13.11. 全卷积网络¶ Open the notebook in SageMaker Studio Lab
如 13.9节中所介绍的那样,语义分割是对图像中的每个像素分类。 全卷积网络(fully convolutional network,FCN)采用卷积神经网络实现了从图像像素到像素类别的变换 (Long et al., 2015)。 与我们之前在图像分类或目标检测部分介绍的卷积神经网络不同,全卷积网络将中间层特征图的高和宽变换回输入图像的尺寸:这是通过在 13.10节中引入的转置卷积(transposed convolution)实现的。 因此,输出的类别预测与输入图像在像素级别上具有一一对应关系:通道维的输出即该位置对应像素的类别预测。
%matplotlib inline
from mxnet import gluon, image, init, np, npx
from mxnet.gluon import nn
from d2l import mxnet as d2l
npx.set_np()
%matplotlib inline
import torch
import torchvision
from torch import nn
from torch.nn import functional as F
from d2l import torch as d2l
%matplotlib inline
import warnings
from d2l import paddle as d2l
warnings.filterwarnings("ignore")
import paddle
import paddle.vision as paddlevision
from paddle import nn
from paddle.nn import functional as F
13.11.1. 构造模型¶
下面我们了解一下全卷积网络模型最基本的设计。 如 图13.11.1所示,全卷积网络先使用卷积神经网络抽取图像特征,然后通过\(1\times 1\)卷积层将通道数变换为类别个数,最后在 13.10节中通过转置卷积层将特征图的高和宽变换为输入图像的尺寸。 因此,模型输出与输入图像的高和宽相同,且最终输出通道包含了该空间位置像素的类别预测。
下面,我们使用在ImageNet数据集上预训练的ResNet-18模型来提取图像特征,并将该网络记为pretrained_net
。
ResNet-18模型的最后几层包括全局平均汇聚层和全连接层,然而全卷积网络中不需要它们。
pretrained_net = gluon.model_zoo.vision.resnet18_v2(pretrained=True)
pretrained_net.features[-3:], pretrained_net.output
[07:25:11] ../src/storage/storage.cc:196: Using Pooled (Naive) StorageManager for CPU
(HybridSequential(
(0): Activation(relu)
(1): GlobalAvgPool2D(size=(1, 1), stride=(1, 1), padding=(0, 0), ceil_mode=True, global_pool=True, pool_type=avg, layout=NCHW)
(2): Flatten
),
Dense(512 -> 1000, linear))
pretrained_net = torchvision.models.resnet18(pretrained=True)
list(pretrained_net.children())[-3:]
Downloading: "https://download.pytorch.org/models/resnet18-f37072fd.pth" to /home/ci/.cache/torch/hub/checkpoints/resnet18-f37072fd.pth
0%| | 0.00/44.7M [00:00<?, ?B/s]
[Sequential(
(0): BasicBlock(
(conv1): Conv2d(256, 512, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
(bn1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(relu): ReLU(inplace=True)
(conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn2): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(downsample): Sequential(
(0): Conv2d(256, 512, kernel_size=(1, 1), stride=(2, 2), bias=False)
(1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
)
(1): BasicBlock(
(conv1): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(relu): ReLU(inplace=True)
(conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn2): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
),
AdaptiveAvgPool2d(output_size=(1, 1)),
Linear(in_features=512, out_features=1000, bias=True)]
pretrained_net = paddlevision.models.resnet18(pretrained=True)
list(pretrained_net.children())[-3:]
W0818 09:21:31.243886 91274 gpu_resources.cc:61] Please NOTE: device: 0, GPU Compute Capability: 7.0, Driver API Version: 11.8, Runtime API Version: 11.8
W0818 09:21:31.275069 91274 gpu_resources.cc:91] device: 0, cuDNN Version: 8.7.
