Keras—embedding嵌入层的用法详解

最近在工作中进行了NLP的内容，使用的还是Keras中embedding的词嵌入来做的。

Keras中embedding层做一下介绍。

中文文档地址：https://keras.io/zh/layers/embeddings/

参数如下：

其中参数重点有input_dim,output_dim,非必选参数input_length.

初始化方法参数设置后面会单独总结一下。

demo使用预训练（使用百度百科（word2vec）的语料库）参考

embedding使用的demo参考：

def create_embedding(word_index, num_words, word2vec_model):
embedding_matrix = np.zeros((num_words, EMBEDDING_DIM))
for word, i in word_index.items():
 try:
  embedding_vector = word2vec_model[word]
  embedding_matrix[i] = embedding_vector
 except:
  continue
return embedding_matrix

#word_index:词典（统计词转换为索引）
#num_word:词典长度+1
#word2vec_model:词向量的model

加载词向量model的方法：

def pre_load_embedding_model(model_file):
# model = gensim.models.Word2Vec.load(model_file)
# model = gensim.models.Word2Vec.load(model_file,binary=True)
model = gensim.models.KeyedVectors.load_word2vec_format(model_file)
return model

model中Embedding层的设置（注意参数，Input层的输入，初始化方法）：

embedding_matrix = create_embedding(word_index, num_words, word2vec_model)

embedding_layer = Embedding(num_words,
       EMBEDDING_DIM,
       embeddings_initializer=Constant(embedding_matrix),
       input_length=MAX_SEQUENCE_LENGTH,
       trainable=False)
sequence_input = Input(shape=(MAX_SEQUENCE_LENGTH,), dtype='int32')
embedded_sequences = embedding_layer(sequence_input)

embedding层的初始化设置

keras embeding设置初始值的两种方式

随机初始化Embedding

from keras.models import Sequential
from keras.layers import Embedding
import numpy as np

model = Sequential()
model.add(Embedding(1000, 64, input_length=10))
# the model will take as input an integer matrix of size (batch, input_length).
# the largest integer (i.e. word index) in the input should be no larger than 999 (vocabulary size).
# now model.output_shape == (None, 10, 64), where None is the batch dimension.

input_array = np.random.randint(1000, size=(32, 10))

model.compile('rmsprop', 'mse')
output_array = model.predict(input_array)
print(output_array)
assert output_array.shape == (32, 10, 64)

使用weights参数指明embedding初始值

import numpy as np
import keras

m = keras.models.Sequential()
"""
可以通过weights参数指定初始的weights参数
因为Embedding层是不可导的
梯度东流至此回,所以把embedding放在中间层是没有意义的,emebedding只能作为第一层
注意weights到embeddings的绑定过程很复杂，weights是一个列表
"""
embedding = keras.layers.Embedding(input_dim=3, output_dim=2, input_length=1, weights=[np.arange(3 * 2).reshape((3, 2))], mask_zero=True)
m.add(embedding) # 一旦add，就会自动调用embedding的build函数,
print(keras.backend.get_value(embedding.embeddings))
m.compile(keras.optimizers.RMSprop(), keras.losses.mse)
print(m.predict([1, 2, 2, 1, 2, 0]))
print(m.get_layer(index=0).get_weights())
print(keras.backend.get_value(embedding.embeddings))

给embedding设置初始值的第二种方式：使用initializer

import numpy as np
import keras

m = keras.models.Sequential()
"""
可以通过weights参数指定初始的weights参数
因为Embedding层是不可导的
梯度东流至此回,所以把embedding放在中间层是没有意义的,emebedding只能作为第一层
给embedding设置权值的第二种方式，使用constant_initializer
"""
embedding = keras.layers.Embedding(input_dim=3, output_dim=2, input_length=1, embeddings_initializer=keras.initializers.constant(np.arange(3 * 2, dtype=np.float32).reshape((3, 2))))
m.add(embedding)
print(keras.backend.get_value(embedding.embeddings))
m.compile(keras.optimizers.RMSprop(), keras.losses.mse)
print(m.predict([1, 2, 2, 1, 2]))
print(m.get_layer(index=0).get_weights())
print(keras.backend.get_value(embedding.embeddings))

关键的难点在于理清weights是怎么传入到embedding.embeddings张量里面去的。

Embedding是一个层，继承自Layer，Layer有weights参数，weights参数是一个list，里面的元素都是numpy数组。在调用Layer的构造函数的时候，weights参数就被存储到了_initial_weights变量

basic_layer.py 之Layer类

 if 'weights' in kwargs:
  self._initial_weights = kwargs['weights']
 else:
  self._initial_weights = None

当把Embedding层添加到模型中、跟模型的上一层进行拼接的时候，会调用layer(上一层)函数，此处layer是Embedding实例，Embedding是一个继承了Layer的类，Embedding类没有重写__call__()方法，Layer实现了__call__()方法。

