Source code of this chpater is in [book/label_semantic_roles](https://github.com/PaddlePaddle/book/tree/develop/label_semantic_roles).
Source code of this chapter is in [book/label_semantic_roles](https://github.com/PaddlePaddle/book/tree/develop/label_semantic_roles).
## Background
Natural Language Analysis contains three components: Lexical Analysis, Syntactic Analysis, and Semantic Analysis. Semantic Role Labelling (SRL) is one way for Shallow Semantic Analysis. A predicate of a sentence is seen as a property that a subject has or is characterized by, such as what it does, what it is or how it is, which mostly corresponds to the core of an event. The noun associated with predicate is called Arugment. Sementic roles express the abstract roles that arguments of a predicate can take in the event, such as Agent, Patient, Theme, Experiencer, Beneficiary, Instrument, Location, Goal and Source etc.
Natural Language Analysis contains three components: Lexical Analysis, Syntactic Analysis, and Semantic Analysis. Semantic Role Labelling (SRL) is one way for Shallow Semantic Analysis. A predicate of a sentence is a property that a subject possesses or is characterized, such as what it does, what it is or how it is, which mostly corresponds to the core of an event. The noun associated with a predicate is called Argument. Semantic roles express the abstract roles that arguments of a predicate can take in the event, such as Agent, Patient, Theme, Experiencer, Beneficiary, Instrument, Location, Goal and Source, etc.
In the following example, “遇到” is Predicate (“Pred”)，“小明” is Agent，“小红” is Patient，“昨天” means when the event occurs (Time), and “公园” means where the event occurs (Location).
In the following example, “遇到” (encounters) is a Predicate (“Pred”)，“小明” (Ming) is an Agent，“小红” (Hong) is a Patient，“昨天” (yesterday) indicates the Time, and “公园” (park) is the Location.
Instead of in-depth analysis on semantic information, the goal of Semantic Role Labeling is to identify the relation of predicate and other constituents, e.g., predicate-argument structure, as specific semantic roles, which is an important intermediate step in a wide range of natural language understanding tasks (Information Extraction, Discourse Analysis, DeepQA etc). Predicates are always assumed to be given, the only thing is to identify arguments and their semantic roles.
Instead of in-depth analysis on semantic information, the goal of Semantic Role Labeling is to identify the relation of predicate and other constituents, e.g., predicate-argument structure, as specific semantic roles, which is an important intermediate step in a wide range of natural language understanding tasks (Information Extraction, Discourse Analysis, DeepQA etc). Predicates are always assumed to be given; the only thing is to identify arguments and their semantic roles.
Standard SRL system mostly build on top of Syntactic Analysis and contains 5 steps:
Standard SRL system mostly builds on top of Syntactic Analysis and contains five steps:
1. Construct a syntactic parse tree, as shown in Fig. 1
2. Identity candidate arguments of given predicate from constructed syntactic parse tree.
3. Prune most unlikely candidate arguments.
4. Identify argument, which is usually solved as a binary classification problem.
4. Identify arguments, often by a binary classifier.
5. Multi-class semantic role labeling. Steps 2-3 usually introduce hand-designed features based on Syntactic Analysis (step 1).
...
...
@@ -77,7 +77,7 @@ Fig 1. Syntactic parse tree
标点-> WP
However, complete syntactic analysis requires to identify the relation among all constitutes and the performance of SRL is sensitive to the precision of syntactic analysis, which make SRL a very challenging task. In order to reduce the complexity and obtain some syntactic structure information, shallow syntactic analysis is proposed. Shallow Syntactic Analysis is also called partial parsing or chunking. Unlike complete syntactic analysis which requires constructing complete parsing tree, Shallow Syntactic Analysis only need to identify some idependent components with relatively simple structure, such as verb phrases (chunk). In order to avoid constructing syntactic tree with high accuracy, some work\[[1](#Reference)\] proposed semantic chunking based SRL methods, which convert SRL as a sequence tagging problem. Sequence tagging tasks classify syntactic chunks using BIO representation. For syntactic chunks forming a chunk of type A, the first chunk receives the B-A tag (Begin), the remaining ones receive the tag I-A (Inside), and all chunks outside receive the tag O-A.