92.3%
接下来,我们创建一个全卷积网络net
。
它复制了ResNet-18中大部分的预训练层,除了最后的全局平均汇聚层和最接近输出的全连接层。
net = nn.HybridSequential()
for layer in pretrained_net.features[:-2]:
net.add(layer)
net = nn.Sequential(*list(pretrained_net.children())[:-2])
net = nn.Sequential(*list(pretrained_net.children())[:-2])
给定高度为320和宽度为480的输入,net
的前向传播将输入的高和宽减小至原来的\(1/32\),即10和15。
X = np.random.uniform(size=(1, 3, 320, 480))
net(X).shape
(1, 512, 10, 15)
X = torch.rand(size=(1, 3, 320, 480))
net(X).shape
torch.Size([1, 512, 10, 15])
X = paddle.rand(shape=(1, 3, 320, 480))
net(X).shape
[1, 512, 10, 15]
接下来使用\(1\times1\)卷积层将输出通道数转换为Pascal VOC2012数据集的类数(21类)。 最后需要将特征图的高度和宽度增加32倍,从而将其变回输入图像的高和宽。 回想一下 6.3节中卷积层输出形状的计算方法: 由于\((320-64+16\times2+32)/32=10\)且\((480-64+16\times2+32)/32=15\),我们构造一个步幅为\(32\)的转置卷积层,并将卷积核的高和宽设为\(64\),填充为\(16\)。 我们可以看到如果步幅为\(s\),填充为\(s/2\)(假设\(s/2\)是整数)且卷积核的高和宽为\(2s\),转置卷积核会将输入的高和宽分别放大\(s\)倍。
num_classes = 21
net.add(nn.Conv2D(num_classes, kernel_size=1),
nn.Conv2DTranspose(
num_classes, kernel_size=64, padding=16, strides=32))
num_classes = 21
net.add_module('final_conv', nn.Conv2d(512, num_classes, kernel_size=1))
net.add_module('transpose_conv', nn.ConvTranspose2d(num_classes, num_classes,
kernel_size=64, padding=16, stride=32))
num_classes = 21
net.add_sublayer('final_conv', nn.Conv2D(512, num_classes, kernel_size=1))
net.add_sublayer('transpose_conv', nn.Conv2DTranspose(num_classes, num_classes,
kernel_size=64, padding=16, stride=32))
Conv2DTranspose(21, 21, kernel_size=[64, 64], stride=[32, 32], padding=16, data_format=NCHW)
13.11.2. 初始化转置卷积层¶
在图像处理中,我们有时需要将图像放大,即上采样(upsampling)。 双线性插值(bilinear interpolation) 是常用的上采样方法之一,它也经常用于初始化转置卷积层。
为了解释双线性插值,假设给定输入图像,我们想要计算上采样输出图像上的每个像素。
将输出图像的坐标\((x,y)\)映射到输入图像的坐标\((x',y')\)上。 例如,根据输入与输出的尺寸之比来映射。 请注意,映射后的\(x′\)和\(y′\)是实数。
在输入图像上找到离坐标\((x',y')\)最近的4个像素。
输出图像在坐标\((x,y)\)上的像素依据输入图像上这4个像素及其与\((x',y')\)的相对距离来计算。
双线性插值的上采样可以通过转置卷积层实现,内核由以下bilinear_kernel
函数构造。
限于篇幅,我们只给出bilinear_kernel
函数的实现,不讨论算法的原理。
def bilinear_kernel(in_channels, out_channels, kernel_size):
factor = (kernel_size + 1) // 2
if kernel_size % 2 == 1:
center = factor - 1
else:
center = factor - 0.5
og = (np.arange(kernel_size).reshape(-1, 1),
np.arange(kernel_size).reshape(1, -1))
filt = (1 - np.abs(og[0] - center) / factor) * \
(1 - np.abs(og[1] - center) / factor)
weight = np.zeros((in_channels, out_channels, kernel_size, kernel_size))
weight[range(in_channels), range(out_channels), :, :] = filt
return np.array(weight)
def bilinear_kernel(in_channels, out_channels, kernel_size):
factor = (kernel_size + 1) // 2
if kernel_size % 2 == 1:
center = factor - 1
else:
center = factor - 0.5
og = (torch.arange(kernel_size).reshape(-1, 1),
torch.arange(kernel_size).reshape(1, -1))
filt = (1 - torch.abs(og[0] - center) / factor) * \
(1 - torch.abs(og[1] - center) / factor)
