序列表示方法

具有时间相关性



[b个波形, 采样100个seq时间点, 对应每个时间点的feature]

最简单的one_hot encoding



意思相近的单词在向量空间上相关性距离越小 表示方式越好

Embedding Layer

In [2]: from tensorflow.keras import layers
In [3]: import tensorflow as tf

In [6]: x = tf.range(5)

In [7]: x = tf.random.shuffle(x)

In [8]: net = layers.Embedding(10, 4)  # 网络最多10个单词,每个单词用一个长度为4的vector表示

In [9]: net(x)  # x为5个单词,则对应[5, 4]
Out[9]: 
<tf.Tensor: shape=(5, 4), dtype=float32, numpy=
array([[-0.04134145,  0.02167301, -0.00487036, -0.03129961],
       [-0.00446678,  0.0411575 , -0.0083902 , -0.04393339],
       [-0.04771481,  0.03433364,  0.03229776,  0.01344705],
       [ 0.03182409, -0.04264296, -0.02120506,  0.00117933],
       [-0.00085657, -0.04112577,  0.04295324, -0.02037839]],
      dtype=float32)>

In [10]: net.trainable
Out[10]: True

In [11]: net.trainable_variables   # trainable
Out[11]: 
[<tf.Variable 'embedding/embeddings:0' shape=(10, 4) dtype=float32, numpy=
 array([[-0.00446678,  0.0411575 , -0.0083902 , -0.04393339],
        [-0.00085657, -0.04112577,  0.04295324, -0.02037839],
        [-0.04134145,  0.02167301, -0.00487036, -0.03129961],
        [ 0.03182409, -0.04264296, -0.02120506,  0.00117933],
        [-0.04771481,  0.03433364,  0.03229776,  0.01344705],
        [ 0.00809318,  0.00439113, -0.02471092,  0.01798195],
        [ 0.02877717,  0.02287935, -0.04870056, -0.01806825],
        [-0.00918406,  0.03466691,  0.04272464, -0.03204942],
        [ 0.03497279,  0.03022062, -0.00239567,  0.00312829],
        [-0.04340002, -0.04850687, -0.02064794, -0.03603333]],
       dtype=float32)>]

最后

以上就是哭泣奇迹最近收集整理的关于序列表示方法的全部内容，更多相关序列表示方法内容请搜索靠谱客的其他文章。