I created my own network model and used data with dimension [batchez, 10, 8] to trained this model.
After that, I wanted to use tensor with dimension [1, 40, 8] to run net.predict(x), ('x' is the name of the tensor), but I got the error:
Input shape axis 1 must equal 10, got shape [1,40,8]
In my opinion, the axis 1 only affects the number of calls to the LSTMCell. Why should it be the same as 10? How can I deal with this problem?
At the same time, I also created a variable in the network to determine whether it is training, because I want to call LSTMCell only once to obtain the output result while it is not training. However, it seems that with the problem above, I can't achieve this goal.
So please help me.
Here is the code.
class lstm_rnn(keras.Model):
def __init__(self, units):
super(lstm_rnn, self).__init__()
self.state0 = [tf.zeros([batchsz, units]),tf.zeros([batchsz, units])]
self.state1 = [tf.zeros([batchsz, units]),tf.zeros([batchsz, units])]
self.lstm_cell0 = layers.LSTMCell(units, dropout = 0.5)
self.lstm_cell1 = layers.LSTMCell(units, dropout = 0.5)
def call(self, inputs):
x = inputs
real_out = 0
axis_size = x.shape[1]
is_training = True
if(axis_size == 2):
is_training = False
print(inputs.shape)
if(is_training):
state0 = self.state0
state1 = self.state1
step_cnt = 0
for word in tf.unstack(x, axis = 1):
out0, state0 = self.lstm_cell0(word, state0)
out1, state1 = self.lstm_cell1(out0, state1)
if(step_cnt == 0):
real_out = tf.reshape(out1, shape = (20, 1 ,6))
else:
real_out = tf.concat([real_out, tf.reshape(out1, shape = (20, 1, 6))], axis = 1)
step_cnt = step_cnt + 1
else:
state0 = [inputs[0], inputs[0]]
state1 = [inputs[0], inputs[0]]
info = inputs[1]
out0, state0 = self.lstm_cell0(info, state0)
out1, state1 = self.lstm_cell1(out0, state1)
real_out = out1
return real_out