However, complete syntactic analysis requires identifying the relation among all constitutes and the performance of SRL is sensitive to the precision of syntactic analysis, which makes SRL a very challenging task. To reduce the complexity and obtain some syntactic structure information, we often use shallow syntactic analysis. Shallow Syntactic Analysis is also called partial parsing or chunking. Unlike complete syntactic analysis which requires the construction of the complete parsing tree, Shallow Syntactic Analysis only need to identify some independent components with relatively simple structure, such as verb phrases (chunk). To avoid difficulties in constructing a syntactic tree with high accuracy, some work\[[1](#Reference)\] proposed semantic chunking based SRL methods, which convert SRL as a sequence tagging problem. Sequence tagging tasks classify syntactic chunks using BIO representation. For syntactic chunks forming a chunk of type A, the first chunk receives the B-A tag (Begin), the remaining ones receive the tag I-A (Inside), and all chunks outside receive the tag O-A.
The BIO representation of above example is shown in Fig.1.
...
...
@@ -91,23 +91,23 @@ Fig 2. BIO represention
标注序列-> label sequence
角色-> role
This example illustrates the simplicity of sequence tagging because (1) shallow syntactic analysis reduces precision requirement of syntactic analysis; (2) pruning candidate arguments is removed; 3) argument identification and tagging are finished at the same time. Such unified methods simplify the precedure, reduce the risk of accumulating errors and boost the performance further.
This example illustrates the simplicity of sequence tagging because (1) shallow syntactic analysis reduces the precision requirement of syntactic analysis; (2) pruning candidate arguments is removed; 3) argument identification and tagging are finished at the same time. Such unified methods simplify the procedure, reduce the risk of accumulating errors and boost the performance further.
In this tutorial, our SRL system is built as an end-to-end system via neural network. We take only text sequences, without using any syntactic parsing results or complex hand-designed features. We give public dataset [CoNLL-2004 and CoNLL-2005 Shared Tasks](http://www.cs.upc.edu/~srlconll/) as an example to illustrate: given a sentence and it's predicates, identify the corresponding arguments and their semantic roles by sequence tagging method.
In this tutorial, our SRL system is built as an end-to-end system via a neural network. We take only text sequences, without using any syntactic parsing results or complex hand-designed features. We give public dataset [CoNLL-2004 and CoNLL-2005 Shared Tasks](http://www.cs.upc.edu/~srlconll/) as an example to illustrate: given a sentence with predicates marked, identify the corresponding arguments and their semantic roles by sequence tagging method.
## Model
Recurrent Nerual Networks are important tools for sequence modeling and have been successfully used in some natural language processing tasks. Unlike Feed-forward neural netowrks, RNNs can model the dependency between elements of sequences. LSTMs as variants of RNNs aim to model long-term dependency in long sequences. We have introduced this in [understand_sentiment](https://github.com/PaddlePaddle/book/tree/develop/understand_sentiment). In this chapter, we continue to use LSTMs to solve SRL problems.
Recurrent Neural Networks are important tools for sequence modeling and have been successfully used in some natural language processing tasks. Unlike Feed-forward neural networks, RNNs can model the dependency between elements of sequences. LSTMs as variants of RNNs aim to model long-term dependency in long sequences. We have introduced this in [understand_sentiment](https://github.com/PaddlePaddle/book/tree/develop/understand_sentiment). In this chapter, we continue to use LSTMs to solve SRL problems.
