Issue about forget_bias - tensorflow

Question

I'm going to write a BasicLSTMCell with C++,and I need check it work well.
I used tf.nn.rnn_cell.BasicLSTMCell to implement a LSTM with 4 cells,and I set
forget_bias to 1.Then I Check the LSTM'bias using this code:
//////////////////////////////////////////////////////////////
with tf.variable_scope("LSTM"):
Cell=tf.nn.rnn_cell.BasicLSTMCell(4,forget_bias=1,state_is_tuple=True)
Sessin=tf.Session()
state=Cell.zero_state(1,dtype=tf.float32)
with tf.variable_scope("Ut_def"):
out,D=tf.nn.dynamic_rnn(
cell=Cell,inputs=Feed,
initial_state=state,
time_major=False)
Sessin.run(tf.global_variables_initializer())
#Saver.save(Sessin,"./123/Var",global_step=1)
out,D=Sessin.run([out,D],feed_dict={Feed:np.arange(8).reshape(1,2,4)})
tf.train.Saver().save(Sessin,"./123/Var",global_step=1)
trainable_vars_dict = {}
for key in tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES):
trainable_vars_dict[key.name] = Sessin.run(key)
# Checking the names of the keys
print(key.name)
lstm_weight_vals = trainable_vars_dict["Ut_def/RNN/BasicLSTMCell/Linear/Matrix:0"]
B=trainable_vars_dict["Ut_def/RNN/BasicLSTMCell/Linear/Bias:0"]
print(B)
/////////////////////////////////////////////////////////////
But I found these bias were all zeros whatever I changed the forget_bias.
Someone know that what it is going on?
For figuring out how lstm work,I just using wights and bias extracting from tensorflow for getting same result. Absolutely they are not equivalent.
w_i, w_C, w_f, w_o = np.split(lstm_weight_vals, 4, axis=1)
w_xi = w_i[:4, :]
w_hi = w_i[4:, :]
w_xC = w_C[:4, :]
w_hC = w_C[4:, :]
w_xf = w_f[:4, :]
w_hf = w_f[4:, :]
w_xo = w_o[4:, :]
w_ho = w_o[4:, :]
Input=tf.range(4,dtype=tf.float32)
Input=tf.reshape(Input,shape=[1,4])
i=tf.sigmoid(tf.matmul(tf.zeros(shape=[1,4]),w_xi)+tf.matmul(Input,w_hi))
o=tf.sigmoid(tf.matmul(tf.zeros(shape=[1,4]),w_xo)+tf.matmul(Input,w_ho))
g=tf.tanh(tf.matmul(tf.zeros(shape=[1,4]),w_xC)+tf.matmul(Input,w_hC))
f=tf.sigmoid(tf.matmul(tf.zeros(shape=[1,4]),w_xf)+tf.matmul(Input,w_hf))
Cstate=tf.zeros(shape=[1,4])*f+i*g
Hstate=tf.tanh(Cstate)*o
Input=Input+4
i=tf.sigmoid(tf.matmul(Cstate,w_xi)+tf.matmul(Input,w_hi))
o=tf.sigmoid(tf.matmul(Cstate,w_xo)+tf.matmul(Input,w_ho))
g=tf.tanh(tf.matmul(Cstate,w_xC)+tf.matmul(Input,w_hC))
f=tf.sigmoid(tf.matmul(Cstate,w_xf)+tf.matmul(Input,w_hf))
Cstate=Cstate*f+i*g
Hstate=tf.tanh(Cstate)*o

Answer 1

I has found wrong code. Code should like this:
i=tf.sigmoid(tf.matmul(Hstate,w_xi)+tf.matmul(Input,w_hi))
o=tf.sigmoid(tf.matmul(Hstate,w_xo)+tf.matmul(Input,w_ho))
g=tf.tanh(tf.matmul(Hstate,w_xC)+tf.matmul(Input,w_hC))
f=tf.sigmoid(tf.matmul(Hstate,w_xf)+tf.matmul(Input,w_hf)+1)
It's Hstate instead of Csatae.