『MXNet』第四弹_Gluon自定义层 『MXNet』第四弹_Gluon自定义层
https://www.cnblogs.com/hellcat/p/9047618.html
一、不含参数层
通过继承Block自定义了一个将输入减掉均值的层：CenteredLayer类，并将层的计算放在forward函数里，
|
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
|
from mxnet import nd, gluon
from mxnet.gluon import nn
class CenteredLayer(nn.Block):
def __init__(self, **kwargs):
super(CenteredLayer, self).__init__(**kwargs)
def forward(self, x):
return x - x.mean()
# 直接使用这个层layer = CenteredLayer()
# layer(nd.array([1, 2, 3, 4, 5]))# 构建更复杂模型net = nn.Sequential()
net.add(nn.Dense(128))
net.add(nn.Dense(10))
net.add(CenteredLayer())# 初始化、运行……net.initialize()y = net(nd.random.uniform(shape=(4, 8)))
|
二、含参数层
注意，本节实现的自定义层不能自动推断输入尺寸，需要手动指定
见上节『MXNet』第三弹_Gluon模型参数在自定义层的时候我们常使用Block自带的ParameterDict类添加成员变量params，如下，
|
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
|
from mxnet import gluon
from mxnet.gluon import nn
class MyDense(nn.Block):
def __init__(self, units, in_units, **kwargs):
super(MyDense, self).__init__(**kwargs)
self.weight = self.params.get('weight', shape=(in_units, units))
self.bias = self.params.get('bias', shape=(units,))
def forward(self, x):
linear = nd.dot(x, self.weight.data()) + self.bias.data()
return nd.relu(linear)
# 实际运行dense = MyDense(5, in_units=10)
|
如果不想使用ParameterDict类则需要一下操作
|
1
2
3
|
# self.weight = self.params.get('weight', shape=(in_units, units))self.weight = gluon.Parameter('weight', shape=(in_units, units))
self.params.update({'weight':self.weight})
|
否则在net.initialize()初始化时是初始化不到ParameterDict外变量的。
有关这一点详见下面：
|
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
|
def __init__(self, conv_arch, dropout_keep_prob, **kwargs):
super(SSD, self).__init__(**kwargs)
self.vgg_conv = nn.Sequential()
self.vgg_conv.add(repeat(*conv_arch[0], pool=False))
[self.vgg_conv.add(repeat(*conv_arch[i])) for i in range(1, len(conv_arch))]
# 迭代器对象只能进行单次迭代，所以将之转化为tuple，否则识别参数处迭代后forward再次迭代直接跳出循环
# self.vgg_conv = tuple([repeat(*conv_arch[i])
# for i in range(len(conv_arch))])
# 只能识别实例属性直接为mx层函数或者mx序列对象的参数，如果使用其他容器，需要将参数收集进参数字典
# _ = [self.params.update(block.collect_params()) for block in self.vgg_conv]
def forward(self, x, feat_layers):
end_points = {'block0': x}
for (index, block) in enumerate(self.vgg_conv):
end_points.update({'block{:d}'.format(index+1): block(end_points['block{:d}'.format(index)])})
return end_points
|
属性对象是mxnet的对象时才能默认识别层中的参数，否则需要显式收集进self.params中。
测试代码：
|
1
2
3
4
5
6
7
8
|
if __name__ == '__main__':
ssd = SSD(conv_arch=((2, 64), (2, 128), (3, 256), (3, 512), (3, 512)),
dropout_keep_prob=0.5)
ssd.initialize()
X = mx.ndarray.random.uniform(shape=(1, 1, 304, 304))
import pprint as pp
pp.pprint([x[1].shape for x in ssd(X).items()])
|
自行验证即可。
一、不含参数层
通过继承Block自定义了一个将输入减掉均值的层：CenteredLayer类，并将层的计算放在forward函数里，
|
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
|
from mxnet import nd, gluon
from mxnet.gluon import nn
class CenteredLayer(nn.Block):
def __init__(self, **kwargs):
super(CenteredLayer, self).__init__(**kwargs)
def forward(self, x):
return x - x.mean()
# 直接使用这个层layer = CenteredLayer()
# layer(nd.array([1, 2, 3, 4, 5]))# 构建更复杂模型net = nn.Sequential()
net.add(nn.Dense(128))
net.add(nn.Dense(10))
net.add(CenteredLayer())# 初始化、运行……net.initialize()y = net(nd.random.uniform(shape=(4, 8)))
|
二、含参数层
注意，本节实现的自定义层不能自动推断输入尺寸，需要手动指定
见上节『MXNet』第三弹_Gluon模型参数在自定义层的时候我们常使用Block自带的ParameterDict类添加成员变量params，如下，
|
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
|
from mxnet import gluon
from mxnet.gluon import nn
class MyDense(nn.Block):
def __init__(self, units, in_units, **kwargs):
super(MyDense, self).__init__(**kwargs)
self.weight = self.params.get('weight', shape=(in_units, units))
self.bias = self.params.get('bias', shape=(units,))
def forward(self, x):
linear = nd.dot(x, self.weight.data()) + self.bias.data()
return nd.relu(linear)
# 实际运行dense = MyDense(5, in_units=10)
|
如果不想使用ParameterDict类则需要一下操作
|
1
2
3
|
# self.weight = self.params.get('weight', shape=(in_units, units))self.weight = gluon.Parameter('weight', shape=(in_units, units))
self.params.update({'weight':self.weight})
|
否则在net.initialize()初始化时是初始化不到ParameterDict外变量的。
有关这一点详见下面：
|
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
|
def __init__(self, conv_arch, dropout_keep_prob, **kwargs):
super(SSD, self).__init__(**kwargs)
self.vgg_conv = nn.Sequential()
self.vgg_conv.add(repeat(*conv_arch[0], pool=False))
[self.vgg_conv.add(repeat(*conv_arch[i])) for i in range(1, len(conv_arch))]
# 迭代器对象只能进行单次迭代，所以将之转化为tuple，否则识别参数处迭代后forward再次迭代直接跳出循环
# self.vgg_conv = tuple([repeat(*conv_arch[i])
# for i in range(len(conv_arch))])
# 只能识别实例属性直接为mx层函数或者mx序列对象的参数，如果使用其他容器，需要将参数收集进参数字典
# _ = [self.params.update(block.collect_params()) for block in self.vgg_conv]
def forward(self, x, feat_layers):
end_points = {'block0': x}
for (index, block) in enumerate(self.vgg_conv):
end_points.update({'block{:d}'.format(index+1): block(end_points['block{:d}'.format(index)])})
return end_points
|
属性对象是mxnet的对象时才能默认识别层中的参数，否则需要显式收集进self.params中。
测试代码：
|
1
2
3
4
5
6
7
8
|
if __name__ == '__main__':
ssd = SSD(conv_arch=((2, 64), (2, 128), (3, 256), (3, 512), (3, 512)),
dropout_keep_prob=0.5)
ssd.initialize()
X = mx.ndarray.random.uniform(shape=(1, 1, 304, 304))
import pprint as pp
pp.pprint([x[1].shape for x in ssd(X).items()])
|
自行验证即可。