### Stacked Recurrent Neural Network
Deep Neural Networks allows to extract hierarchical represetations, higher layer can form more abstract/complex representations on top of lower layers. LSTMs when unfolded in time is deep, because a computational path between the input at time $k <t$totheoutputattime$t$crossesseveralnonlinearlayers.However,thecomputationcarriedoutateachtime-stepisonlylineartransformation,whichmakesLSTMsashallowmodel.DeepLSTMsaretypicallyconstructedbystackingmultipleLSTMlayersontopofeachotherandtakingtheoutputfromlowerLSTMlayerattime$t$astheinputofupperLSTMlayerattime$t$.Deep,hierarchicalnerualnetworkscanbemuchefficientatrepresentingsomefunctionsandmodelingvarying-lengthdependencies\[[2](#Reference)\].
Deep Neural Networks allows extracting hierarchical representations. Higher layers can form more abstract/complex representations on top of lower layers. LSTMs, when unfolded in time, is a deep feed-forward neural network, because a computational path between the input at time $k <t$totheoutputattime$t$crossesseveralnonlinearlayers.However,thecomputationcarriedoutateachtime-stepisonlylineartransformation,whichmakesLSTMsashallowmodel.DeepLSTMsaretypicallyconstructedbystackingmultipleLSTMlayersontopofeachotherandtakingtheoutputfromlowerLSTMlayerattime$t$astheinputofupperLSTMlayerattime$t$.Deep,hierarchicalneuralnetworkscanbemuchefficientatrepresentingsomefunctionsandmodelingvarying-lengthdependencies\[[2](#Reference)\].
LSTMs can summarize the history of previous inputs seen up to now, but can not see the future. In most of natural language processing tasks, the entire sentences are ready to use. Therefore, sequencal learning might be much effecient if the future can be encoded as well like histories.
LSTMs can summarize the history of previous inputs seen up to now, but can not see the future. In most of NLP (natural language processing) tasks, the entire sentences are ready to use. Therefore, sequential learning might be much efficient if the future can be encoded as well like histories.
To address the above drawbacks, we can design bidirectional recurrent neural networks by making a minor modification. Higher LSTM layers process the sequence in reversed direction with previous lower LSTM layers, i.e., Deep LSTMs operate from left-to-right, right-to-left, left-to-right,..., in depth. Therefore, LSTM layers at time-step $t$ can see both histories and the future since the second layer. Fig. 4 illustrates the bidirectional recurrent neural networks.
...
...
@@ -133,10 +133,10 @@ Fig 4. Bidirectional LSTMs
线性变换-> linear transformation
输入层到隐层-> input-to-hidden
正向处理输出序列->process sequence in forward direction
反向处理上一层序列-> process sequence from previous layer in backward direction
正向处理输出序列->process sequence in the forward direction
反向处理上一层序列-> process sequence from the previous layer in backward direction
Note that, this bidirectional RNNs is different with the one proposed by Bengio et al in machine translation tasks \[[3](#Reference), [4](#Reference)\]. We will introduce another bidirectional RNNs in the following tasks[machine translation](https://github.com/PaddlePaddle/book/blob/develop/machine_translation/README.md)
Note that, this bidirectional RNNs is different with the one proposed by Bengio et al. in machine translation tasks \[[3](#Reference), [4](#Reference)\]. We will introduce another bidirectional RNNs in the following tasks[machine translation](https://github.com/PaddlePaddle/book/blob/develop/machine_translation/README.md)
### Conditional Random Field
...
...
@@ -145,7 +145,7 @@ The basic pipeline of Neural Networks solving problems is 1) all lower layers ai
CRF is a probabilistic graph model (undirected) with nodes denoting random variables and edges denoting dependencies between nodes. To be simplicity, CRFs learn conditional probability $P(Y|X)$, where $X = (x_1, x_2, ... , x_n)$ are sequences of input, $Y = (y_1, y_2, ... , y_n)$ are label sequences; Decoding is to search sequence $Y$ to maximize conditional probability $P(Y|X)$, i.e., $Y^* = \mbox{arg max}_{Y} P(Y | X)$。
Sequence tagging tasks only consider input and output as linear sequences without extra dependent assumptions on graph model. Thus, the graph model of sequence tagging tasks is simple chain or line, which results in a LinearChain Conditional Random Field, shown in Fig.5.
