.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
pretrained_net = gluon.model_zoo.vision.resnet18_v2(pretrained=True)
pretrained_net.features[-3:], pretrained_net.output
.. raw:: latex
\diilbookstyleoutputcell
.. parsed-literal::
:class: output
[07:25:11] ../src/storage/storage.cc:196: Using Pooled (Naive) StorageManager for CPU
.. raw:: latex
\diilbookstyleoutputcell
.. parsed-literal::
:class: output
(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))
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
pretrained_net = torchvision.models.resnet18(pretrained=True)
list(pretrained_net.children())[-3:]
.. raw:: latex
\diilbookstyleoutputcell
.. parsed-literal::
:class: output
Downloading: "https://download.pytorch.org/models/resnet18-f37072fd.pth" to /home/ci/.cache/torch/hub/checkpoints/resnet18-f37072fd.pth
.. raw:: latex
\diilbookstyleoutputcell
.. parsed-literal::
:class: output
0%| | 0.00/44.7M [00:00
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
pretrained_net = paddlevision.models.resnet18(pretrained=True)
list(pretrained_net.children())[-3:]
.. raw:: latex
\diilbookstyleoutputcell
.. parsed-literal::
:class: output
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%
.. raw:: html
.. raw:: html
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。
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
net = nn.HybridSequential()
for layer in pretrained_net.features[:-2]:
net.add(layer)
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
net = nn.Sequential(*list(pretrained_net.children())[:-2])
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
net = nn.Sequential(*list(pretrained_net.children())[:-2])
.. raw:: html
.. raw:: html
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\ :math:`1/32`\ ，即10和15。
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
X = np.random.uniform(size=(1, 3, 320, 480))
net(X).shape
.. raw:: latex
\diilbookstyleoutputcell
.. parsed-literal::
:class: output
(1, 512, 10, 15)
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
X = torch.rand(size=(1, 3, 320, 480))
net(X).shape
.. raw:: latex
\diilbookstyleoutputcell
.. parsed-literal::
:class: output
torch.Size([1, 512, 10, 15])
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
X = paddle.rand(shape=(1, 3, 320, 480))
net(X).shape
.. raw:: latex
\diilbookstyleoutputcell
.. parsed-literal::
:class: output
[1, 512, 10, 15]
.. raw:: html
.. raw:: html
接下来使用\ :math:`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 :numref:`sec_padding`\ 中卷积层输出形状的计算方法：
由于\ :math:`(320-64+16\times2+32)/32=10`\ 且\ :math:`(480-64+16\times2+32)/32=15`\ ，我们构造一个步幅为\ :math:`32`\ 的转置卷积层，并将卷积核的高和宽设为\ :math:`64`\ ，填充为\ :math:`16`\ 。
我们可以看到如果步幅为\ :math:`s`\ ，填充为\ :math:`s/2`\ （假设\ :math:`s/2`\ 是整数）且卷积核的高和宽为\ :math:`2s`\ ，转置卷积核会将输入的高和宽分别放大\ :math:`s`\ 倍。
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
num_classes = 21
net.add(nn.Conv2D(num_classes, kernel_size=1),
nn.Conv2DTranspose(
num_classes, kernel_size=64, padding=16, strides=32))
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
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))
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
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))
.. raw:: latex
\diilbookstyleoutputcell
.. parsed-literal::
:class: output
Conv2DTranspose(21, 21, kernel_size=[64, 64], stride=[32, 32], padding=16, data_format=NCHW)
.. raw:: html
.. raw:: html
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。
1. 将输出图像的坐标\ :math:`(x,y)`\ 映射到输入图像的坐标\ :math:`(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\ :math:`x′`\ 和\ :math:`y′`\ 是实数。
2. 在输入图像上找到离坐标\ :math:`(x',y')`\ 最近的4个像素。
3. 输出图像在坐标\ :math:`(x,y)`\ 上的像素依据输入图像上这4个像素及其与\ :math:`(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。
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
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)
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
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
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
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
.. raw:: html
.. raw:: html
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。
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
conv_trans = nn.Conv2DTranspose(3, kernel_size=4, padding=1, strides=2)
conv_trans.initialize(init.Constant(bilinear_kernel(3, 3, 4)))
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
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));
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
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));
.. raw:: html
.. raw:: html
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。
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
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)
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
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()
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
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()
.. raw:: html
.. raw:: html
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
:numref:`sec_bbox`\ 中打印出的原图看上去没什么两样。
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
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());
.. raw:: latex
\diilbookstyleoutputcell
.. parsed-literal::
:class: output
input image shape: (561, 728, 3)
output image shape: (1122, 1456, 3)