weight = torch.zeros((in_channels, out_channels,
kernel_size, kernel_size))
weight[range(in_channels), range(out_channels), :, :] = filt
return weight
def bilinear_kernel(in_channels, out_channels, kernel_size):
factor = (kernel_size + 1) // 2
if kernel_size % 2 == 1:
center = factor - 1
else:
center = factor - 0.5
og = (paddle.arange(kernel_size).reshape([-1, 1]),
paddle.arange(kernel_size).reshape([1, -1]))
filt = (1 - paddle.abs(og[0] - center) / factor) * \
(1 - paddle.abs(og[1] - center) / factor)
weight = paddle.zeros((in_channels, out_channels,
kernel_size, kernel_size))
weight[range(in_channels), range(out_channels), :, :] = filt
return weight
让我们用双线性插值的上采样实验它由转置卷积层实现。
我们构造一个将输入的高和宽放大2倍的转置卷积层,并将其卷积核用bilinear_kernel
函数初始化。
conv_trans = nn.Conv2DTranspose(3, kernel_size=4, padding=1, strides=2)
conv_trans.initialize(init.Constant(bilinear_kernel(3, 3, 4)))
conv_trans = nn.ConvTranspose2d(3, 3, kernel_size=4, padding=1, stride=2,
bias=False)
conv_trans.weight.data.copy_(bilinear_kernel(3, 3, 4));
conv_trans = nn.Conv2DTranspose(3, 3, kernel_size=4, padding=1, stride=2,
bias_attr=False)
conv_trans.weight.set_value(bilinear_kernel(3, 3, 4));
读取图像X
,将上采样的结果记作Y
。为了打印图像,我们需要调整通道维的位置。
img = image.imread('../img/catdog.jpg')
X = np.expand_dims(img.astype('float32').transpose(2, 0, 1), axis=0) / 255
Y = conv_trans(X)
out_img = Y[0].transpose(1, 2, 0)
img = torchvision.transforms.ToTensor()(d2l.Image.open('../img/catdog.jpg'))
X = img.unsqueeze(0)
Y = conv_trans(X)
out_img = Y[0].permute(1, 2, 0).detach()
img = paddlevision.transforms.ToTensor()(d2l.Image.open('../img/catdog.jpg'))
X = img.unsqueeze(0)
Y = conv_trans(X)
out_img = Y[0].transpose([1, 2, 0]).detach()
可以看到,转置卷积层将图像的高和宽分别放大了2倍。 除了坐标刻度不同,双线性插值放大的图像和在 13.3节中打印出的原图看上去没什么两样。
d2l.set_figsize()
print('input image shape:', img.shape)
d2l.plt.imshow(img.asnumpy());
print('output image shape:', out_img.shape)
d2l.plt.imshow(out_img.asnumpy());
input image shape: (561, 728, 3)
output image shape: (1122, 1456, 3)
d2l.set_figsize()
print('input image shape:', img.permute(1, 2, 0).shape)
d2l.plt.imshow(img.permute(1, 2, 0));
print('output image shape:', out_img.shape)
d2l.plt.imshow(out_img);
input image shape: torch.Size([561, 728, 3])
output image shape: torch.Size([1122, 1456, 3])
d2l.set_figsize()
print('input image shape:', img.transpose([1, 2, 0]).shape)
d2l.plt.imshow(img.transpose([1, 2, 0]));
print('output image shape:', out_img.shape)
d2l.plt.imshow(out_img);
input image shape: [561, 728, 3]
output image shape: [1122, 1456, 3]
全卷积网络用双线性插值的上采样初始化转置卷积层。对于\(1\times 1\)卷积层,我们使用Xavier初始化参数。
W = bilinear_kernel(num_classes, num_classes, 64)
net[-1].initialize(init.Constant(W))
net[-2].initialize(init=init.Xavier())
W = bilinear_kernel(num_classes, num_classes, 64)
net.transpose_conv.weight.data.copy_(W);
