def _net_conf(word_ids, target):
"""
Configure the network
"""
word_embedding = fluid.layers.embedding(
input=word,
size=[word_dict_len, word_emb_dim],
dtype='float32',
is_sparse=IS_SPARSE,
param_attr=fluid.ParamAttr(
learning_rate=emb_lr,
name="word_emb",
initializer=fluid.initializer.Uniform(
low=-init_bound, high=init_bound)))
# add elmo embedding
elmo_emb = elmo_encoder(word_ids, args.elmo_l2_coef)
input_feature = layers.concat(input=[elmo_emb, word_embedding], axis=1)
for i in range(bigru_num):
bigru_output = _bigru_layer(input_feature)
input_feature = bigru_output
emission = fluid.layers.fc(
size=label_dict_len,
input=bigru_output,
param_attr=fluid.ParamAttr(
initializer=fluid.initializer.Uniform(
low=-init_bound, high=init_bound),
regularizer=fluid.regularizer.L2DecayRegularizer(
regularization_coeff=1e-4)))
crf_cost = fluid.layers.linear_chain_crf(
input=emission,
label=target,
param_attr=fluid.ParamAttr(
name='crfw', learning_rate=crf_lr))
crf_decode = fluid.layers.crf_decoding(
input=emission, param_attr=fluid.ParamAttr(name='crfw'))
avg_cost = fluid.layers.mean(x=crf_cost)
return avg_cost, crf_decode
def rc_model(hidden_size, vocab, args):
emb_shape = [vocab.size(), vocab.embed_dim]
start_labels = layers.data(
name="start_lables", shape=[1], dtype='float32', lod_level=1)
end_labels = layers.data(
name="end_lables", shape=[1], dtype='float32', lod_level=1)
vocab_size=52445
# stage 1:encode
q_id0 = get_data('q_id0', 1, args)
q_ids = get_data('q_ids', 2, args)
p_ids_name = 'p_ids'
p_ids = get_data('p_ids', 2, args)
q_ids_elmo = get_data('q_ids_elmo', 2, args)
p_ids_elmo = get_data('p_ids_elmo', 2, args)
#layers.Print(p_ids_elmo, message='p_ids_elmo', summarize=10)
#layers.Print(p_ids, message='p_ids', summarize=10)
#layers.Print(q_ids_elmo, message='q_ids_elmo', summarize=10)
#layers.Print(q_ids, message='q_ids', summarize=10)
p_embs = embedding(p_ids, emb_shape, args)
q_embs = embedding(q_ids, emb_shape, args)
if args.elmo==True:
q_embs_elmo = emb(q_ids_elmo)
p_embs_elmo = emb(p_ids_elmo)
drnn = layers.DynamicRNN()
with drnn.block():
p_emb = drnn.step_input(p_embs)
q_emb = drnn.step_input(q_embs)
if args.elmo==True:
q_emb_elmo = drnn.step_input(q_embs_elmo)
p_emb_elmo = drnn.step_input(p_embs_elmo)
p_encs_elmo= elmo_encoder(p_emb_elmo)
q_encs_elmo= elmo_encoder(q_emb_elmo)
#layers.Print(p_encs_elmo, message='p_encs_elmo', summarize=10)
#layers.Print(q_encs_elmo, message='q_encs_elmo', summarize=10)
#layers.Print(p_emb, message='p_emb', summarize=10)
p_emb=layers.concat(input=[p_emb, p_emb_elmo], axis=1)
q_emb=layers.concat(input=[q_emb, q_emb_elmo], axis=1)
p_enc = encoder(p_emb,'p_enc', hidden_size, args)
q_enc = encoder(q_emb, 'q_enc', hidden_size, args)
g_i = attn_flow(q_enc, p_enc, p_ids_name, args)
# stage 3:fusion
m_i = fusion(g_i, args)
drnn.output(m_i, q_enc)
ms, q_encs = drnn()
p_vec = layers.lod_reset(x=ms, y=start_labels)
q_vec = layers.lod_reset(x=q_encs, y=q_id0)
# stage 4:decode
start_probs, end_probs = point_network_decoder(
p_vec=p_vec, q_vec=q_vec, hidden_size=hidden_size, args=args)
cost0 = layers.sequence_pool(
layers.cross_entropy(
input=start_probs, label=start_labels, soft_label=True),
'sum')
cost1 = layers.sequence_pool(
layers.cross_entropy(
input=end_probs, label=end_labels, soft_label=True),
'sum')
cost0 = layers.mean(cost0)
cost1 = layers.mean(cost1)
cost = cost0 + cost1
cost.persistable = True
feeding_list=[]
if args.elmo==True:
feeding_list = ["q_ids", "start_lables", "end_lables", "p_ids", "q_id0","q_ids_elmo","p_ids_elmo"]
else:
feeding_list = ["q_ids", "start_lables", "end_lables", "p_ids", "q_id0"]
return cost, start_probs, end_probs, ms, feeding_list