dummy

RAWCODE

1 W,U合并

在进入代码环节之前, 先介绍一下代码中出现的W,U参数的合并, 如下图

              |<---H---->|<---------K---------->|
           ---+---------------------------------+
           ^  |          |                      |
           |  |          |                      |
           H  |    W     |          U           |
           |  |          |                      |
           v  |          |                      |
           ---+---------------------------------+
                 Wh_{t-1}          Ux_t

前: \(s_t = \phi(Wh_{t-1} + Ux_t + b)\)

后: $s_t = (W' <h_{t-1}, x_t> + b) = (W' z_t + b) $

从空间维度分析:

\(W \in \mathbf R^{H \times H}\), \(U \in \mathbf R^{H \times K}\), \(W' \in \mathbf R^{H \times (H+K)}\)

\(h_{t-1} \in \mathbf R^{H \times 1}\), \(x_t \in \mathbf R^{K \times 1}\), \(z_t \in \mathbf R^{(H+K) \times 1}\)

2 Logits

在代码中出现, 把这个单词拆开log + it(它odds) + s(复数), 即对它们(odds)取log. odds是"几率, 胜率", 一个例子 搞定:

assume: \(\text{ if } P = \dfrac {1}{5}, 1-P = \dfrac{4}{5} \text { then } Odds(A) = log \dfrac{1/5}{4/5}\)

logits (未归一化的概率)一般作为softmax的输入参数(归一化).

3 Codes

百度云盘Datasets

4 References