1.  什么是LSTM

长短期记忆网络 LSTM（long short-term memory）是 RNN 的一种变体，其核心概念在于细胞状态以及“门”结构。细胞状态相当于信息传输的路径，让信息能在序列连中传递下去。你可以将其看作网络的“记忆”。理论上讲，细胞状态能够将序列处理过程中的相关信息一直传递下去。因此，即使是较早时间步长的信息也能携带到较后时间步长的细胞中来，这克服了短时记忆的影响。信息的添加和移除我们通过“门”结构来实现，“门”结构在训练过程中会去学习该保存或遗忘哪些信息。
 

 

2. 实验代码

2.1. 搭建一个只有一层RNN和Dense网络的模型。

2.2. 验证LSTM里的逻辑

 假设我的输入数据是x = [1,0], 

kernel = [[[2, 1, 1, 0, 0, 0, 0, 1, 1, 0, 1, 0],

              [1, 1, 0, 1, 1, 0, 0, 1, 1 ,0, 0, 0],]]

recurrent_kernel = [[1, 0, 0, 1, 2,1,0,1,2,0,1,0],

                              [1, 1, 0, 0, 2,1,0,1,2,2,0,0],

                              [1, 0, 1, 2, 0,1,0,1,1,0,1,0]]

biase = [3, 1, 0, 1, 1,0,0,1,0,2,0.0,0]

通过下面手算，h的结果是[0, 4,1], c 的结果是[0,4,1].  注意无激活函数。

代码验证上面的结果

def change_weight():
    # Create a simple Dense layer
    lstm_layer = LSTM(units=3, input_shape=(3, 2), activation=None, recurrent_activation=None, return_sequences=True,
                      return_state= True)

    # Simulate input data (batch size of 1 for demonstration)
    input_data = np.array([
                [[1.0, 2], [2, 3], [3, 4]],
                [[5, 6], [6, 7], [7, 8]],
                [[9, 10], [10, 11], [11, 12]]
        ])

    # Pass the input data through the layer to initialize the weights and biases
    lstm_layer(input_data)

    kernel, recurrent_kernel, biases = lstm_layer.get_weights()

    # Print the initial weights and biases
    print("recurrent_kernel:", recurrent_kernel, recurrent_kernel.shape ) # (3,3)
    print('kernal:',kernel, kernel.shape) #(2,3)
    print('biase: ',biases , biases.shape) # (3)


    kernel = np.array([[2, 1, 1, 0, 0, 0, 0, 1, 1, 0, 1, 0],
                       [1, 1, 0, 1, 1, 0, 0, 1, 1 ,0, 0, 0],])

    recurrent_kernel = np.array([[1, 0, 0, 1, 2,1,0,1,2,0,1,0],
                                 [1, 1, 0, 0, 2,1,0,1,2,2,0,0],
                                 [1, 0, 1, 2, 0,1,0,1,1,0,1,0]])

    biases = np.array([3, 1, 0, 1, 1,0,0,1,0,2,0.0,0])

    lstm_layer.set_weights([kernel, recurrent_kernel, biases])
    print(lstm_layer.get_weights())

    # test_data = np.array([
    #     [[1.0, 3], [1, 1], [2, 3]]
    # ])

    test_data = np.array([
        [[1,0.0]]
    ])

    output, memory_state, carry_state  = lstm_layer(test_data)

    print(output)
    print(memory_state)
    print(carry_state)
if __name__ == '__main__':
    change_weight()

执行结果：

recurrent_kernel: [[-0.36744034 -0.11181469 -0.10642298  0.5450207  -0.30208975  0.5405432
   0.09643812 -0.14983998  0.1859854   0.2336958  -0.16187981  0.11621032]
 [ 0.07727922 -0.226477    0.1491096  -0.03933501  0.31236103 -0.12963092
   0.10522162 -0.4815724  -0.2093935   0.34740582 -0.60979587 -0.15877807]
 [ 0.15371156  0.01244636 -0.09840634 -0.32093546  0.06523462  0.18934932
   0.38859126 -0.3261706  -0.05138849  0.42713478  0.49390993  0.37013963]] (3, 12)
kernal: [[-0.47606698 -0.43589187 -0.5371355  -0.07337284  0.30526626 -0.18241835
  -0.03675252  0.2873094   0.33218485  0.24838251  0.17765659  0.4312396 ]
 [ 0.4007727   0.41280174  0.40750778 -0.6245315   0.6382301   0.42889225
   0.11961156 -0.6021105  -0.43556038  0.39798307  0.6390712   0.16719025]] (2, 12)
biase:  [0. 0. 0. 1. 1. 1. 0. 0. 0. 0. 0. 0.] (12,)
[array([[2., 1., 1., 0., 0., 0., 0., 1., 1., 0., 1., 0.],
       [1., 1., 0., 1., 1., 0., 0., 1., 1., 0., 0., 0.]], dtype=float32), array([[1., 0., 0., 1., 2., 1., 0., 1., 2., 0., 1., 0.],
       [1., 1., 0., 0., 2., 1., 0., 1., 2., 2., 0., 0.],
       [1., 0., 1., 2., 0., 1., 0., 1., 1., 0., 1., 0.]], dtype=float32), array([3., 1., 0., 1., 1., 0., 0., 1., 0., 2., 0., 0.], dtype=float32)]
tf.Tensor([[[0. 4. 0.]]], shape=(1, 1, 3), dtype=float32)
tf.Tensor([[0. 4. 0.]], shape=(1, 3), dtype=float32)
tf.Tensor([[0. 4. 1.]], shape=(1, 3), dtype=float32)

可以看出h=[0,4,0], c=[0,4,1]