Sequence tagging tasks only consider input and output as linear sequences without extra dependent assumptions on graph model. Thus, the graph model of sequence tagging tasks is simple chain or line, which results in a Linear-Chain Conditional Random Field, shown in Fig.5.
@@ -174,25 +174,25 @@ This objective function can be solved via back-propagation in an end-to-end mann
Given predicates and a sentence, SRL tasks aim to identify arguments of the given predicate and their semantic roles. If a sequence has n predicates, we will process this sequence n times. One model is as follows:
1. Construct inputs;
- output 1: predicate, output 2: sentence
- input 1: predicate, input 2: sentence
- expand input 1 as a sequence with the same length with input 2 using one-hot representation;
2. Convert one-hot sequences from step 1 to real-vector sequences via lookup table;
3. Learn the representation of input sequences by taking real-vector sequences from step 2 as inputs;
2. Convert one-hot sequences from step 1 to vector sequences via lookup table;
3. Learn the representation of input sequences by taking vector sequences from step 2 as inputs;
4. Take representations from step 3 as inputs, label sequence as supervision signal, do sequence tagging tasks
We can try above method. Here, we propose some modifications by introducing two simple but effective features:
- predicate context (ctx-p): A single predicate word can not exactly describe the predicate information, especially when the same words appear more than one times in a sentence. With the expanded context, the ambiguity can be largely eliminated. Thus, we extract $n$ words before and after predicate to construct a window chunk.
- region mark ($m_r$): $m_r = 1$ to denote the argument position if it locates in the predicate context region, or $m_r = 0$ if not.
- region mark ($m_r$): $m_r = 1$ to denote word in that position locates in the predicate context region, or $m_r = 0$ if not.
After modification, the model is as follows:
1. Construct inputs
- input 1: sentence, input 2: predicate sequence, input 3: predicate context, extract $n$ words before and after predicate and get one-hot representation, input 4: region mark, annotate argument position if it locates in the predicate context region
- Input 1: word sequence. Input 2: predicate. Input 3: predicate context, extract $n$ words before and after predicate. Input 4: region mark sequence, element value will be 1 if word locates in the predicate context region, 0 otherwise.
- expand input 2~3 as sequences with the same length with input 1
2. Convert input 1~4 to real-vector sequences via lookup table; input 1 and 3 share the same lookup table, input 2 and 4 have separate lookup tables
3. Take four real-vector sequences from step 2 as inputs of bidirectional LSTMs; Train LSTMs to update representations
2. Convert input 1~4 to vector sequences via lookup table; input 1 and 3 shares the same lookup table, input 2 and 4 have separate lookup tables
3. Take four vector sequences from step 2 as inputs of bidirectional LSTMs; Train LSTMs to update representations
4. Take representation from step 3 as input of CRF, label sequence as supervision signal, do sequence tagging tasks
...
...
@@ -209,12 +209,11 @@ Fig 6. DB-LSTM for SRL tasks
原句-> sentence
反向LSTM-> LSTM Reverse
## 数据准备
### 数据介绍与下载
## Data Preparation
在此教程中，我们选用[CoNLL 2005](http://www.cs.upc.edu/~srlconll/)SRL任务开放出的数据集作为示例。运行 `sh ./get_data.sh` 会自动从官方网站上下载原始数据。需要特别说明的是，CoNLL 2005 SRL任务的训练数集和开发集在比赛之后并非免费进行公开，目前，能够获取到的只有测试集，包括Wall Street Journal的23节和Brown语料集中的3节。在本教程中，我们以测试集中的WSJ数据为训练集来讲解模型。但是，由于测试集中样本的数量远远不够，如果希望训练一个可用的神经网络SRL系统，请考虑付费获取全量数据。
In the tutorial, we use [CoNLL 2005](http://www.cs.upc.edu/~srlconll/) SRL task open dataset as an example. It is important to note that the training set and development set of the CoNLL 2005 SRL task are not free to download after the competition. Currently, only the test set can be obtained, including 23 sections of the Wall Street Journal and three sections of the Brown corpus. In this tutorial, we use the WSJ corpus as the training dataset to explain the model. However, since the training set is small, if you want to train a usable neural network SRL system, consider paying for the full corpus.
