataloader
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
train_augs = paddlevision.transforms.Compose([
paddlevision.transforms.RandomHorizontalFlip(),
paddlevision.transforms.ToTensor()])
test_augs = paddlevision.transforms.Compose([
paddlevision.transforms.ToTensor()])
def load_cifar10(is_train, augs, batch_size):
dataset = paddlevision.datasets.Cifar10(mode="train",
transform=augs, download=True)
dataloader = paddle.io.DataLoader(dataset, batch_size=batch_size,
num_workers=d2l.get_dataloader_workers(), shuffle=is_train)
return dataloader
.. raw:: html
.. raw:: html
多GPU训练
~~~~~~~~~
我们在CIFAR-10数据集上训练 :numref:`sec_resnet`\ 中的ResNet-18模型。
回想一下 :numref:`sec_multi_gpu_concise`\ 中对多GPU训练的介绍。
接下来，我们定义一个函数，使用多GPU对模型进行训练和评估。
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
#@save
def train_batch_ch13(net, features, labels, loss, trainer, devices,
split_f=d2l.split_batch):
"""用多GPU进行小批量训练"""
X_shards, y_shards = split_f(features, labels, devices)
with autograd.record():
pred_shards = [net(X_shard) for X_shard in X_shards]
ls = [loss(pred_shard, y_shard) for pred_shard, y_shard
in zip(pred_shards, y_shards)]
for l in ls:
l.backward()
# True标志允许使用过时的梯度，这很有用（例如，在微调BERT中）
trainer.step(labels.shape[0], ignore_stale_grad=True)
train_loss_sum = sum([float(l.sum()) for l in ls])
train_acc_sum = sum(d2l.accuracy(pred_shard, y_shard)
for pred_shard, y_shard in zip(pred_shards, y_shards))
return train_loss_sum, train_acc_sum
#@save
def train_ch13(net, train_iter, test_iter, loss, trainer, num_epochs,
devices=d2l.try_all_gpus(), split_f=d2l.split_batch):
"""用多GPU进行模型训练"""
timer, num_batches = d2l.Timer(), len(train_iter)
animator = d2l.Animator(xlabel='epoch', xlim=[1, num_epochs], ylim=[0, 1],
legend=['train loss', 'train acc', 'test acc'])
for epoch in range(num_epochs):
# 4个维度：储存训练损失，训练准确度，实例数，特点数
metric = d2l.Accumulator(4)
for i, (features, labels) in enumerate(train_iter):
timer.start()
l, acc = train_batch_ch13(
net, features, labels, loss, trainer, devices, split_f)
metric.add(l, acc, labels.shape[0], labels.size)
timer.stop()
if (i + 1) % (num_batches // 5) == 0 or i == num_batches - 1:
animator.add(epoch + (i + 1) / num_batches,
(metric[0] / metric[2], metric[1] / metric[3],
None))
test_acc = d2l.evaluate_accuracy_gpus(net, test_iter, split_f)
animator.add(epoch + 1, (None, None, test_acc))
print(f'loss {metric[0] / metric[2]:.3f}, train acc '
f'{metric[1] / metric[3]:.3f}, test acc {test_acc:.3f}')
print(f'{metric[2] * num_epochs / timer.sum():.1f} examples/sec on '
f'{str(devices)}')
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
#@save
def train_batch_ch13(net, X, y, loss, trainer, devices):
"""用多GPU进行小批量训练"""
if isinstance(X, list):
# 微调BERT中所需
X = [x.to(devices[0]) for x in X]
else:
X = X.to(devices[0])
y = y.to(devices[0])
net.train()
trainer.zero_grad()
pred = net(X)
l = loss(pred, y)
l.sum().backward()
trainer.step()
train_loss_sum = l.sum()
train_acc_sum = d2l.accuracy(pred, y)
return train_loss_sum, train_acc_sum
#@save
def train_ch13(net, train_iter, test_iter, loss, trainer, num_epochs,
devices=d2l.try_all_gpus()):
"""用多GPU进行模型训练"""
timer, num_batches = d2l.Timer(), len(train_iter)
animator = d2l.Animator(xlabel='epoch', xlim=[1, num_epochs], ylim=[0, 1],
legend=['train loss', 'train acc', 'test acc'])
