def _attention(self, direct, cur_token, prev, to_apply, to_apply_proj):
with layer.mixed(size=cur_token.size,
bias_attr=Attr.Param(direct + '.bp', initial_std=0.),
act=Act.Linear()) as proj:
proj += layer.full_matrix_projection(input=cur_token,
param_attr=Attr.Param(direct +
'.wp'))
proj += layer.full_matrix_projection(input=prev,
param_attr=Attr.Param(direct +
'.wr'))
expanded = layer.expand(input=proj, expand_as=to_apply)
att_context = layer.addto(input=[expanded, to_apply_proj],
act=Act.Tanh(),
bias_attr=False)
att_weights = layer.fc(input=att_context,
param_attr=Attr.Param(direct + '.w'),
bias_attr=Attr.Param(direct + '.b',
initial_std=0.),
act=Act.SequenceSoftmax(),
size=1)
scaled = layer.scaling(input=to_apply, weight=att_weights)
applied = layer.pooling(input=scaled,
pooling_type=paddle.pooling.Sum())
return applied
def network(self):
"""
Implements the whole network of Match-LSTM.
Returns:
A tuple of LayerOutput objects containing the start and end
probability distributions respectively.
"""
self.check_and_create_data()
self.create_shared_params()
q_enc = self.get_enc(self.q_ids, type='q')
p_encs = []
p_matches = []
for p in self.p_ids:
p_encs.append(self.get_enc(p, type='p'))
q_proj_left = layer.fc(size=self.emb_dim * 2,
bias_attr=False,
param_attr=Attr.Param(self.name + '_left_' +
'.wq'),
input=q_enc)
q_proj_right = layer.fc(size=self.emb_dim * 2,
bias_attr=False,
param_attr=Attr.Param(self.name + '_right_' +
'.wq'),
input=q_enc)
for i, p in enumerate(p_encs):
left_out = self.recurrent_group(
self.name + '_left_' + str(i),
[layer.StaticInput(q_enc),
layer.StaticInput(q_proj_left), p],
reverse=False)
right_out = self.recurrent_group(
self.name + '_right_' + str(i),
[layer.StaticInput(q_enc),
layer.StaticInput(q_proj_right), p],
reverse=True)
match_seq = layer.concat(input=[left_out, right_out])
match_seq_dropped = self.drop_out(match_seq, drop_rate=0.5)
bi_match_seq = paddle.networks.bidirectional_lstm(
input=match_seq_dropped,
size=match_seq.size,
fwd_mat_param_attr=Attr.Param('pn_f_enc_mat.w'),
fwd_bias_param_attr=Attr.Param('pn_f_enc.bias',
initial_std=0.),
fwd_inner_param_attr=Attr.Param('pn_f_enc_inn.w'),
bwd_mat_param_attr=Attr.Param('pn_b_enc_mat.w'),
bwd_bias_param_attr=Attr.Param('pn_b_enc.bias',
initial_std=0.),
bwd_inner_param_attr=Attr.Param('pn_b_enc_inn.w'),
return_seq=True)
p_matches.append(bi_match_seq)
all_docs = reduce(lambda x, y: layer.seq_concat(a=x, b=y), p_matches)
all_docs_dropped = self.drop_out(all_docs, drop_rate=0.5)
start = self.decode('start', all_docs_dropped)
end = self.decode('end', all_docs_dropped)
return start, end
def _step_basic(self, h_cur, u):
expanded_h = layer.expand(input=h_cur, expand_as=u)
hu = layer.concat(input=[expanded_h, u])
with layer.mixed(bias_attr=False) as dot_hu:
dot_hu += layer.dotmul_operator(a=expanded_h, b=u)
cat_all = layer.concat(input=[hu, dot_hu])
s = layer.fc(size=1,
bias_attr=False,
param_attr=Attr.Param(self.name + '.ws'),
input=cat_all)
return s
def _step(self, name, h_q_all, q_proj, h_p_cur):
"""
Match-LSTM step. This function performs operations done in one
time step.