.. figure:: output_fcn_ce3435_99_1.svg
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
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);
.. raw:: latex
\diilbookstyleoutputcell
.. parsed-literal::
:class: output
input image shape: torch.Size([561, 728, 3])
output image shape: torch.Size([1122, 1456, 3])
.. figure:: output_fcn_ce3435_102_1.svg
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
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);
.. raw:: latex
\diilbookstyleoutputcell
.. parsed-literal::
:class: output
input image shape: [561, 728, 3]
output image shape: [1122, 1456, 3]
.. figure:: output_fcn_ce3435_105_1.svg
.. raw:: html
.. raw:: html
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\ :math:`1\times 1`\ 卷积层，我们使用Xavier初始化参数。
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
W = bilinear_kernel(num_classes, num_classes, 64)
net[-1].initialize(init.Constant(W))
net[-2].initialize(init=init.Xavier())
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
W = bilinear_kernel(num_classes, num_classes, 64)
net.transpose_conv.weight.data.copy_(W);
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
W = bilinear_kernel(num_classes, num_classes, 64)
net.transpose_conv.weight.set_value(W);
.. raw:: html
.. raw:: html
读取数据集
----------
我们用
:numref:`sec_semantic_segmentation`\ 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\ :math:`320\times 480`\ ：高和宽都可以被\ :math:`32`\ 整除。
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
batch_size, crop_size = 32, (320, 480)
train_iter, test_iter = d2l.load_data_voc(batch_size, crop_size)
.. raw:: latex
\diilbookstyleoutputcell
.. parsed-literal::
:class: output
read 1114 examples
read 1078 examples
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
batch_size, crop_size = 32, (320, 480)
train_iter, test_iter = d2l.load_data_voc(batch_size, crop_size)
.. raw:: latex
\diilbookstyleoutputcell
.. parsed-literal::
:class: output
read 1114 examples
read 1078 examples
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
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)
.. raw:: latex
\diilbookstyleoutputcell
.. parsed-literal::
:class: output
read 1114 examples
read 1078 examples
.. raw:: html
.. raw:: html
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。
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
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)
.. raw:: latex
\diilbookstyleoutputcell
.. parsed-literal::
:class: output
loss 0.325, train acc 0.892, test acc 0.846
131.8 examples/sec on [gpu(0), gpu(1)]
.. figure:: output_fcn_ce3435_135_1.svg
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
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)
.. raw:: latex
\diilbookstyleoutputcell
.. parsed-literal::
:class: output
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)]
.. figure:: output_fcn_ce3435_138_1.svg
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
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])
.. raw:: latex
\diilbookstyleoutputcell
.. parsed-literal::
:class: output
loss 0.370, train acc 0.882, test acc 0.841
210.8 examples/sec on [Place(gpu:0)]
.. figure:: output_fcn_ce3435_141_1.svg
.. raw:: html
.. raw:: html
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。
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
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])
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
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])
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
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]])
.. raw:: html
.. raw:: html
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。
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
def label2image(pred):
colormap = np.array(d2l.VOC_COLORMAP, ctx=devices[0], dtype='uint8')
X = pred.astype('int32')
return colormap[X, :]
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
def label2image(pred):
colormap = torch.tensor(d2l.VOC_COLORMAP, device=devices[0])
X = pred.long()
return colormap[X, :]
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
def label2image(pred):
colormap = paddle.to_tensor(d2l.VOC_COLORMAP)
X = pred.astype(paddle.int32)
return colormap[X]
.. raw:: html
.. raw:: html
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\ :math:`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。
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
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);
.. figure:: output_fcn_ce3435_171_0.svg
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
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);
.. figure:: output_fcn_ce3435_174_0.svg
.. raw:: html
.. raw:: html
.. raw:: latex
\diilbookstyleinputcell
.. code:: python
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);
.. figure:: output_fcn_ce3435_177_0.svg
.. raw:: html
.. raw:: html
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\ :math:`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。
练习
----
1. 如果将转置卷积层改用Xavier随机初始化，结果有什么变化？
2. 调节超参数，能进一步提升模型的精度吗？
3. 预测测试图像中所有像素的类别。
4. 最初的全卷积网络的论文中
:cite:`Long.Shelhamer.Darrell.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。
.. raw:: html
.. raw:: html
`Discussions `__
.. raw:: html
.. raw:: html
`Discussions `__
.. raw:: html
.. raw:: html
`Discussions `__
.. raw:: html
.. raw:: html