def output_hidden_state(C, o):
'''
Output the hidden state.
Input:
C: the current cell state of the LSTM, a float torch Tensor of shape n by h. Here h is the number of cell units.
o: the output gate values of the batch of training instances, a float matrix of shape n by h. Here h is the number of hidden units.
Output:
H: the output hidden state values of the batch of training instances, a float matrix of shape n by h. Here h is the number of hidden units.
'''
#########################################
## INSERT YOUR CODE HERE
H = o * tanh(C)
#########################################
return H
def output_hidden_state(C, o):
'''
Outputing the hidden state.
Input:
C: the current cell state of the LSTM, a float torch Tensor of shape n by h. Here h is the number of cell units.
o: the output gate values of the batch of training instances, a float matrix of shape n by h. Here h is the number of hidden units.
Output:
H: the output hidden state values of the batch of training instances, a float matrix of shape n by h. Here h is the number of hidden units.
'''
c1 = tanh(C)
H = th.mul(c1,o)
return H
def gates(self, x, H):
'''
Given a batch of training instances (with one time step), computing the gating functions: forget gate f, input gate i, output gate o; and candidate cells state.
Note: W_f is matrix consisting weights for both inputs and hidden states. Here we assume the first p rows of W_f coresponds to input weights, last h rows of W_f coresponds to the weights of hidden states.
Input:
x: a batch of training instance, a float torch Tensor of shape n by p. Here n is the batch size. p is the number of features.
H: the hidden state of the LSTM, a float torch Tensor of shape n by h. Here h is the number of hidden units.
Output:
f: the forget gate values of the batch of training instances, a float matrix of shape n by h. Here h is the number of hidden units.
i: the input gate values of the batch of training instances, a float matrix of shape n by h. Here h is the number of hidden units.
o: the output gate values of the batch of training instances, a float matrix of shape n by h. Here h is the number of hidden units.
C_c: the candidate cell state values of the batch of training instances, a float matrix of shape n by h. Here h is the number of hidden units.
Hint: you could solve this problem using 4-5 lines of code.
'''
f = th.sigmoid(th.mm((th.cat([x,H],1)),self.W_f)+self.b_f)
i = th.sigmoid(th.mm(th.cat([x,H],1),self.W_i)+self.b_i)
o = th.sigmoid(th.mm(th.cat([x,H],1),self.W_o)+self.b_o)
C_c = tanh(th.mm(th.cat([x,H],1),self.W_c)+self.b_c)
return f, i, o, C_c
def gates(self, x, H):
'''
Given a batch of training instances (with one time step), compute the gating functions: forget gate f, input gate i, output gate o; and candidate cells state.
Note: W_f is matrix consisting weights for both inputs and hidden states. Here we assume the first p rows of W_f coresponds to input weights, last h rows of W_f coresponds to the weights of hidden states.
Input:
x: a batch of training instance, a float torch Tensor of shape n by p. Here n is the batch size. p is the number of features.
H: the hidden state of the LSTM, a float torch Tensor of shape n by h. Here h is the number of hidden units.
Output:
f: the forget gate values of the batch of training instances, a float matrix of shape n by h. Here h is the number of hidden units.
i: the input gate values of the batch of training instances, a float matrix of shape n by h. Here h is the number of hidden units.
o: the output gate values of the batch of training instances, a float matrix of shape n by h. Here h is the number of hidden units.
C_c: the candidate cell state values of the batch of training instances, a float matrix of shape n by h. Here h is the number of hidden units.
Hint: you could solve this problem using 4-5 lines of code.
'''
#########################################
## INSERT YOUR CODE HERE
n, p = x.size()
f = th.cat((x, H), dim=1)
f = th.mm(f, self.W_f)
f = f + self.b_f.expand(n, self.h)
f = th.sigmoid(f)
i = th.cat((x, H), dim=1)
i = th.mm(i, self.W_i)
i = i + self.b_i.expand(n, self.h)
i = th.sigmoid(i)
o = th.cat((x, H), dim=1)
o = th.mm(o, self.W_o)
o = o + self.b_o.expand(n, self.h)
o = th.sigmoid(o)
C_c = th.cat((x, H), dim=1)
C_c = th.mm(C_c, self.W_c)
C_c = C_c + self.b_c.expand(n, self.h)
C_c = tanh(C_c)
#########################################
return f, i, o, C_c