The original data includes a variety of information such as POS tagging, naming entity recognition, parsing tree, and so on. In this tutorial, we only use the data under the words folder (text sequence) and the props folder (label results) inside test.wsj parent folder. The data directory used in this tutorial is as follows:
The annotation information is derived from the results of Penn TreeBank\[[7](#references)\] and PropBank \[[8](# references)\]. The label of the PropBank is different from the label that we used in the example at the beginning of the article, but the principle is the same. For the description of the label, please refer to the paper \[[9](#references)\].
除数据之外，`get_data.sh`同时下载了以下资源：
The raw data needs to be preprocessed before used by PaddlePaddle. The preprocessing consists of the following steps:
1. Merge the text sequence and the tag sequence into the same record;
2. If a sentence contains $n$ predicates, the sentence will be processed $n$ times into $n$ separate training samples, each sample with a different predicate;
3. Extract the predicate context and construct the predicate context region marker;
4. Construct the markings in BIO format;
5. Obtain the integer index corresponding to the word according to the dictionary.
After preprocessing completes, a training sample contains nine features, namely: word sequence, predicate, predicate context (5 columns), region mark sequence, label sequence. Following table is an example of a training sample.
| 句子序列 | 谓词 | 谓词上下文（窗口 = 5） | 谓词上下文区域标记 | 标注序列 |
| word sequence | predicate | predicate context（5 columns） | region mark sequence | label sequence|
In addition to the data, we provide following resources:
| filename | explanation |
|---|---|
| word_dict | dictionary of input sentences, total 44068 words |
| label_dict | dictionary of labels, total 106 labels |
| predicate_dict | predicate dictionary, total 3162 predicates |
| emb | a pre-trained word vector lookup table, 32-dimentional |
We trained in the English Wikipedia language model to get a word vector lookup table used to initialize the SRL model. During the SRL model training process, the word vector lookup table is no longer updated. About the language model and the word vector lookup table can refer to [word vector](https://github.com/PaddlePaddle/book/blob/develop/word2vec/README.md) tutorial. There are 995,000,000 token in training corpus, and the dictionary size is 4900,000 words. In the CoNLL 2005 training corpus, 5% of the words are not in the 4900,000 words, and we see them all as unknown words, represented by `<unk>`.
4. We will concatenate the output of top LSTM unit with it's input, and project into a hidden layer. Then put a fully connected layer on top of it to get the final vector representation.
```python
feature_out = paddle.layer.mixed(
size=label_dict_len,
bias_attr=std_default,
input=[
paddle.layer.full_matrix_projection(
input=input_tmp[0], param_attr=hidden_para_attr),
paddle.layer.full_matrix_projection(
input=input_tmp[1], param_attr=lstm_para_attr)
], )
```
5. We use CRF as cost function, the parameter of CRF cost will be named `crfw`.
```python
crf_cost = paddle.layer.crf(
size=label_dict_len,
input=feature_out,
label=target,
param_attr=paddle.attr.Param(
name='crfw',
initial_std=default_std,
learning_rate=mix_hidden_lr))
```
6. CRF decoding layer is used for evaluation and inference. It shares parameter with CRF layer. The sharing of parameters among multiple layers is specified by the same parameter name in these layers.
```python
crf_dec = paddle.layer.crf_decoding(
name='crf_dec_l',
size=label_dict_len,
input=feature_out,
label=target,
param_attr=paddle.attr.Param(name='crfw'))
```
## Train model
### Create Parameters
All necessary parameters will be traced created given output layers that we need to use.
We will create trainer given model topology, parameters and optimization method. We will use most basic SGD method (momentum optimizer with 0 momentum). In the meantime, we will set learning rate and regularization.
As mentioned in data preparation section, we will use CoNLL 2005 test corpus as training data set. `conll05.test()` outputs one training instance at a time. It will be shuffled, and batched into mini batches as input.
