Skip to content

Commit

Permalink
Supplement doc for RNN (PaddlePaddle#214)
Browse files Browse the repository at this point in the history
  • Loading branch information
Zrachel authored and luotao1 committed Oct 18, 2016
1 parent 28bc05b commit 45280a0
Showing 1 changed file with 137 additions and 1 deletion.
138 changes: 137 additions & 1 deletion doc_cn/algorithm/rnn/hierarchical-rnn.md
Original file line number Diff line number Diff line change
Expand Up @@ -260,7 +260,143 @@ out = recurrent_group(step=outer_step, input=SubsequenceInput(emb))

## 示例3:双进双出,输入不等长

TBD
**输入不等长**是指recurrent_group的多个输入在各时刻的长度可以不相等, 但需要指定一个和输出长度一致的input,用<font color="red">targetInlink</font>表示。参考配置:单层RNN(`sequence_rnn_multi_unequalength_inputs.conf`),双层RNN(`sequence_nest_rnn_multi_unequalength_inputs.conf`

### 读取双层序列的方法

我们看一下单双层序列的数据组织形式和dataprovider(见`rnn_data_provider.py`
```python
data2 = [
[[[1, 2], [4, 5, 2]], [[5, 4, 1], [3, 1]] ,0],
[[[0, 2], [2, 5], [0, 1, 2]],[[1, 5], [4], [2, 3, 6, 1]], 1],
]

@provider(input_types=[integer_value_sub_sequence(10),
integer_value_sub_sequence(10),
integer_value(2)],
should_shuffle=False)
def process_unequalength_subseq(settings, file_name): #双层RNN的dataprovider
for d in data2:
yield d


@provider(input_types=[integer_value_sequence(10),
integer_value_sequence(10),
integer_value(2)],
should_shuffle=False)
def process_unequalength_seq(settings, file_name): #单层RNN的dataprovider
for d in data2:
words1=reduce(lambda x,y: x+y, d[0])
words2=reduce(lambda x,y: x+y, d[1])
yield words1, words2, d[2]
```

data2 中有两个样本,每个样本有两个特征, 记fea1, fea2。

- 单层序列:两个样本分别为[[1, 2, 4, 5, 2], [5, 4, 1, 3, 1]][[0, 2, 2, 5, 0, 1, 2], [1, 5, 4, 2, 3, 6, 1]]
- 双层序列:两个样本分别为
- **样本1**[[[1, 2], [4, 5, 2]], [[5, 4, 1], [3, 1]]]。fea1和fea2都分别有2个子句,fea1=[[1, 2], [4, 5, 2]], fea2=[[5, 4, 1], [3, 1]]
- **样本2**[[[0, 2], [2, 5], [0, 1, 2]],[[1, 5], [4], [2, 3, 6, 1]]]。fea1和fea2都分别有3个子句, fea1=[[0, 2], [2, 5], [0, 1, 2]], fea2=[[1, 5], [4], [2, 3, 6, 1]]。<br/>
- **注意**:每个样本中,各特征的子句数目需要相等。这里说的“双进双出,输入不等长”是指fea1在i时刻的输入的长度可以不等于fea2在i时刻的输入的长度。如对于第1个样本,时刻i=2, fea1[2]=[4, 5, 2],fea2[2]=[3, 1],3≠2。
- 单双层序列中,两个样本的label都分别是0和1

### 模型中的配置

单层RNN(`sequence_rnn_multi_unequalength_inputs.conf`)和双层RNN(`sequence_nest_rnn_multi_unequalength_inputs.conf`)两个模型配置达到的效果完全一样,区别只在于输入为单层还是双层序列,现在我们来看它们内部分别是如何实现的。