net = nn.DataParallel(net, device_ids=devices).to(devices[0])
for epoch in range(num_epochs):
# 4个维度：储存训练损失，训练准确度，实例数，特点数
metric = d2l.Accumulator(4)
for i, (features, labels) in enumerate(train_iter):
timer.start()
l, acc = train_batch_ch13(
net, features, labels, loss, trainer, devices)
metric.add(l, acc, labels.shape[0], labels.numel())
timer.stop()
if (i + 1) % (num_batches // 5) == 0 or i == num_batches - 1:
animator.add(epoch + (i + 1) / num_batches,
(metric[0] / metric[2], metric[1] / metric[3],
None))
test_acc = d2l.evaluate_accuracy_gpu(net, test_iter)
animator.add(epoch + 1, (None, None, test_acc))
print(f'loss {metric[0] / metric[2]:.3f}, train acc '
f'{metric[1] / metric[3]:.3f}, test acc {test_acc:.3f}')
print(f'{metric[2] * num_epochs / timer.sum():.1f} examples/sec on '
f'{str(devices)}')
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
#@save
def train_batch_ch13(net, X, y, loss, trainer, devices):
"""用多GPU进行小批量训练
飞桨不支持在notebook上进行多GPU训练
Defined in :numref:`sec_image_augmentation`"""
if isinstance(X, list):
# 微调BERT中所需（稍后讨论）
X = [paddle.to_tensor(x, place=devices[0]) for x in X]
else:
X = paddle.to_tensor(X, place=devices[0])
y = paddle.to_tensor(y, place=devices[0])
net.train()
trainer.clear_grad()
pred = net(X)
l = loss(pred, y)
l.sum().backward()
trainer.step()
train_loss_sum = l.sum()
train_acc_sum = d2l.accuracy(pred, y)
return train_loss_sum, train_acc_sum
#@save
def train_ch13(net, train_iter, test_iter, loss, trainer, num_epochs,
devices=d2l.try_all_gpus()):
"""用多GPU进行模型训练
Defined in :numref:`sec_image_augmentation`"""
timer, num_batches = d2l.Timer(), len(train_iter)
animator = d2l.Animator(xlabel='epoch', xlim=[1, num_epochs], ylim=[0, 1],
legend=['train loss', 'train acc', 'test acc'])
net = paddle.DataParallel(net)
for epoch in range(num_epochs):
# 4个维度：储存训练损失，训练准确度，实例数，特点数
metric = d2l.Accumulator(4)
for i, (features, labels) in enumerate(train_iter):
timer.start()
l, acc = train_batch_ch13(
net, features, labels, loss, trainer, devices)
metric.add(l, acc, labels.shape[0], labels.numel())
timer.stop()
if (i + 1) % (num_batches // 5) == 0 or i == num_batches - 1:
animator.add(epoch + (i + 1) / num_batches,
(metric[0] / metric[2], metric[1] / metric[3],
None))
test_acc = d2l.evaluate_accuracy_gpu(net, test_iter)
animator.add(epoch + 1, (None, None, test_acc))
print(f'loss {metric[0] / metric[2]:.3f}, train acc '
f'{metric[1] / metric[3]:.3f}, test acc {test_acc:.3f}')
print(f'{metric[2] * num_epochs / timer.sum():.1f} examples/sec on '
f'{str(devices)}')
.. raw:: html
.. raw:: html
现在，我们可以定义\ ``train_with_data_aug``\ 函数，使用图像增广来训练模型。该函数获取所有的GPU，并使用Adam作为训练的优化算法，将图像增广应用于训练集，最后调用刚刚定义的用于训练和评估模型的\ ``train_ch13``\ 函数。
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
batch_size, devices, net = 256, d2l.try_all_gpus(), d2l.resnet18(10)
net.initialize(init=init.Xavier(), ctx=devices)
def train_with_data_aug(train_augs, test_augs, net, lr=0.001):
train_iter = load_cifar10(True, train_augs, batch_size)
test_iter = load_cifar10(False, test_augs, batch_size)
loss = gluon.loss.SoftmaxCrossEntropyLoss()
trainer = gluon.Trainer(net.collect_params(), 'adam',
{'learning_rate': lr})
train_ch13(net, train_iter, test_iter, loss, trainer, 10, devices)
.. raw:: latex
\diilbookstyleoutputcell
.. parsed-literal::
