LSTM 后跟均值池化

Question

我使用的是 Keras 1.0。我的问题与这个 (How to implement a Mean Pooling layer in Keras) 相同，但那里的答案对我来说似乎还不够。

我想实现这个网络：

以下代码不起作用：

sequence = Input(shape=(max_sent_len,), dtype='int32')
embedded = Embedding(vocab_size, word_embedding_size)(sequence)
lstm = LSTM(hidden_state_size, activation='sigmoid', inner_activation='hard_sigmoid', return_sequences=True)(embedded)
pool = AveragePooling1D()(lstm)
output = Dense(1, activation='sigmoid')(pool)

如果我没有设置return_sequences=True，当我调用AveragePooling1D()时会出现这个错误：

Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/PATH/keras/engine/topology.py", line 462, in __call__
    self.assert_input_compatibility(x)
  File "/PATH/keras/engine/topology.py", line 382, in assert_input_compatibility
    str(K.ndim(x)))
Exception: ('Input 0 is incompatible with layer averagepooling1d_6: expected ndim=3', ' found ndim=2')

否则，当我调用Dense() 时会出现此错误：

Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/PATH/keras/engine/topology.py", line 456, in __call__
    self.build(input_shapes[0])
  File "/fs/clip-arqat/mossaab/trec/liveqa/cmu/venv/lib/python2.7/site-packages/keras/layers/core.py", line 512, in build
    assert len(input_shape) == 2
AssertionError

Answer 1

我只是尝试实现与原始海报相同的模型，我使用的是Keras 2.0.3。当我使用GlobalAveragePooling1D 时，LSTM 之后的平均池化工作，只要确保 LSTM 层中的return_sequences=True 即可。试试看！

Answer 2

添加TimeDistributed(Dense(1)) 有帮助：

sequence = Input(shape=(max_sent_len,), dtype='int32')
embedded = Embedding(vocab_size, word_embedding_size)(sequence)
lstm = LSTM(hidden_state_size, activation='sigmoid', inner_activation='hard_sigmoid', return_sequences=True)(embedded)
distributed = TimeDistributed(Dense(1))(lstm)
pool = AveragePooling1D()(distributed)
output = Dense(1, activation='sigmoid')(pool)

Answer 3

我认为接受的答案基本上是错误的。在以下位置找到了解决方案： https://github.com/fchollet/keras/issues/2151 但是，它仅适用于 theano 后端。我已经修改了代码，使其同时支持 theano 和 tensorflow。

from keras.engine.topology import Layer, InputSpec
from keras import backend as T

class TemporalMeanPooling(Layer):
    """
This is a custom Keras layer. This pooling layer accepts the temporal
sequence output by a recurrent layer and performs temporal pooling,
looking at only the non-masked portion of the sequence. The pooling
layer converts the entire variable-length hidden vector sequence
into a single hidden vector, and then feeds its output to the Dense
layer.

input shape: (nb_samples, nb_timesteps, nb_features)
output shape: (nb_samples, nb_features)
"""
def __init__(self, **kwargs):
    super(TemporalMeanPooling, self).__init__(**kwargs)
    self.supports_masking = True
    self.input_spec = [InputSpec(ndim=3)]

def get_output_shape_for(self, input_shape):
    return (input_shape[0], input_shape[2])

def call(self, x, mask=None): #mask: (nb_samples, nb_timesteps)
    if mask is None:
        mask = T.mean(T.ones_like(x), axis=-1)
    ssum = T.sum(x,axis=-2) #(nb_samples, np_features)
    mask = T.cast(mask,T.floatx())
    rcnt = T.sum(mask,axis=-1,keepdims=True) #(nb_samples)
    return ssum/rcnt
    #return rcnt

def compute_mask(self, input, mask):
    return None

Answer 4

谢谢，我也遇到了这个问题，但我认为 TimeDistributed 层无法按您的意愿工作，您可以尝试 Luke Guye 的 TemporalMeanPooling 层，它对我有用。示例如下：

sequence = Input(shape=(max_sent_len,), dtype='int32')
embedded = Embedding(vocab_size, word_embedding_size)(sequence)
lstm = LSTM(hidden_state_size, return_sequences=True)(embedded)
pool = TemporalMeanPooling()(lstm)
output = Dense(1, activation='sigmoid')(pool)

Answer 5

聚会迟到了，但tf.keras.layers.AveragePooling1D 与合适的pool_size 参数似乎也返回了正确的结果。

处理bobchennan 在此issue 上共享的示例。

# create sample data
A=np.array([[1,2,3],[4,5,6],[0,0,0],[0,0,0],[0,0,0]])
B=np.array([[1,3,0],[4,0,0],[0,0,1],[0,0,0],[0,0,0]])
C=np.array([A,B]).astype("float32")
# expected answer (for temporal mean)
np.mean(C, axis=1)

输出是

array([[1. , 1.4, 1.8],
       [1. , 0.6, 0.2]], dtype=float32)

现在使用AveragePooling1D，

model = keras.models.Sequential(
        tf.keras.layers.AveragePooling1D(pool_size=5)
)
model.predict(C)

输出是，

array([[[1. , 1.4, 1.8]],
       [[1. , 0.6, 0.2]]], dtype=float32)

需要考虑的几点，

• pool_size 应该等于循环层的步长/时间步长。
• 输出的形状为(batch_size, downsampled_steps, features)，其中包含一个额外的downsampled_steps 维度。如果您将 pool_size 设置为等于循环层中的时间步长，则该值始终为 1。