def attend(self, input_mat, state, w1dt, input_len, batch_size):
global attention_w2
global attention_v
w2 = dy.parameter(attention_w2)
v = dy.parameter(attention_v)
w2dt = w2 * dy.concatenate(list(state.s()))
unnormalized = dy.transpose(v * dy.tanh(dy.colwise_add(w1dt, w2dt)))
unnormalized = dy.reshape(unnormalized, (input_len, ), batch_size)
att_weights = dy.softmax(unnormalized)
context = input_mat * att_weights
return context, att_weights
def attend_batch(input_mat, state, w1dt, batch_size, input_length):
#print "in attend batch"
global attention_w2
global attention_v
w2 = dy.parameter(attention_w2)
v = dy.parameter(attention_v)
#print "Calculating w2dt"
w2dt = w2*dy.concatenate(list(state.s()))
unnormalized = dy.transpose(v * dy.tanh(dy.colwise_add(w1dt, w2dt)))
attention_reshaped = dy.reshape(unnormalized, (input_length, ), batch_size)
att_weights = dy.softmax(attention_reshaped)
context = input_mat * att_weights
return context
def attend(input_mat, state, w1dt): global attention_w2 global attention_v w2 = dy.parameter(attention_w2) v = dy.parameter(attention_v) # input_mat: (encoder_state x seqlen) => input vecs concatenated as cols # w1dt: (attdim x seqlen) # w2dt: (attdim x attdim) w2dt = w2*dy.concatenate(list(state.s())) # att_weights: (seqlen,) row vector unnormalized = dy.transpose(v * dy.tanh(dy.colwise_add(w1dt, w2dt))) att_weights = dy.softmax(unnormalized) # context: (encoder_state) context = input_mat * att_weights return context
def _mlp(self, W1_att_f, W1_att_e, w2_att, h_fs_matrix, h_e, F):
E = W1_att_e * h_e
a = dy.colwise_add(W1_att_f * h_fs_matrix, E)
res = dy.transpose(dy.tanh(a)) * w2_att
return res