Keras之LSTM源码阅读笔记。
Recurrent是LSTM的父类(实际是通过SimpleRNN间接继承),定义所有RNNs的统一接口。
implementation: one of {0, 1, or 2}.
If set to 0, the RNN will use an implementation that uses fewer, larger matrix products, thus running faster on CPU but consuming more memory. If set to 1, the RNN will use more matrix products, but smaller ones, thus running slower (may actually be faster on GPU) while consuming less memory. If set to 2 (LSTM/GRU only), the RNN will combine the input gate, the forget gate and the output gate into a single matrix, enabling more time-efficient parallelization on the GPU. Note: RNN dropout must be shared for all gates, resulting in a slightly reduced regularization.
博主主要使用GPU加速且不在意内存的代码,所以通常设置implementation=2,源码阅读也主要集中在implementation=2的部分。
weights: list of Numpy arrays to set as initial weights.
The list should have 3 elements, of shapes: [(input_dim, output_dim), (output_dim, output_dim), (output_dim,)]`.
Activation function to use for the recurrent step.
注意: 默认值是’hard_sigmoid’,而原论文中用的’sigmoid’。
1 . 初始化
self.kernel = self.add_weight(shape=(self.input_dim, self.units * 4), name='kernel', initializer=self.kernel_initializer, regularizer=self.kernel_regularizer, constraint=self.kernel_constraint)
2 . 分块意义
self.kernel_i = self.kernel[:, :self.units] self.kernel_f = self.kernel[:, self.units: self.units * 2] self.kernel_c = self.kernel[:, self.units * 2: self.units * 3] self.kernel_o = self.kernel[:, self.units * 3:]
3 . kernel是用于和输入x做乘法的矩阵
1 . 初始化:
self.recurrent_kernel = self.add_weight( shape=(self.units, self.units * 4), name='recurrent_kernel', initializer=self.recurrent_initializer, regularizer=self.recurrent_regularizer, constraint=self.recurrent_constraint)
2 . 分块意义
self.recurrent_kernel_i = self.recurrent_kernel[:, :self.units] self.recurrent_kernel_f = self.recurrent_kernel[:, self.units: self.units * 2] self.recurrent_kernel_c = self.recurrent_kernel[:, self.units * 2: self.units * 3] self.recurrent_kernel_o = self.recurrent_kernel[:, self.units * 3:]
3 . recurrent_kernel是用于和前一时刻隐层输出h做乘法的矩阵
迭代部分代码如下:
if self.implementation == 2: z = K.dot(inputs * dp_mask[0], self.kernel) z += K.dot(h_tm1 * rec_dp_mask[0], self.recurrent_kernel) if self.use_bias: z = K.bias_add(z, self.bias) z0 = z[:, :self.units] z1 = z[:, self.units: 2 * self.units] z2 = z[:, 2 * self.units: 3 * self.units] z3 = z[:, 3 * self.units:] i = self.recurrent_activation(z0) f = self.recurrent_activation(z1) c = f * c_tm1 + i * self.activation(z2) o = self.recurrent_activation(z3) h = o * self.activation(c)
可见activation 作用于i,f,o的生成,recurrent_activation作用于g的生成以及在c的输出部分做微调。如果要模拟原论文的话,应该设置activation = tanh, recurrent_activation = sigmoid。