Args:
h_p_cur: Current hidden of paragraph encodings: h_i.
This is the `REAL` input of the group, like
x_t in normal rnn.
h_q_all: Question encodings.
Returns:
The $h^{r}_{i}$ in the paper.
"""
direct = 'left' if 'left' in name else 'right'
h_r_prev = paddle.layer.memory(name=name + '_out_',
size=h_q_all.size,
boot_layer=None)
q_expr = self._attention(direct, h_p_cur, h_r_prev, h_q_all, q_proj)
z_cur = self.fusion_layer(h_p_cur, q_expr)
with layer.mixed(size=h_q_all.size * 4,
act=Act.Tanh(),
bias_attr=False) as match_input:
match_input += layer.full_matrix_projection(
input=z_cur,
param_attr=Attr.Param('match_input_%s.w0' % direct))
step_out = paddle.networks.lstmemory_unit(
name=name + '_out_',
out_memory=h_r_prev,
param_attr=Attr.Param('step_lstm_%s.w' % direct),
input_proj_bias_attr=Attr.Param('step_lstm_mixed_%s.bias' % direct,
initial_std=0.),
lstm_bias_attr=Attr.Param('step_lstm_%s.bias' % direct,
initial_std=0.),
input=match_input,
size=h_q_all.size)
return step_out
def network(self):
"""
Implements the detail of the model.
"""
self.check_and_create_data()
self.create_shared_params()
q_enc = self.get_enc(self.q_ids, type='q')
a_enc = self.get_enc(self.a_ids, type='q')
q_proj_left = layer.fc(size=self.emb_dim * 2,
bias_attr=False,
param_attr=Attr.Param(self.name + '_left.wq'),
input=q_enc)
q_proj_right = layer.fc(size=self.emb_dim * 2,
bias_attr=False,
param_attr=Attr.Param(self.name + '_right.wq'),
input=q_enc)
left_match = self.recurrent_group(
self.name + '_left',
[layer.StaticInput(q_enc),
layer.StaticInput(q_proj_left), a_enc],
reverse=False)
right_match = self.recurrent_group(
self.name + '_right',
[layer.StaticInput(q_enc),
layer.StaticInput(q_proj_right), a_enc],
reverse=True)
match_seq = layer.concat(input=[left_match, right_match])
with layer.mixed(size=match_seq.size,
act=Act.Identity(),
layer_attr=Attr.ExtraLayerAttribute(drop_rate=0.2),
bias_attr=False) as dropped:
dropped += layer.identity_projection(match_seq)
match_result = layer.pooling(input=dropped,
pooling_type=paddle.pooling.Max())
cls = layer.fc(input=match_result,
act=Act.Softmax(),
size=self.label_dim)
return cls
def network(self, question, evidence, qe_comm, ee_comm, conf):
"""
Implements the whole network of Match-LSTM.
Returns:
A tuple of LayerOutput objects containing the start and end
probability distributions respectively.