- 单层序列:
- 过了一个简单的recurrent_group。每一个时间步,当前的输入y和上一个时间步的输出rnn_state做了一个全连接,功能与示例2中`sequence_rnn.conf``step`函数完全相同。这里,两个输入x1,x2分别通过calrnn返回最后时刻的状态。结果得到的encoder1_rep和encoder2_rep分别是单层序列,最后取encoder1_rep的最后一个时刻和encoder2_rep的所有时刻分别相加得到context。
- 注意到这里recurrent_group输入的每个样本中,fea1和fea2的长度都分别相等,这并非偶然,而是因为recurrent_group要求输入为单层序列时,所有输入的长度都必须相等。

```python
def step(x1, x2):
def calrnn(y):
mem = memory(name = 'rnn_state_' + y.name, size = hidden_dim)
out = fc_layer(input = [y, mem],
size = hidden_dim,
act = TanhActivation(),
bias_attr = True,
name = 'rnn_state_' + y.name)
return out

encoder1 = calrnn(x1)
encoder2 = calrnn(x2)
return [encoder1, encoder2]

encoder1_rep, encoder2_rep = recurrent_group(
name="stepout",
step=step,
input=[emb1, emb2])

encoder1_last = last_seq(input = encoder1_rep)
encoder1_expandlast = expand_layer(input = encoder1_last,
expand_as = encoder2_rep)
context = mixed_layer(input = [identity_projection(encoder1_expandlast),
identity_projection(encoder2_rep)],
size = hidden_dim)
```
- 双层序列:
- 双层RNN中,对输入的两个特征分别求时序上的连续全连接(`inner_step1``inner_step2`分别处理fea1和fea2),其功能与示例2中`sequence_nest_rnn.conf``outer_step`函数完全相同。不同之处是,此时输入`[SubsequenceInput(emb1), SubsequenceInput(emb2)]`在各时刻并不等长。
- 函数`outer_step`中可以分别处理这两个特征,但我们需要用<font color=red>targetInlink</font>指定recurrent_group的输出的格式(各子句长度)只能和其中一个保持一致,如这里选择了和emb2的长度一致。
- 最后,依然是取encoder1_rep的最后一个时刻和encoder2_rep的所有时刻分别相加得到context。

```python
def outer_step(x1, x2):
outer_mem1 = memory(name = "outer_rnn_state1", size = hidden_dim)
outer_mem2 = memory(name = "outer_rnn_state2", size = hidden_dim)
def inner_step1(y):
inner_mem = memory(name = 'inner_rnn_state_' + y.name,
size = hidden_dim,
boot_layer = outer_mem1)
out = fc_layer(input = [y, inner_mem],
size = hidden_dim,
act = TanhActivation(),
bias_attr = True,
name = 'inner_rnn_state_' + y.name)
return out

def inner_step2(y):
inner_mem = memory(name = 'inner_rnn_state_' + y.name,
size = hidden_dim,
boot_layer = outer_mem2)
out = fc_layer(input = [y, inner_mem],
size = hidden_dim,
act = TanhActivation(),
bias_attr = True,
name = 'inner_rnn_state_' + y.name)
return out

encoder1 = recurrent_group(
step = inner_step1,
name = 'inner1',
input = x1)

encoder2 = recurrent_group(
step = inner_step2,
name = 'inner2',
input = x2)

sentence_last_state1 = last_seq(input = encoder1, name = 'outer_rnn_state1')
sentence_last_state2_ = last_seq(input = encoder2, name = 'outer_rnn_state2')

encoder1_expand = expand_layer(input = sentence_last_state1,
expand_as = encoder2)

return [encoder1_expand, encoder2]

encoder1_rep, encoder2_rep = recurrent_group(
name="outer",
step=outer_step,
input=[SubsequenceInput(emb1), SubsequenceInput(emb2)],
targetInlink=emb2)

encoder1_last = last_seq(input = encoder1_rep)
encoder1_expandlast = expand_layer(input = encoder1_last,
expand_as = encoder2_rep)
context = mixed_layer(input = [identity_projection(encoder1_expandlast),
identity_projection(encoder2_rep)],
size = hidden_dim)
```

## 示例4:beam_search的生成

Expand Down

0 comments on commit 45280a0

Please sign in to comment.