def update_lstm(input, hiddens, cells):
change = np.tanh(concat_and_multiply(params['change'], input, hiddens))
forget = sigmoid(concat_and_multiply(params['forget'], input, hiddens))
ingate = sigmoid(concat_and_multiply(params['ingate'], input, hiddens))
outgate = sigmoid(concat_and_multiply(params['outgate'], input, hiddens))
cells = cells * forget + ingate * change
hiddens = outgate * np.tanh(cells)
return hiddens, cells
def update_lstm(input, hiddens, cells):
change = np.tanh(concat_and_multiply(params['change'], input, hiddens))
forget = sigmoid(concat_and_multiply(params['forget'], input, hiddens))
ingate = sigmoid(concat_and_multiply(params['ingate'], input, hiddens))
outgate = sigmoid(concat_and_multiply(params['outgate'], input, hiddens))
cells = cells * forget + ingate * change
hiddens = outgate * np.tanh(cells)
return hiddens, cells
def update_gru(input, hiddens):
update = sigmoid(
concat_and_multiply(params['transion']['update'], input, hiddens))
reset = sigmoid(
concat_and_multiply(params['transion']['reset'], input, hiddens))
hiddens = (1 - update) * hiddens + update * sigmoid(
concat_and_multiply(params['transion']['hiddenOut'], input,
hiddens * reset))
return hiddens
def hiddens_to_output_probs(hiddens):
output = concat_and_multiply(params['predict'], hiddens)
# Normalize log-probs.
return output - logsumexp(output, axis=1, keepdims=True)
def update_rnn(input, hiddens):
return np.tanh(
concat_and_multiply(params['rnn']['change'], input, hiddens))
def hiddens_to_output_probs(hiddens):
output = concat_and_multiply(params['predict'], hiddens)
noise = npr.gumbel(loc=0.0, scale=1.0, size=output.shape)
output = TEMP * (output + noise)
return output - logsumexp(output, axis=1,
keepdims=True) # Normalize log-probs.
def hiddens_to_output_probs(hiddens):
output = concat_and_multiply(params['predict'], hiddens)
return 1.0 / (1.0 + np.exp(-output))
def hiddens_to_output_probs(hiddens):
output = concat_and_multiply(params['predict'], hiddens)
return output - logsumexp(output, axis=1, keepdims=True) # Normalize log-probs.