"""
q_enc = self.get_enc(question, conf, type='q')
p_enc = self.get_enc(evidence, conf, type='q')
q_proj_left = layer.fc(size=conf.word_vec_dim * 2,
bias_attr=False,
param_attr=Attr.Param(self.name + '_left_' +
'.wq'),
input=q_enc)
q_proj_right = layer.fc(size=conf.word_vec_dim * 2,
bias_attr=False,
param_attr=Attr.Param(self.name + '_right_' +
'.wq'),
input=q_enc)
# StaticInput 定义了一个只读的Memory，由StaticInput指定的输入不会被recurrent_group拆解，
# recurrent_group 循环展开的每个时间步总是能够引用所有输入，可以是一个非序列，或者一个单层序列。
left_out = self.recurrent_group(self.name + '_left', [
layer.StaticInput(q_enc),
layer.StaticInput(q_proj_left), p_enc, qe_comm, ee_comm
],
reverse=False)
right_out = self.recurrent_group(self.name + '_right_', [
layer.StaticInput(q_enc),
layer.StaticInput(q_proj_right), p_enc, qe_comm, ee_comm
],
reverse=True)
match_seq = layer.concat(input=[left_out, right_out])
return self.drop_out(match_seq, drop_rate=0.5)
def _get_enc(self, input, type='q'):
embs = self.get_embs(input)
enc = networks.bidirectional_lstm(
input=embs,
size=self.emb_dim,
fwd_mat_param_attr=Attr.Param(self.name + '_f_enc_mat.w' + type),
fwd_bias_param_attr=Attr.Param(self.name + '_f_enc.bias' + type,
initial_std=0.),
fwd_inner_param_attr=Attr.Param(self.name + '_f_enc_inn.w' + type),
bwd_mat_param_attr=Attr.Param(self.name + '_b_enc_mat.w' + type),
bwd_bias_param_attr=Attr.Param(self.name + '_b_enc.bias' + type,
initial_std=0.),
bwd_inner_param_attr=Attr.Param(self.name + '_b_enc_inn.w' + type),
return_seq=True)
enc_dropped = self.drop_out(enc, drop_rate=0.25)
return enc_dropped
def get_enc(self, input, type='q'):
"""
Encodes the input by feeding it into a bidirectional lstm and
concatenates the forward and backward expression of each time step.
"""
embs = self.get_embs(input)
enc = paddle.networks.bidirectional_lstm(
input=embs,
size=self.emb_dim,
fwd_mat_param_attr=Attr.Param('f_enc_mat.w' + type),
fwd_bias_param_attr=Attr.Param('f_enc.bias' + type,
initial_std=0.),
fwd_inner_param_attr=Attr.Param('f_enc_inn.w' + type),
bwd_mat_param_attr=Attr.Param('b_enc_mat.w' + type),
bwd_bias_param_attr=Attr.Param('b_enc.bias' + type,
initial_std=0.),
bwd_inner_param_attr=Attr.Param('b_enc_inn.w' + type),
return_seq=True)
enc_dropped = self.drop_out(enc, drop_rate=0.5)
return enc_dropped
def network(self):
"""
Implements the whole network.
Returns:
A tuple of LayerOutput objects containing the start and end
probability distributions respectively.
"""
self.check_and_create_data()
self.create_shared_params()
u = self._get_enc(self.q_ids, type='q')
m1s = []
m2s = []
for p in self.p_ids:
h = self._get_enc(p, type='q')
g = self._attention_flow(h, u)
m1 = networks.bidirectional_lstm(
fwd_mat_param_attr=Attr.Param('_f_m1_mat.w'),
fwd_bias_param_attr=Attr.Param('_f_m1.bias', initial_std=0.),
fwd_inner_param_attr=Attr.Param('_f_m1_inn.w'),
bwd_mat_param_attr=Attr.Param('_b_m1_mat.w'),
bwd_bias_param_attr=Attr.Param('_b_m1.bias', initial_std=0.),
bwd_inner_param_attr=Attr.Param('_b_m1_inn.w'),
input=g,
size=self.emb_dim,
return_seq=True)
m1_dropped = self.drop_out(m1, drop_rate=0.)
cat_g_m1 = layer.concat(input=[g, m1_dropped])
m2 = networks.bidirectional_lstm(
fwd_mat_param_attr=Attr.Param('_f_m2_mat.w'),
fwd_bias_param_attr=Attr.Param('_f_m2.bias', initial_std=0.),
fwd_inner_param_attr=Attr.Param('_f_m2_inn.w'),
bwd_mat_param_attr=Attr.Param('_b_m2_mat.w'),
bwd_bias_param_attr=Attr.Param('_b_m2.bias', initial_std=0.),
bwd_inner_param_attr=Attr.Param('_b_m2_inn.w'),
input=m1,
size=self.emb_dim,
return_seq=True)
m2_dropped = self.drop_out(m2, drop_rate=0.)