父类Layer的__call__方法调用子类的call()方法来获取结果。

所以最终调用的是Layer.__call__()。在这个方法中，会自动检测该层是否build过（根据self.built布尔变量）。

Layer.__call__函数非常重要。

def __call__(self, inputs, **kwargs):
 """Wrapper around self.call(), for handling internal references.
 If a Keras tensor is passed:
  - We call self._add_inbound_node().
  - If necessary, we `build` the layer to match
   the _keras_shape of the input(s).
  - We update the _keras_shape of every input tensor with
   its new shape (obtained via self.compute_output_shape).
   This is done as part of _add_inbound_node().
  - We update the _keras_history of the output tensor(s)
   with the current layer.
   This is done as part of _add_inbound_node().
 # Arguments
  inputs: Can be a tensor or list/tuple of tensors.
  **kwargs: Additional keyword arguments to be passed to `call()`.
 # Returns
  Output of the layer's `call` method.
 # Raises
  ValueError: in case the layer is missing shape information
   for its `build` call.
 """
 if isinstance(inputs, list):
  inputs = inputs[:]
 with K.name_scope(self.name):
  # Handle laying building (weight creating, input spec locking).
  if not self.built:#如果未曾build，那就要先执行build再调用call函数
   # Raise exceptions in case the input is not compatible
   # with the input_spec specified in the layer constructor.
   self.assert_input_compatibility(inputs)

# Collect input shapes to build layer.
   input_shapes = []
   for x_elem in to_list(inputs):
    if hasattr(x_elem, '_keras_shape'):
     input_shapes.append(x_elem._keras_shape)
    elif hasattr(K, 'int_shape'):
     input_shapes.append(K.int_shape(x_elem))
    else:
     raise ValueError('You tried to call layer "' +
          self.name +
          '". This layer has no information'
          ' about its expected input shape, '
          'and thus cannot be built. '
          'You can build it manually via: '
          '`layer.build(batch_input_shape)`')
   self.build(unpack_singleton(input_shapes))
   self.built = True#这句话其实有些多余，因为self.build函数已经把built置为True了

# Load weights that were specified at layer instantiation.
   if self._initial_weights is not None:#如果传入了weights，把weights参数赋值到每个变量，此处会覆盖上面的self.build函数中的赋值。
    self.set_weights(self._initial_weights)

# Raise exceptions in case the input is not compatible
  # with the input_spec set at build time.
  self.assert_input_compatibility(inputs)

# Handle mask propagation.
  previous_mask = _collect_previous_mask(inputs)
  user_kwargs = copy.copy(kwargs)
  if not is_all_none(previous_mask):
   # The previous layer generated a mask.
   if has_arg(self.call, 'mask'):
    if 'mask' not in kwargs:
     # If mask is explicitly passed to __call__,
     # we should override the default mask.
     kwargs['mask'] = previous_mask
  # Handle automatic shape inference (only useful for Theano).
  input_shape = _collect_input_shape(inputs)

# Actually call the layer,
  # collecting output(s), mask(s), and shape(s).
  output = self.call(inputs, **kwargs)
  output_mask = self.compute_mask(inputs, previous_mask)

# If the layer returns tensors from its inputs, unmodified,
  # we copy them to avoid loss of tensor metadata.
  output_ls = to_list(output)
  inputs_ls = to_list(inputs)
  output_ls_copy = []
  for x in output_ls:
   if x in inputs_ls:
    x = K.identity(x)
   output_ls_copy.append(x)
  output = unpack_singleton(output_ls_copy)

# Inferring the output shape is only relevant for Theano.
  if all([s is not None
    for s in to_list(input_shape)]):
   output_shape = self.compute_output_shape(input_shape)
  else:
   if isinstance(input_shape, list):
    output_shape = [None for _ in input_shape]
   else:
    output_shape = None

if (not isinstance(output_mask, (list, tuple)) and
    len(output_ls) > 1):
   # Augment the mask to match the length of the output.
   output_mask = [output_mask] * len(output_ls)

# Add an inbound node to the layer, so that it keeps track
  # of the call and of all new variables created during the call.
  # This also updates the layer history of the output tensor(s).
  # If the input tensor(s) had not previous Keras history,
  # this does nothing.
  self._add_inbound_node(input_tensors=inputs,
        output_tensors=output,
        input_masks=previous_mask,
        output_masks=output_mask,
        input_shapes=input_shape,
        output_shapes=output_shape,
        arguments=user_kwargs)

# Apply activity regularizer if any:
  if (hasattr(self, 'activity_regularizer') and
    self.activity_regularizer is not None):
   with K.name_scope('activity_regularizer'):
    regularization_losses = [
     self.activity_regularizer(x)
     for x in to_list(output)]
   self.add_loss(regularization_losses,
       inputs=to_list(inputs))
 return output

如果没有build过，会自动调用Embedding类的build()函数。Embedding.build()这个函数并不会去管weights，如果它使用的initializer没有传入，self.embeddings_initializer会变成随机初始化。

如果传入了，那么在这一步就能够把weights初始化好。

如果同时传入embeddings_initializer和weights参数，那么weights参数稍后会把Embedding#embeddings覆盖掉。

embedding.py Embedding类的build函数

def build(self, input_shape):
 self.embeddings = self.add_weight(
  shape=(self.input_dim, self.output_dim),
  initializer=self.embeddings_initializer,
  name='embeddings',
  regularizer=self.embeddings_regularizer,
  constraint=self.embeddings_constraint,
  dtype=self.dtype)
 self.built = True

综上，在keras中，使用weights给Layer的变量赋值是一个比较通用的方法，但是不够直观。keras鼓励多多使用明确的initializer，而尽量不要触碰weights。

来源：https://blog.csdn.net/qq_33472765/article/details/86561245