```python
crf_dec_l = crf_decoding_layer(
name='crf_dec_l',
size=label_dict_len,
input=feature_out,
param_attr=ParameterAttribute(name='crfw'))
reader = paddle.reader.batched(
paddle.reader.shuffle(
conll05.test(), buf_size=8192), batch_size=20)
```
运行`python predict.py`脚本，便可使用指定的模型进行预测。
```bash
python predict.py
-c db_lstm.py # 指定配置文件
-w output/pass-00100 # 指定预测使用的模型所在的路径
-l data/targetDict.txt # 指定标记的字典
-p data/verbDict.txt # 指定谓词的词典
-d data/wordDict.txt # 指定输入文本序列的字典
-i data/feature # 指定输入数据的路径
-o predict.res # 指定标记结果输出到文件的路径
`reader_dict` is used to specify relationship between data instance and layer layer. For example, according to following `reader_dict`, the 0th column of data instance produced by`conll05.test()` correspond to data layer named `word_data`.
```python
reader_dict = {
'word_data': 0,
'ctx_n2_data': 1,
'ctx_n1_data': 2,
'ctx_0_data': 3,
'ctx_p1_data': 4,
'ctx_p2_data': 5,
'verb_data': 6,
'mark_data': 7,
'target': 8
}
```
预测结束后，在 - o 参数所指定的标记结果文件中，我们会得到如下格式的输出：每行是一条样本，以 “\t” 分隔的 2 列，第一列是输入文本，第二列是标记的结果。通过BIO标记可以直接得到论元的语义角色标签。
`event_handle` can be used as callback for training events, it will be used as an argument for `train`. Following `event_handle` prints cost during training.
```text
The interest-only securities were priced at 35 1\/2 to yield 10.72 % . B-A0 I-A0 I-A0 O O O O O O B-V B-A1 I-A1 O
```python
def event_handler(event):
if isinstance(event, paddle.event.EndIteration):
if event.batch_id % 100 == 0:
print "Pass %d, Batch %d, Cost %f" % (
event.pass_id, event.batch_id, event.cost)
```
`trainer.train` will train the model.
```python
trainer.train(
reader=reader,
event_handler=event_handler,
num_passes=10000,
reader_dict=reader_dict)
```
## Conclusion
Semantic Role Labeling is an important intermediate step in a wide range of natural language processing tasks. In this tutorial, we give SRL as an example to introduce how to use PaddlePaddle to do sequence tagging tasks. Proposed models are from our published paper\[[10](#Reference)\]. We only use test data as illustration since train data on CoNLL 2005 dataset is not completely public. We hope to propose an end-to-end neural network model with less dependencies on natural language processing tools, but is comparable, or even better than trandional models. Please check out our paper for more information and discussions.
Semantic Role Labeling is an important intermediate step in a wide range of natural language processing tasks. In this tutorial, we give SRL as an example to introduce how to use PaddlePaddle to do sequence tagging tasks. Proposed models are from our published paper\[[10](#Reference)\]. We only use test data as an illustration since train data on CoNLL 2005 dataset is not completely public. We hope to propose an end-to-end neural network model with fewer dependencies on natural language processing tools but is comparable, or even better than traditional models. Please check out our paper for more information and discussions.
## Reference
1. Sun W, Sui Z, Wang M, et al. [Chinese semantic role labeling with shallow parsing](http://www.aclweb.org/anthology/D09-1#page=1513)[C]//Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing: Volume 3-Volume 3. Association for Computational Linguistics, 2009: 1475-1483.
PaddlePaddle provides a special layer `layer.data` for reading data. Let us create a data layer for reading images and connect it to a classification network created using one of above three functions. We also need a cost layer for training the model.
Now, it is time to specify training parameters. The number 0.9 in the following `Momentum` optimizer means that 90% of the current the momentum comes from the momentum of the previous iteration.
