使用paddle2.0实现LSTM

多层LSTM连接示意图：

单层LSTM的动态实现：

输入门、遗忘门、输出门：

import paddle# 输入数据向量维数6, 隐藏层维度8, 2个LSTM层串联(如果是1，可以省略，默认为1)
rnn = paddle.nn.LSTM(input_size=6, hidden_size=8, num_layers=2)# batch_size=4, 序列长度seq_len=3, 输入数据向量维数=6
x = paddle.randn((4, 3, 6))
print('x',x)# 初始化的隐藏层（prev_h）和记忆单元（prev_c）,通常它们的维度是一样的
# prev_h = paddle.randn((2, 4, 8)) # 2个LSTM层，batch_size=4,隐藏元维度8
# prev_c = paddle.randn((2, 4, 8)) # 2个LSTM层，batch_size=4,隐藏元维度8
# 输入数据向量维数、隐藏层维度、网路层数、序列长度确定后，prev_h和prev_c的维度已经固定，所以以上两行不需要亦可# y, (h, c) = rnn(x, (prev_h, prev_c))
y, (h, c) = rnn(x) 
print(rnn)# y/h/c表示的都是最后的输出，但输出不同# y是最后一层lstm的每个词向量对应隐藏层的输出，h,c是所有层最后一个隐藏层和记忆元的输出print('y',y)
print(y.shape)  # [4, 3, 8] # 这里有2层lstm，其与层数无关，只与序列长度相关（y是其实是句子的向量化表示）print('h',h)
print(h.shape)  # [2, 4, 8]
# 注意最后的大结论：对于batch中的每一个句子，最后一层lstm网络的h向量结果，和y(也就是output)最后一个时间节点（或单词）的输出，是一样的。
print('c',c)
print(c.shape)  # [2, 4, 8]paddle.summary(rnn,(4, 3, 6))
# 该网络中，共有1088个参数，计算过程如下：
# 2*4（14*8+3*8)=1088，其中：#  2表示：2层网络
#  4表示：batch_size为4，也就是有一次计算4个句子#  8表示：h的维度
# 14表示：8+6，h的维度（8）和单词维度(6)之和
#  3表示：单词（或者时间t）的个数
# 14*8表示：参数w 
(6)之和
#  3表示：单词（或者时间t）的个数
# 14*8表示：参数w 
#  3*8表示：4个单词