cat_g_m2 = layer.concat(input=[g, m2_dropped])
m1s.append(cat_g_m1)
m2s.append(cat_g_m2)
all_m1 = reduce(lambda x, y: layer.seq_concat(a=x, b=y), m1s)
all_m2 = reduce(lambda x, y: layer.seq_concat(a=x, b=y), m2s)
start = self.decode('start', all_m1)
end = self.decode('end', all_m2)
return start, end
import paddle.v2.data_type as data_type
import paddle.v2.layer as layer
import paddle.v2.pooling as pooling
import paddle.v2.networks as networks
pixel = layer.data(name='pixel', type=data_type.dense_vector(128))
label = layer.data(name='label', type=data_type.integer_value(10))
weight = layer.data(name='weight', type=data_type.dense_vector(1))
combine_weight = layer.data(name='weight_combine',
type=data_type.dense_vector(10))
score = layer.data(name='score', type=data_type.dense_vector(1))
hidden = layer.fc(input=pixel,
size=100,
act=activation.Sigmoid(),
param_attr=attr.Param(name='hidden'))
inference = layer.fc(input=hidden, size=10, act=activation.Softmax())
conv = layer.img_conv(input=pixel,
filter_size=1,
filter_size_y=1,
num_channels=8,
num_filters=16,
act=activation.Linear())
class ImageLayerTest(unittest.TestCase):
def test_conv_layer(self):
conv_shift = layer.conv_shift(a=pixel, b=score)
print layer.parse_network(conv, conv_shift)
def test_pooling_layer(self):
def _step(self, name, h_q_all, q_proj, h_p_cur, qe_comm, ee_comm):
"""
Match-LSTM step. This function performs operations done in one
time step.
Args:
h_p_cur: Current hidden of paragraph encodings: h_i.
This is the `REAL` input of the group, like
x_t in normal rnn.
h_q_all: Question encodings.
Returns:
The $h^{r}_{i}$ in the paper.
"""
conf = mLSTM_crf_config.TrainingConfig()
direct = 'left' if 'left' in name else 'right'
# 获取上一个时间步的输出
h_r_prev = paddle.layer.memory(name=name + '_out_',
size=h_q_all.size,
boot_layer=None)
# h_p_cur :: Current hidden of paragraph encodings
# h_q_all :: q wordEmbedding
# q_proj :: q_proj_(left or right)
q_expr = self._attention(direct, h_p_cur, h_r_prev, h_q_all, q_proj)
z_cur = self.fusion_layer(h_p_cur, q_expr)
# feature embeddings
comm_initial_std = 1 / math.sqrt(64.0)
qe_comm_emb = paddle.layer.embedding(input=qe_comm,
size=conf.com_vec_dim,
param_attr=paddle.attr.ParamAttr(
name="_cw_embedding.w0",
initial_std=comm_initial_std,
l2_rate=conf.default_l2_rate))
ee_comm_emb = paddle.layer.embedding(input=ee_comm,
size=conf.com_vec_dim,
param_attr=paddle.attr.ParamAttr(
name="_eecom_embedding.w0",
initial_std=comm_initial_std,
l2_rate=conf.default_l2_rate))
# layer.mixed :: 综合输入映射到指定维度，为 lstm 的输入做准备！
with layer.mixed(size=h_q_all.size * 4,
act=Act.Tanh(),
bias_attr=False) as match_input:
match_input += layer.full_matrix_projection(
input=z_cur,
param_attr=Attr.Param('match_input_z_%s.w0' % direct))
match_input += layer.full_matrix_projection(
input=qe_comm_emb,
param_attr=Attr.Param('match_input_qe_%s.w0' % direct))
match_input += layer.full_matrix_projection(
input=ee_comm_emb,
param_attr=Attr.Param('match_input_ee_%s.w0' % direct))
step_out = paddle.networks.lstmemory_unit(
name=name + '_out_',
out_memory=h_r_prev,
param_attr=Attr.Param('step_lstm_%s.w' % direct),
input_proj_bias_attr=Attr.Param('step_lstm_mixed_%s.bias' % direct,
initial_std=0.),
lstm_bias_attr=Attr.Param('step_lstm_%s.bias' % direct,
initial_std=0.),
input=match_input,
size=h_q_all.size)
return step_out
