如何自定义 Keras 图层名称并使其自动递增 layer.name

Question

我目前正在尝试使用名称为 cust_sig 的自定义激活来创建多层。但是当我尝试编译模型时，我得到了一个 ValueError，因为多个层具有相同的名称 cust_sig。我知道我可以手动更改每一层的名称，但想知道是否可以像对内置层那样自动将 _1, _2, ... 添加到名称中。可以在下面找到模型定义。

# Creating a model
from tensorflow.python.keras import keras
from tensorflow.python.keras.models import Model
from tensorflow.python.keras.layers import Dense

# Custom activation function
from tensorflow.python.keras.layers import Activation
from tensorflow.python.keras import backend as K
from keras.utils.generic_utils import get_custom_objects

def custom_activation(x):
    return (K.sigmoid(x) * 5) - 1

get_custom_objects().update({'custom_activation': Activation(custom_activation)})

data_format = 'channels_first'

spec_input = keras.layers.Input(shape=(1, 3, 256), name='spec')
x = keras.layers.Flatten(data_format)(spec_input)

for layer in range(3):
  x = Dense(512)(x)
  x = Activation('custom_activation', name='cust_sig')(x)

out = Dense(256, activation="sigmoid", name='out')(x)
model = Model(inputs=spec_input, outputs=out)

错误信息如下所示

Traceback (most recent call last):
  File "/home/xyz/anaconda3/envs/ctf/lib/python3.7/site-packages/tensorflow/python/training/tracking/base.py", line 457, in _method_wrapper
    result = method(self, *args, **kwargs)
  File "/home/xyz/anaconda3/envs/ctf/lib/python3.7/site-packages/tensorflow/python/keras/engine/network.py", line 315, in _init_graph_network
    self.inputs, self.outputs)
  File "/home/xyz/anaconda3/envs/ctf/lib/python3.7/site-packages/tensorflow/python/keras/engine/network.py", line 1861, in _map_graph_network
    str(all_names.count(name)) + ' times in the model. '
ValueError: The name "cust_sig" is used 3 times in the model. All layer names should be unique.

Answer 1

下面应该做的：

def custom_activation(x):
    return (K.sigmoid(x) * 5) - 1

class CustSig(Layer):
    def __init__(self, my_activation, **kwargs):
        super(CustSig, self).__init__(**kwargs)
        self.supports_masking = True
        self.activation = my_activation

    def call(self, inputs):
        return self.activation(inputs)

    def get_config(self):
        config = {'activation': activations.serialize(self.activation)}
        base_config = super(Activation, self).get_config()
        return dict(list(base_config.items()) + list(config.items()))

    def compute_output_shape(self, input_shape):
        return input_shape

解释:

自source code起，自动命名工作如下：

if not name:
  self._name = backend.unique_object_name(
      generic_utils.to_snake_case(self.__class__.__name__),
      zero_based=zero_based)
else:
  self._name = name

检查 Keras 图是否存在与您定义的对象同名的现有对象 - 如果存在，则继续递增 1，直到 none 匹配。问题是，您不能指定 name=，因为这会消除上述条件的自动命名。

唯一的解决方法可能是使用所需的名称定义您自己的自定义激活层 class 名称，如上所示，其表现如下：

ipt = Input(shape=(1, 3, 256), name='spec')
x   = Flatten('channels_last')(ipt)
for _ in range(3):
    x   = Dense(512)(x)
    x   = CustSig(custom_activation)(x)
out = Dense(256, activation='sigmoid', name='out')(x)

model = Model(ipt, out)

print(model.layers[3].name)
print(model.layers[5].name)
print(model.layers[7].name)

cust_sig
cust_sig_1
cust_sig_2

Answer 2

如果你查看Layer class的源代码，你可以找到这些决定图层名称的行。

if not name:
    prefix = self.__class__.__name__
    name = _to_snake_case(prefix) + '_' + str(K.get_uid(prefix))
self.name = name

K.get_uid(prefix) 将从图表中获取唯一 ID，这就是您看到 activation_1、activation_2.

的原因

如果你想对你的自定义激活函数有同样的效果，更好的方法是定义你自己的 class 继承自 Layer。

class MyAct(Layer):
    def __init__(self):
        super().__init__()

    def call(self, inputs):
        return (K.sigmoid(inputs) * 5) - 1 

spec_input = Input(shape=(10,10))
x = Flatten()(spec_input)
for layer in range(3):
    x = Dense(512)(x)
    x = MyAct()(x)

model = Model(spec_input, x)
model.summary()

输出

# Layer (type)                 Output Shape              Param #   
# =================================================================
# input_1 (InputLayer)         (None, 10, 10)            0         
# _________________________________________________________________
# flatten_1 (Flatten)          (None, 100)               0         
# _________________________________________________________________
# dense_1 (Dense)              (None, 512)               51712     
# _________________________________________________________________
# my_act_1 (MyAct)             (None, 512)               0         
# _________________________________________________________________
# dense_2 (Dense)              (None, 512)               262656    
# _________________________________________________________________
# my_act_2 (MyAct)             (None, 512)               0         
# _________________________________________________________________
# dense_3 (Dense)              (None, 512)               262656    
# _________________________________________________________________
# my_act_3 (MyAct)             (None, 512)               0         
# =================================================================
# Total params: 577,024
# Trainable params: 577,024
# Non-trainable params: 0

Answer 3

如果您想多次使用带数字后缀的 specific_name，请使用：

tf.get_default_graph().unique_name("specific_name")

或

tf.compat.v1.get_default_graph().unique_name("specific_name")

你的情况：

...
for layer in range(3):
  x = Dense(512)(x)
  x = Activation('custom_activation', name=tf.get_default_graph().unique_name("cust_sig"))(x)
...

如何自定义 Keras 图层名称并使其自动递增 layer.name

How to customize Keras layer names and also have it automatically increment layer.name

python

tensorflow

keras

activation-function