x Tensor(shape=[4, 3, 6], dtype=float32, place=CPUPlace, stop_gradient=True,[[[-0.39994577, -0.16884394, -1.09118211,  0.80282438,  0.77592939,  1.98887908],[-1.37776613, -0.50116003, -1.59069943, -0.92857558, -0.32358357,  0.40606323],[ 0.51607883, -0.90631920, -2.49934697,  0.52673978,  0.35466543, -1.23233390]],[[-0.35236493, -0.80131197,  0.41734406,  0.70300984,  2.56589055, -0.57770211],[ 0.52168477, -1.14594662,  0.89456636,  0.60237592, -0.64571965, -0.92338151],[-0.89197439,  2.07416654,  0.09576637,  0.51839197, -0.35192293, -0.34265402]],[[-1.20565021, -0.13362758, -1.21440089,  0.36546969,  0.29902026,  1.14574087],[ 0.16499370, -1.05054629,  0.60000432, -0.73830217,  1.45883906,  0.80631649],[ 0.90025264, -0.88333410,  1.86590922,  0.94177932,  0.25186336,  0.73045748]],[[ 0.59448457,  0.20400517,  0.10334436,  0.64043117,  1.15682471,  0.09200878],[-0.67695278,  0.52044302, -0.78882062,  1.27999902,  0.06613204,  1.43959534],[-0.00347163,  1.17208421,  0.45402086, -0.14898437, -0.22504820,  0.67211795]]])
LSTM(6, 8, num_layers=2(0): RNN((cell): LSTMCell(6, 8))(1): RNN((cell): LSTMCell(8, 8))
)
y Tensor(shape=[4, 3, 8], dtype=float32, place=CPUPlace, stop_gradient=False,[[[ 0.12606533,  0.03723524, -0.01543412, -0.05093168, -0.00058493,  0.02041843, -0.02015576, -0.03794828],[ 0.19590674,  0.09097288, -0.01683361, -0.09791360,  0.02531381,  0.04274859, -0.04400444, -0.05620224],[ 0.22171436,  0.14395498, -0.01471146, -0.12150520,  0.06584308,  0.05339855, -0.07185388, -0.05342202]],[[ 0.09426511,  0.04475664, -0.01056199, -0.07163173,  0.00677959,  0.02386722, -0.01371512, -0.04715024],[ 0.14545080,  0.07383338, -0.00895626, -0.13342401,  0.01213170,  0.04805123, -0.00590389, -0.06722270],[ 0.18524487,  0.11151871,  0.00289736, -0.19252071,  0.02871165,  0.06043112, -0.01222652, -0.06402460]],[[ 0.12229927,  0.05400331, -0.00953178, -0.06553464,  0.01734342,  0.02204929, -0.02571913, -0.04117016],[ 0.16857712,  0.07484694, -0.02250303, -0.09614565, -0.00239868,  0.04445631, -0.03118773, -0.05886544],[ 0.20006333,  0.06436764, -0.02840774, -0.11968626, -0.03018227,  0.06358111, -0.01926485, -0.06980018]],[[ 0.10642181,  0.05075689, -0.01515573, -0.05271235,  0.01304606,  0.02216698, -0.02666905, -0.04353783],[ 0.18087025,  0.06951316, -0.01432738, -0.10333572, -0.00218153,  0.04213399, -0.02817569, -0.05427728],[ 0.21686205,  0.08704377, -0.01545346, -0.12414743, -0.00640737,  0.05959129, -0.03789495, -0.05795782]]])
[4, 3, 8]
h Tensor(shape=[2, 4, 8], dtype=float32, place=CPUPlace, stop_gradient=False,[[[ 0.17736050, -0.08764581,  0.05920768,  0.03660521,  0.17524016, -0.19638550,  0.34747770,  0.21448627],[ 0.06513791, -0.11409879, -0.21574451, -0.14325079,  0.02657837, -0.27729952, -0.19942993, -0.03731265],[ 0.04778732,  0.21589978, -0.11827739, -0.12047609,  0.27885476,  0.02591145, -0.15165608, -0.39237073],[ 0.01800619,  0.05139154, -0.17368346,  0.08708881,  0.35007477, -0.10484657, -0.12915266, -0.13835609]],[[ 0.22171436,  0.14395498, -0.01471146, -0.12150520,  0.06584308,  0.05339855, -0.07185388, -0.05342202],[ 0.18524487,  0.11151871,  0.00289736, -0.19252071,  0.02871165,  0.06043112, -0.01222652, -0.06402460],[ 0.20006333,  0.06436764, -0.02840774, -0.11968626, -0.03018227,  0.06358111, -0.01926485, -0.06980018],[ 0.21686205,  0.08704377, -0.01545346, -0.12414743, -0.00640737,  0.05959129, -0.03789495, -0.05795782]]])
[2, 4, 8]
c Tensor(shape=[2, 4, 8], dtype=float32, place=CPUPlace, stop_gradient=False,[[[ 0.37910637, -0.42640921,  0.12153221,  0.06514800,  0.49467742, -0.33455023,  0.49630848,  0.28509313],[ 0.10826016, -0.28854835, -0.58194470, -0.24991161,  0.03781226, -0.45884287, -0.37416759, -0.11784617],[ 0.16163930,  0.67656636, -0.33257541, -0.33084434,  0.36936939,  0.13234609, -0.24183141, -0.92222750],[ 0.04265260,  0.11543229, -0.35903591,  0.16842273,  0.48566097, -0.27855623, -0.24916294, -0.36821401]],[[ 0.46252376,  0.27226013, -0.02847151, -0.19650725,  0.13580972,  0.11520113, -0.17235029, -0.10225918],[ 0.38834193,  0.20850745,  0.00485603, -0.34224576,  0.06490447,  0.13443303, -0.02630240, -0.11699687],[ 0.44824469,  0.13285753, -0.05207641, -0.23715332, -0.06911413,  0.14895318, -0.04506143, -0.13610318],[ 0.47666115,  0.16542867, -0.02775677, -0.23145792, -0.01375821,  0.13625184, -0.08468147, -0.10931204]]])
[2, 4, 8]
------------------------------------------------------------------------------------------Layer (type)       Input Shape                 Output Shape                 Param #    
==========================================================================================LSTM-2          [[4, 3, 6]]      [[4, 3, 8], [[2, 4, 8], [2, 4, 8]]]      1,088     
==========================================================================================
Total params: 1,088
Trainable params: 1,088
Non-trainable params: 0
------------------------------------------------------------------------------------------
Input size (MB): 0.00
Forward/backward pass size (MB): 0.00
Params size (MB): 0.00
Estimated Total Size (MB): 0.01
------------------------------------------------------------------------------------------{'total_params': 1088, 'trainable_params': 1088}