def fit(words, tags, labels, model, builders):
"""
compute joint error of the
:param words: list of indices
:param tags: list of indices
:param labels: index
:param model: current model to access parameters
:param builders: builder to create state combinations
:return: joint error
"""
# retrieve model parameters
if MLP:
H = pycnn.parameter(pH)
O = pycnn.parameter(pO)
else:
O = pycnn.parameter(pO)
errs = []
for (forward_state, backward_state), tag in zip(build_tagging_graph(words, model, builders), tags):
f_b = pycnn.concatenate([forward_state, backward_state])
if MLP:
# TODO: add bias terms
r_t = O * (pycnn.tanh(H * f_b))
else:
r_t = O * f_b
err = pycnn.pickneglogsoftmax(r_t, tag)
errs.append(err)
return pycnn.esum(errs)
def predict(sent, model, builders):
"""
predict tags and demographic labels
:param sent:
:param model:
:param builders:
:return: tag and label predictions
"""
if MLP:
H = pycnn.parameter(pH)
O = pycnn.parameter(pO)
else:
O = pycnn.parameter(pO)
tags = []
for forward_state, backward_state in build_tagging_graph(words, model, builders):
if MLP:
r_t = O * (pycnn.tanh(H * pycnn.concatenate([forward_state, backward_state])))
else:
r_t = O * pycnn.concatenate([forward_state, backward_state])
out = pycnn.softmax(r_t)
chosen = np.argmax(out.npvalue())
tags.append(vocab_tags.i2w[chosen])
return tags
def add_input(self, input_vec):
"""
Note that this function updates the existing State object!
"""
x = pycnn.concatenate([input_vec, self.h])
i = pycnn.logistic(self.W_i * x + self.b_i)
f = pycnn.logistic(self.W_f * x + self.b_f)
g = pycnn.tanh(self.W_c * x + self.b_c)
o = pycnn.logistic(self.W_o * x + self.b_o)
c = pycnn.cwise_multiply(f, self.c) + pycnn.cwise_multiply(i, g)
h = pycnn.cwise_multiply(o, pycnn.tanh(c))
self.c = c
self.h = h
self.outputs.append(h)
return self
def attend(model, input_vectors, state):
w1 = pc.parameter(model['attention_w1'])
w2 = pc.parameter(model['attention_w2'])
v = pc.parameter(model['attention_v'])
attention_weights = []
w2dt = w2*pc.concatenate(list(state.s()))
for input_vector in input_vectors:
attention_weight = v*pc.tanh(w1*input_vector + w2dt)
attention_weights.append(attention_weight)
attention_weights = pc.softmax(pc.concatenate(attention_weights))
output_vectors = pc.esum([vector*attention_weight for vector, attention_weight in zip(input_vectors, attention_weights)])
return output_vectors
def attend(model, input_vectors, state):
w1 = pc.parameter(model['attention_w1'])
w2 = pc.parameter(model['attention_w2'])
v = pc.parameter(model['attention_v'])
attention_weights = []
w2dt = w2*pc.concatenate(list(state.s()))
for input_vector in input_vectors:
attention_weight = v*pc.tanh(w1*input_vector + w2dt)
attention_weights.append(attention_weight)
attention_weights = pc.softmax(pc.concatenate(attention_weights))
output_vectors = pc.esum([vector*attention_weight for vector, attention_weight in zip(input_vectors, attention_weights)])
return output_vectors
def predict(word_indices, model, builder, target):
"""
predict demographic label
:param word_indices:
:param model:
:param builder:
:return: tag and label predictions
"""
forward_states = build_tagging_graph(word_indices, model, builder)
final_state = forward_states[-1]
H = pycnn.parameter(pH)
bias_H = pycnn.parameter(biasH)
H2 = pycnn.parameter(pH2)
bias_H2 = pycnn.parameter(biasH2)
if target in ['age', 'both', 'joint']:
O = pycnn.parameter(pOutAge)
bias_O = pycnn.parameter(biasOutAge)
elif target == 'gender':
O = pycnn.parameter(pOutGender)
bias_O = pycnn.parameter(biasOutGender)
if target == 'both':
O2 = pycnn.parameter(pOutGender)
bias_O2 = pycnn.parameter(biasOutGender)
if target == 'both':
# hidden = bias_H2 + pycnn.tanh(H2 * (bias_H + pycnn.tanh(H * final_state)))
hidden = bias_H + pycnn.tanh(H * final_state)
r_age = bias_O + (O * hidden)
r_gender = bias_O2 + (O2 * hidden)
out_age = pycnn.softmax(r_age)
out_gender = pycnn.softmax(r_gender)
return [np.argmax(out_age.npvalue()), np.argmax(out_gender.npvalue())]
else:
# r_t = bias_O + (O * (bias_H2 + pycnn.tanh(H2 * (bias_H + pycnn.tanh(H * final_state)))))
r_t = bias_O + (O * (bias_H + (H * final_state)))
out = pycnn.softmax(r_t)
chosen = np.argmax(out.npvalue())
return chosen
def __init__(self, lstm):
self.lstm = lstm
self.outputs = []
self.c = pycnn.parameter(self.lstm.c0)
self.h = pycnn.tanh(self.c)
self.W_i = pycnn.parameter(self.lstm.W_i)
self.b_i = pycnn.parameter(self.lstm.b_i)
self.W_f = pycnn.parameter(self.lstm.W_f)
self.b_f = pycnn.parameter(self.lstm.b_f)
self.W_c = pycnn.parameter(self.lstm.W_c)
self.b_c = pycnn.parameter(self.lstm.b_c)
self.W_o = pycnn.parameter(self.lstm.W_o)
self.b_o = pycnn.parameter(self.lstm.b_o)