Then we specify the training data `paddle.dataset.movielens.train()` and testing data `paddle.dataset.movielens.test()`. These two functions are *reader creators*, once called, returns a *reader*. A reader is a Python function, which, once called, returns a Python generator, which yields instances of data.
...
...
@@ -233,48 +232,48 @@ Here `shuffle` is a reader decorator, which takes a reader A as its parameter, a
`batch` is a special decorator, whose input is a reader and output is a *batch reader*, which doesn't yield an instance at a time, but a minibatch.
# Test with Pass 0, Cost 0.326659, {'classification_error_evaluator': 0.09470000118017197}
```
After the training, we can check the model's prediction accuracy.
```
# find the best pass
best = sorted(lists, key=lambda list: float(list[1]))[0]
print 'Best pass is %s, testing Avgcost is %s' % (best[0], best[1])
print 'The classification accuracy is %.2f%%' % (100 - float(best[2]) * 100)
# find the best pass
best = sorted(lists, key=lambda list: float(list[1]))[0]
print 'Best pass is %s, testing Avgcost is %s' % (best[0], best[1])
print 'The classification accuracy is %.2f%%' % (100 - float(best[2]) * 100)
```
Usually, with MNIST data, the softmax regression model can get accuracy around 92.34%, MLP can get about 97.66%, and convolution network can get up to around 99.20%. Convolution layers have been widely considered a great invention for image processsing.
@@ -156,15 +156,8 @@ For more information, please refer to [Activation functions on Wikipedia](https:
## Data Preparation
### Data Download
PaddlePaddle provides a Python module, `paddle.dataset.mnist`, which downloads and caches the [MNIST dataset](http://yann.lecun.com/exdb/mnist/). The cache is under `/home/username/.cache/paddle/dataset/mnist`:
Execute the following command to download the [MNIST](http://yann.lecun.com/exdb/mnist/) dataset and unzip. Add paths to the training set and the test set to train.list and test.list respectively for PaddlePaddle to read.
```bash
./data/get_mnist_data.sh
```
`gzip` downloaded data. The following files can be found in `data/raw_data`:
In the model configuration, use `define_py_data_sources2` to define reading of data from `dataprovider`. If this configuration is used for prediction, data definition is not necessary.
```python
if not is_predict:
data_dir = './data/'
define_py_data_sources2(
train_list=data_dir + 'train.list',
test_list=data_dir + 'test.list',
module='mnist_provider',
obj='process')
```
### Algorithm Configuration
## Model Configuration
Set training related parameters.
- batch_size: use 128 samples in each training step.
- learning_rate: determines step taken in each iteration, it determines how fast the model converges.
- learning_method: use optimizer `MomentumOptimizer` for training. The parameter 0.9 indicates momentum keeps 0.9 of previous speed.
- regularization: A method to prevent overfitting. Here L2 regularization is used.
A PaddlePaddle program starts from importing the API package:
```python
settings(
batch_size=128,
learning_rate=0.1 / 128.0,
learning_method=MomentumOptimizer(0.9),
regularization=L2Regularization(0.0005 * 128))
import paddle.v2 as paddle
```
### Model Architecture
We want to use this program to demonstrate multiple kinds of models. Let define each of them as a Python function:
#### Overview
First get reference labels from `data_layer`, and get classification results (predictions) from classifier. Here we provide three different classifiers. In training, we compute loss function, which is usually cross entropy for classification problem. In prediction, we can directly output the results (predictions).
The following code implements a Multilayer Perceptron with two fully connected hidden layers and a ReLU activation function. The output layer has a Softmax activation function.
- multi-layer perceptron: this network has two hidden fully-connected layers, one with LeRU and the other with softmax activation:
The following is the LeNet-5 network architecture. A 2D input image is first fed into two sets of convolutional layers and pooling layers, this result is then fed to a fully connected layer, and another fully connected layer with a softmax activation.
- convolution network LeNet-5: the input image is fed through two convolution-pooling layer, a fully-connected layer, and the softmax output layer:
After configuring parameters, execute `./train.sh`. Training log is as follows.