W = bilinear_kernel(num_classes, num_classes, 64)
net.transpose_conv.weight.set_value(W);
13.11.3. 读取数据集¶
我们用 13.9节中介绍的语义分割读取数据集。 指定随机裁剪的输出图像的形状为\(320\times 480\):高和宽都可以被\(32\)整除。
batch_size, crop_size = 32, (320, 480)
train_iter, test_iter = d2l.load_data_voc(batch_size, crop_size)
read 1114 examples
read 1078 examples
batch_size, crop_size = 32, (320, 480)
train_iter, test_iter = d2l.load_data_voc(batch_size, crop_size)
read 1114 examples
read 1078 examples
import os
def load_data_voc(batch_size, crop_size):
"""加载VOC语义分割数据集
Defined in :numref:`sec_semantic_segmentation`"""
voc_dir = d2l.download_extract('voc2012', os.path.join(
'VOCdevkit', 'VOC2012'))
train_iter = paddle.io.DataLoader(
d2l.VOCSegDataset(True, crop_size, voc_dir), batch_size=batch_size,
shuffle=True, return_list=True, drop_last=True, num_workers=0)
test_iter = paddle.io.DataLoader(
d2l.VOCSegDataset(False, crop_size, voc_dir), batch_size=batch_size,
drop_last=True, return_list=True, num_workers=0)
return train_iter, test_iter
batch_size, crop_size = 32, (320, 480)
train_iter, test_iter = load_data_voc(batch_size, crop_size)
read 1114 examples
read 1078 examples
13.11.4. 训练¶
现在我们可以训练全卷积网络了。 这里的损失函数和准确率计算与图像分类中的并没有本质上的不同,因为我们使用转置卷积层的通道来预测像素的类别,所以需要在损失计算中指定通道维。 此外,模型基于每个像素的预测类别是否正确来计算准确率。
num_epochs, lr, wd, devices = 5, 0.1, 1e-3, d2l.try_all_gpus()
loss = gluon.loss.SoftmaxCrossEntropyLoss(axis=1)
net.collect_params().reset_ctx(devices)
trainer = gluon.Trainer(net.collect_params(), 'sgd',
{'learning_rate': lr, 'wd': wd})
d2l.train_ch13(net, train_iter, test_iter, loss, trainer, num_epochs, devices)
loss 0.325, train acc 0.892, test acc 0.846
131.8 examples/sec on [gpu(0), gpu(1)]
def loss(inputs, targets):
return F.cross_entropy(inputs, targets, reduction='none').mean(1).mean(1)
num_epochs, lr, wd, devices = 5, 0.001, 1e-3, d2l.try_all_gpus()
trainer = torch.optim.SGD(net.parameters(), lr=lr, weight_decay=wd)
d2l.train_ch13(net, train_iter, test_iter, loss, trainer, num_epochs, devices)
loss 0.443, train acc 0.863, test acc 0.848
254.0 examples/sec on [device(type='cuda', index=0), device(type='cuda', index=1)]
def loss(inputs, targets):
return F.cross_entropy(inputs.transpose([0, 2, 3, 1]), targets, reduction='none').mean(1).mean(1)
num_epochs, lr, wd, devices = 5, 0.001, 1e-3, d2l.try_all_gpus()
trainer = paddle.optimizer.SGD(learning_rate=lr, parameters=net.parameters(), weight_decay=wd)
d2l.train_ch13(net, train_iter, test_iter, loss, trainer, num_epochs, devices[:1])
loss 0.370, train acc 0.882, test acc 0.841
210.8 examples/sec on [Place(gpu:0)]
13.11.5. 预测¶
在预测时,我们需要将输入图像在各个通道做标准化,并转成卷积神经网络所需要的四维输入格式。
def predict(img):
X = test_iter._dataset.normalize_image(img)
X = np.expand_dims(X.transpose(2, 0, 1), axis=0)
pred = net(X.as_in_ctx(devices[0])).argmax(axis=1)
