def global_attention(query):
# linear map
y = Linear(query, global_attention_vec_size, True)
y = y.view(-1, 1, 1, global_attention_vec_size)
# Attention mask is a softmax of v_g^{\top} * tanh(...)
s = torch.sum(global_v *
torch.tanh(global_hidden_features + y),
dim=[1, 3])
a = tf.softmax(s)
return a
def local_attention(query):
# linear map
y = Linear(query, local_attention_vec_size, True)
y = y.view(-1, 1, 1, local_attention_vec_size)
# Attention mask is a softmax of v_l^{\top} * tanh(...)
#print((local_v * torch.tanh(local_hidden_features + y)).size())
s = torch.sum(local_v * torch.tanh(local_hidden_features + y),
dim=[1, 3])
# Now calculate the attention-weighted vector, i.e., alpha in eq.[2]
a = tf.softmax(s)
return a
def attention(query):
# linear map
y = Linear(query, attention_vec_size, True)
y = y.view(-1, 1, 1, attention_vec_size)
# Attention mask is a softmax of v_d^{\top} * tanh(...).
s = torch.sum(v * torch.tanh(hidden_features + y), dim=[1, 3])
# Now calculate the attention-weighted vector, i.e., gamma in eq.[7]
a = tf.softmax(s)
# eq. [8]
#print(hidden.size())
#print((a.view(-1, 1, attn_length, 1)).size())
d = torch.sum(a.view(-1, 1, attn_length, 1) * hidden, dim=[2, 3])
return d.view(-1, attn_size)