我建模了一个双层LSTM Keras模型，然后通过输入相同的权重和输入，将第一个LSTM层的输出与LSTM层的简单python实现进行比较 . 批次的第一个序列的结果相似但不相同，并且从第二个序列得到的结果偏差太大 . 下面是我的keras模型：
为了比较Keras模型，我首先创建了一个中间层，其中中间层输出第一层的结果，第一个序列的结果为 print(intermediate_output[0,0]) ，同一批次的第二个序列的 print(intermediate_output[0][1]) ，然后是最后一个序列的 print(intermediate_output[0][127]) .
inputs = Input(shape=(128,9))
f1=LSTM((n_hidden),return_sequences=True,name='lstm1')(inputs)
f2=LSTM((n_hidden), return_sequences=False,name='lstm2')(f1)
fc=Dense(6,activation='softmax',kernel_regularizer=regularizers.l2(lambda_loss_amount),name='fc')(f2)
model2 = Model(inputs=inputs, outputs=fc)
layer_name = 'lstm1'
intermediate_layer_model = Model(inputs=model2.input,
outputs=model2.get_layer(layer_name).output)
intermediate_output = intermediate_layer_model.predict(X_single_sequence[0,:,:])
print(intermediate_output[0,0]) # takes input[0][9]
print(intermediate_output[0][1]) # takes input[1][9] and hidden layer output of intermediate_output[0,0]
print(intermediate_output[0][127])
重新实现相同模型的第一层：我定义了 LSTMlayer 函数，它执行相同的计算....之后 weightLSTM 加载保存的权重和 x_t 相同的输入序列，稍后 h_t 包含下一个序列的输出 . intermediate_out 是与LSTM层相对应的函数 .
def sigmoid(x):
return(1.0/(1.0+np.exp(-x)))
def LSTMlayer(warr,uarr, barr,x_t,h_tm1,c_tm1):
'''
c_tm1 = np.array([0,0]).reshape(1,2)
h_tm1 = np.array([0,0]).reshape(1,2)
x_t = np.array([1]).reshape(1,1)
warr.shape = (nfeature,hunits*4)
uarr.shape = (hunits,hunits*4)
barr.shape = (hunits*4,)
'''
s_t = (x_t.dot(warr) + h_tm1.dot(uarr) + barr)
hunit = uarr.shape[0]
i = sigmoid(s_t[:,:hunit])
f = sigmoid(s_t[:,1*hunit:2*hunit])
_c = np.tanh(s_t[:,2*hunit:3*hunit])
o = sigmoid(s_t[:,3*hunit:])
c_t = i*_c + f*c_tm1
h_t = o*np.tanh(c_t)
return(h_t,c_t)
weightLSTM = model2.layers[1].get_weights()
warr,uarr, barr = weightLSTM
warr.shape,uarr.shape,barr.shape
def intermediate_out(n,warr,uarr,barr,X_test):
for i in range(0, n+1):
if i==0:
c_tm1 = np.array([0]*hunits, dtype=np.float32).reshape(1,32)
h_tm1 = np.array([0]*hunits, dtype=np.float32).reshape(1,32)
h_t,ct = LSTMlayer(warr,uarr, barr,X_test[0][0:1][0:9],h_tm1,c_tm1)
else:
h_t,ct = LSTMlayer(warr,uarr, barr,X_test[0][i:i+1][0:9],h_t,ct)
return h_t
# 1st sequence
ht0 = intermediate_out(0,warr,uarr,barr,X_test)
# 2nd sequence
ht1 = intermediate_out(1,warr,uarr,barr,X_test)
# 128th sequence
ht127 = intermediate_out(127,warr,uarr,barr,X_test)
来自 print(intermediate_output[0,0]) 的keras LSTM层的输出如下：
array([-0.05616369, -0.02299516, -0.00801201, 0.03872827, 0.07286803,
-0.0081161 , 0.05235862, -0.02240333, 0.0533984 , -0.08501752,
-0.04866522, 0.00254417, -0.05269946, 0.05809477, -0.08961852,
0.03975506, 0.00334282, -0.02813114, 0.01677909, -0.04411673,
-0.06751891, -0.02771493, -0.03293832, 0.04311397, -0.09430656,