return pred.reshape(pred.shape[1], pred.shape[2])
def predict(img):
X = test_iter.dataset.normalize_image(img).unsqueeze(0)
pred = net(X.to(devices[0])).argmax(dim=1)
return pred.reshape(pred.shape[1], pred.shape[2])
def predict(img):
X = paddle.to_tensor(test_iter.dataset.normalize_image(img),dtype='float32').unsqueeze(0)
pred = net(X).argmax(axis=1)
return pred.reshape([pred.shape[1], pred.shape[2]])
为了可视化预测的类别给每个像素,我们将预测类别映射回它们在数据集中的标注颜色。
def label2image(pred):
colormap = np.array(d2l.VOC_COLORMAP, ctx=devices[0], dtype='uint8')
X = pred.astype('int32')
return colormap[X, :]
def label2image(pred):
colormap = torch.tensor(d2l.VOC_COLORMAP, device=devices[0])
X = pred.long()
return colormap[X, :]
def label2image(pred):
colormap = paddle.to_tensor(d2l.VOC_COLORMAP)
X = pred.astype(paddle.int32)
return colormap[X]
测试数据集中的图像大小和形状各异。
由于模型使用了步幅为32的转置卷积层,因此当输入图像的高或宽无法被32整除时,转置卷积层输出的高或宽会与输入图像的尺寸有偏差。
为了解决这个问题,我们可以在图像中截取多块高和宽为32的整数倍的矩形区域,并分别对这些区域中的像素做前向传播。
请注意,这些区域的并集需要完整覆盖输入图像。
当一个像素被多个区域所覆盖时,它在不同区域前向传播中转置卷积层输出的平均值可以作为softmax
运算的输入,从而预测类别。
为简单起见,我们只读取几张较大的测试图像,并从图像的左上角开始截取形状为\(320\times480\)的区域用于预测。 对于这些测试图像,我们逐一打印它们截取的区域,再打印预测结果,最后打印标注的类别。
voc_dir = d2l.download_extract('voc2012', 'VOCdevkit/VOC2012')
test_images, test_labels = d2l.read_voc_images(voc_dir, False)
n, imgs = 4, []
for i in range(n):
crop_rect = (0, 0, 480, 320)
X = image.fixed_crop(test_images[i], *crop_rect)
pred = label2image(predict(X))
imgs += [X, pred, image.fixed_crop(test_labels[i], *crop_rect)]
d2l.show_images(imgs[::3] + imgs[1::3] + imgs[2::3], 3, n, scale=2);
voc_dir = d2l.download_extract('voc2012', 'VOCdevkit/VOC2012')
test_images, test_labels = d2l.read_voc_images(voc_dir, False)
n, imgs = 4, []
for i in range(n):
crop_rect = (0, 0, 320, 480)
X = torchvision.transforms.functional.crop(test_images[i], *crop_rect)
pred = label2image(predict(X))
imgs += [X.permute(1,2,0), pred.cpu(),
torchvision.transforms.functional.crop(
test_labels[i], *crop_rect).permute(1,2,0)]
d2l.show_images(imgs[::3] + imgs[1::3] + imgs[2::3], 3, n, scale=2);
voc_dir = d2l.download_extract('voc2012', 'VOCdevkit/VOC2012')
test_images, test_labels = d2l.read_voc_images(voc_dir, False)
n, imgs = 4, []
for i in range(n):
crop_rect = (0, 0, 320, 480)
X = paddlevision.transforms.crop(test_images[i], *crop_rect)
pred = label2image(predict(X))
imgs += [X.transpose([1,2,0]).astype('uint8'), pred,
paddlevision.transforms.crop(
test_labels[i], *crop_rect).transpose([1, 2, 0]).astype("uint8")]
d2l.show_images(imgs[::3] + imgs[1::3] + imgs[2::3], 3, n, scale=2);
13.11.6. 小结¶
全卷积网络先使用卷积神经网络抽取图像特征,然后通过\(1\times 1\)卷积层将通道数变换为类别个数,最后通过转置卷积层将特征图的高和宽变换为输入图像的尺寸。
在全卷积网络中,我们可以将转置卷积层初始化为双线性插值的上采样。
13.11.7. 练习¶
如果将转置卷积层改用Xavier随机初始化,结果有什么变化?
调节超参数,能进一步提升模型的精度吗?
预测测试图像中所有像素的类别。
最初的全卷积网络的论文中 (Long et al., 2015)还使用了某些卷积神经网络中间层的输出。试着实现这个想法。