PaddlePaddle provides a special layer `layer.data` for reading data. Let us create a data layer for reading images and connect it to a classification network created using one of above three functions. We also need a cost layer for training the model.
Now, it is time to specify training parameters. The number 0.9 in the following `Momentum` optimizer means that 90% of the current the momentum comes from the momentum of the previous iteration.
From the evaluation results, the best pass for softmax regression model is pass-00013, where the classification accuracy is 90.01%, and the last pass-00099 has an accuracy of 89.3%. From Fig. 7, we also see that the best accuracy may not appear in the last pass. This is because during training, the model may already arrive at a local optimum, and it just swings around nearby in the following passes, or it gets a lower local optimum.
Then we specify the training data `paddle.dataset.movielens.train()` and testing data `paddle.dataset.movielens.test()`. These two functions are *reader creators*, once called, returns a *reader*. A reader is a Python function, which, once called, returns a Python generator, which yields instances of data.
### Results of Multilayer Perceptron
Here `shuffle` is a reader decorator, which takes a reader A as its parameter, and returns a new reader B, where B calls A to read in `buffer_size` data instances everytime into a buffer, then shuffles and yield instances in the buffer. If you want very shuffled data, try use a larger buffer size.
From the evaluation results, the final training accuracy is 94.95%. It is significantly better than the softmax regression model. This is because the softmax regression is simple, and it cannot fit complex data. The Multilayer Perceptron with hidden layers has better capacity to fit complex data than the softmax regression.
### Training results for Convolutional Neural Network
During training, `trainer.train` invokes `event_handler` for certain events. This gives us a chance to print the training progress.
Results of model evaluation：
```text
Best pass is 00076, testing Avgcost is 0.0244684
The classification accuracy is 99.20%
```
From the evaluation result, the best accuracy of Convolutional Neural Network is 99.20%. So for image classification, a Convolutional Neural Network has better recognition results than a fully connected network. This is related to the local connection and parameter sharing of convolutional layers. In Fig. 9, the Convolutional Neural Network achieves good results in early steps, which indicates that it converges faster.
## Application Model
### Prediction Commands and Results
Script `predict.py` can make prediction for trained models. For example, in softmax regression:
# Test with Pass 0, Cost 0.326659, {'classification_error_evaluator': 0.09470000118017197}
```
- -c sets model architecture
- -d sets data for prediction
- -m sets model parameters, here the best trained model is used for prediction
Follow the instructions to input image ID for prediction. The classifier can output probabilities for each digit, predictions with the highest probability, and ground truth label.
After the training, we can check the model's prediction accuracy.
best = sorted(lists, key=lambda list: float(list[1]))[0]
print 'Best pass is %s, testing Avgcost is %s' % (best[0], best[1])
print 'The classification accuracy is %.2f%%' % (100 - float(best[2]) * 100)
```
From the result, this classifier recognizes the digit on the third image as digit 0 with near to 100% probability. This predicted result is consistent with the ground truth label.
Usually, with MNIST data, the softmax regression model can get accuracy around 92.34%, MLP can get about 97.66%, and convolution network can get up to around 99.20%. Convolution layers have been widely considered a great invention for image processsing.
## Conclusion
This tutorial describes a few basic Deep Learning models viz. Softmax regression, Multilayer Perceptron Network and Convolutional Neural Network. The subsequent tutorials will derive more sophisticated models from these. So it is crucial to understand these models for future learning. When our model evolved from a simple softmax regression to slightly complex Convolutional Neural Network, the recognition accuracy on the MNIST data set achieved large improvement in accuracy. This is due to the Convolutional layers' local connections and parameter sharing. While learning new models in the future, we encourage the readers to understand the key ideas that lead a new model to improve results of an old one. Moreover, this tutorial introduced the basic flow of PaddlePaddle model design, starting with a dataprovider, model layer construction, to final training and prediction. Readers can leverage the flow used in this MNIST handwritten digit classification example and experiment with different data and network architectures to train models for classification tasks of their choice.