-0.00269871, -0.07775293, -0.11201388, -0.08271968, -0.07464679,
-0.03533605, -0.0112953 ], dtype=float32)
我从 print(ht0) 执行的输出是：
array([[-0.05591469, -0.02280132, -0.00797964, 0.03681555, 0.06771626,
-0.00855897, 0.05160453, -0.02309707, 0.05746563, -0.08988875,
-0.05093143, 0.00264367, -0.05087904, 0.06033305, -0.0944235 ,
0.04066657, 0.00344291, -0.02881387, 0.01696692, -0.04101779,
-0.06718517, -0.02798996, -0.0346873 , 0.04402719, -0.10021093,
-0.00276826, -0.08390114, -0.1111543 , -0.08879325, -0.07953986,
-0.03261982, -0.01175724]], dtype=float32)
来自 print(intermediate_output[0][1]) 的输出：
array([-0.13193817, -0.03231169, -0.02096735, 0.07571879, 0.12657365,
0.00067896, 0.09008797, -0.05597101, 0.09581321, -0.1696091 ,
-0.08893952, -0.0352162 , -0.07936387, 0.11100324, -0.19354928,
0.09691346, -0.0057206 , -0.03619875, 0.05680932, -0.08598096,
-0.13047703, -0.06360915, -0.05707538, 0.09686109, -0.18573627,
0.00711019, -0.1934243 , -0.21811798, -0.15629804, -0.17204499,
-0.07108577, 0.01727455], dtype=float32)
print(ht1) ：
array([[-1.34333193e-01, -3.36792655e-02, -2.06091907e-02,
7.15097040e-02, 1.18231244e-01, 7.98894180e-05,
9.03479978e-02, -5.85013032e-02, 1.06357656e-01,
-1.82848617e-01, -9.50253978e-02, -3.67032290e-02,
-7.70251378e-02, 1.16113290e-01, -2.08772928e-01,
9.89214852e-02, -5.82863577e-03, -3.79538871e-02,
6.01535551e-02, -7.99121782e-02, -1.31876275e-01,
-6.66067824e-02, -6.15542643e-02, 9.91254672e-02,
-2.00229391e-01, 7.51443207e-03, -2.13641390e-01,
-2.18286291e-01, -1.70858681e-01, -1.88928470e-01,
-6.49823472e-02, 1.72227081e-02]], dtype=float32)
print(intermediate_output[0][127]) ：
array([-0.46212202, 0.280646 , 0.514289 , -0.21109435, 0.53513926,
0.20116206, 0.24579187, 0.10773794, -0.6350403 , -0.0052841 ,
-0.15971565, 0.00309152, 0.04909453, 0.29789132, 0.24909772,
0.12323025, 0.15282209, 0.34281147, -0.2948742 , 0.03674917,
-0.22213924, 0.17646286, -0.12948939, 0.06568322, 0.04172657,
-0.28638166, -0.29086435, -0.6872528 , -0.12620741, 0.63395363,
-0.37212485, -0.6649531 ], dtype=float32)
print(ht127) ：
array([[-0.47431907, 0.29702517, 0.5428258 , -0.21381126, 0.6053808 ,
0.22849198, 0.25656056, 0.10378123, -0.6960949 , -0.09966939,
-0.20533416, -0.01677105, 0.02512029, 0.37508538, 0.35703233,
0.14703275, 0.24901289, 0.35873395, -0.32249793, 0.04093777,
-0.20691746, 0.20096642, -0.11741923, 0.06169611, 0.01019177,
-0.33316574, -0.08499744, -0.6748463 , -0.06659956, 0.71961826,
-0.4071832 , -0.6804066 ]], dtype=float32)
来自（ print(intermediate_output[0,0]) ， print(h_t[0]) ）和（ print(intermediate_output[0][1]) ， print(h_t1) ）的输出类似......但 print(intermediate_output[0][127]) 和 print(h_t127) 的输出不相同，并且两种算法都在同一个gpu上实现...
我看到了keras文档，对我来说，似乎我没有做错任何事情......请对此发表评论并让我知道我还缺少什么？