:class: output
[07:08:22] ../src/storage/storage.cc:196: Using Pooled (Naive) StorageManager for GPU
[07:08:22] ../src/storage/storage.cc:196: Using Pooled (Naive) StorageManager for GPU
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
batch_size, devices, net = 256, d2l.try_all_gpus(), d2l.resnet18(10, 3)
def init_weights(m):
if type(m) in [nn.Linear, nn.Conv2d]:
nn.init.xavier_uniform_(m.weight)
net.apply(init_weights)
def train_with_data_aug(train_augs, test_augs, net, lr=0.001):
train_iter = load_cifar10(True, train_augs, batch_size)
test_iter = load_cifar10(False, test_augs, batch_size)
loss = nn.CrossEntropyLoss(reduction="none")
trainer = torch.optim.Adam(net.parameters(), lr=lr)
train_ch13(net, train_iter, test_iter, loss, trainer, 10, devices)
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
batch_size, devices, net = 256, d2l.try_all_gpus(), d2l.resnet18(10, 3)
def init_weights(m):
if type(m) in [nn.Linear, nn.Conv2D]:
nn.initializer.XavierUniform(m.weight)
net.apply(init_weights)
def train_with_data_aug(train_augs, test_augs, net, lr=0.001):
train_iter = load_cifar10(True, train_augs, batch_size)
test_iter = load_cifar10(False, test_augs, batch_size)
loss = nn.CrossEntropyLoss(reduction="none")
trainer = paddle.optimizer.Adam(learning_rate=lr, parameters=net.parameters())
train_ch13(net, train_iter, test_iter, loss, trainer, 10, devices[:1])
.. raw:: latex
\diilbookstyleoutputcell
.. parsed-literal::
:class: output
W0818 09:18:14.902511 73688 gpu_resources.cc:61] Please NOTE: device: 0, GPU Compute Capability: 7.0, Driver API Version: 11.8, Runtime API Version: 11.8
W0818 09:18:14.932250 73688 gpu_resources.cc:91] device: 0, cuDNN Version: 8.7.
.. raw:: html
.. raw:: html
让我们使用基于随机左右翻转的图像增广来训练模型。
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
train_with_data_aug(train_augs, test_augs, net)
.. raw:: latex
\diilbookstyleoutputcell
.. parsed-literal::
:class: output
loss 0.173, train acc 0.940, test acc 0.855
2158.6 examples/sec on [gpu(0), gpu(1)]
.. figure:: output_image-augmentation_7d0887_175_1.svg
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
train_with_data_aug(train_augs, test_augs, net)
.. raw:: latex
\diilbookstyleoutputcell
.. parsed-literal::
:class: output
loss 0.173, train acc 0.941, test acc 0.854
4183.9 examples/sec on [device(type='cuda', index=0), device(type='cuda', index=1)]
.. figure:: output_image-augmentation_7d0887_178_1.svg
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
train_with_data_aug(train_augs, test_augs, net)
.. raw:: latex
\diilbookstyleoutputcell
.. parsed-literal::
:class: output
loss 0.166, train acc 0.944, test acc 0.936
3465.5 examples/sec on [Place(gpu:0)]
.. figure:: output_image-augmentation_7d0887_181_1.svg
.. raw:: html
.. raw:: html
小结
----
- 图像增广基于现有的训练数据生成随机图像，来提高模型的泛化能力。
- 为了在预测过程中得到确切的结果，我们通常对训练样本只进行图像增广，而在预测过程中不使用带随机操作的图像增广。
- 深度学习框架提供了许多不同的图像增广方法，这些方法可以被同时应用。
练习
----
1. 在不使用图像增广的情况下训练模型：\ ``train_with_data_aug(no_aug, no_aug)``\ 。比较使用和不使用图像增广的训练结果和测试精度。这个对比实验能支持图像增广可以减轻过拟合的论点吗？为什么？
2. 在基于CIFAR-10数据集的模型训练中结合多种不同的图像增广方法。它能提高测试准确性吗？
3. 参阅深度学习框架的在线文档。它还提供了哪些其他的图像增广方法？
.. raw:: html
.. raw:: html
`Discussions `__
.. raw:: html
.. raw:: html
`Discussions `__
.. raw:: html
.. raw:: html
`Discussions `__
.. raw:: html
.. raw:: html