diff --git a/doc/ui/api/py_data_provider_wrapper.rst b/doc/ui/api/py_data_provider_wrapper.rst deleted file mode 100644 index 91222dd96819f54ef3f1af0a475549937b480476..0000000000000000000000000000000000000000 --- a/doc/ui/api/py_data_provider_wrapper.rst +++ /dev/null @@ -1,6 +0,0 @@ -PyDataProviderWrapper API -========================= - - -.. automodule:: paddle.trainer.PyDataProviderWrapper - :members: diff --git a/doc/ui/data_provider/index.md b/doc/ui/data_provider/index.md deleted file mode 100644 index 49f6e8fbc87edd6fba0c04323140122448277b02..0000000000000000000000000000000000000000 --- a/doc/ui/data_provider/index.md +++ /dev/null @@ -1,55 +0,0 @@ -# DataProvider Tutorial # - -DataProvider is responsible for data management in PaddlePaddle, corresponding to Data Layer. - -## Input Data Format ## -PaddlePaddle uses **Slot** to describe the data layer of neural network. One slot describes one data layer. Each slot stores a series of samples, and each sample contains a set of features. There are three attributes of a slot: -+ **Dimension**: dimenstion of features -+ **SlotType**: there are 5 different slot types in PaddlePaddle, following table compares the four commonly used ones. - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
SlotTypeFeature DescriptionVector Description
DenseSlotContinuous FeaturesDense Vector
SparseNonValueSlotDiscrete Features without weightsSparse Vector with all non-zero elements equaled to 1
SparseValueSlotDiscrete Features with weightsSparse Vector
IndexSlotmostly the same as SparseNonValueSlot, but especially for a single labelSparse Vector with only one value in each time step
-
- -And the remained one is **StringSlot**. It stores Character String, and can be used for debug or to describe data Id for prediction, etc. -+ **SeqType**: a **sequence** is a sample whose features are expanded in time scale. And a **sub-sequence** is a continous ordered subset of a sequence. For example, (a1, a2) and (a3, a4, a5) are two sub-sequences of one sequence (a1, a2, a3, a4, a5). Following are 3 different sequence types in PaddlePaddle: - - **NonSeq**: input sample is not sequence - - **Seq**: input sample is a sequence without sub-sequence - - **SubSeq**: input sample is a sequence with sub-sequence - -## Python DataProvider - -PyDataProviderWrapper is a python decorator in PaddlePaddle, used to read custom python DataProvider class. It currently supports all SlotTypes and SeqTypes of input data. User should only concern how to read samples from file. Feel easy with its [Use Case](python_case.md) and API Reference. diff --git a/doc/ui/data_provider/index.rst b/doc/ui/data_provider/index.rst new file mode 100644 index 0000000000000000000000000000000000000000..db890c2ab1c7198f45e67a802ae64d45adddb701 --- /dev/null +++ b/doc/ui/data_provider/index.rst @@ -0,0 +1,42 @@ +PaddlePaddle DataProvider Introduction +================================ +DataProvider is a module that loads training or testing data into cpu or gpu +memory for the following triaining or testing process. + +For simple use, users can use Python :code:`PyDataProvider` to dynamically reads +the original data in any format or in any form, and then transfer them into a +data format PaddlePaddle requires. The process is extremly flexible and highly +customized, with sacrificing the efficiency only a little. This is extremly +useful when you have to dynamically generate certain kinds of data according to, +for example, the training performance. + +Besides, users also can also customize a C++ :code:`DataProvider` for a more +complex usage, or for a higher efficiency. + +The following parameters are required to define in the PaddlePaddle network +configuration file (trainer_config.py): which DataProvider is chosen to used, +and specific parameters for DataProvider, including training file list +(train.list) and testing file list (test.list). + +Train.list and test.list are simply two plain text files, which defines path +of training or testing data. It is recommended that directly placing them into +the training directory, and reference to them by using a relative path ( +relative to the PaddePaddle program). + +Testing or evaluating will not be performed during training if the test.list is +not set or set to None. Otherwise, PaddlePaddle will evaluate the trained model +by the specified tesing data while training, every testing period (a user +defined command line parameter in PaddlePaddle) to prevent over-fitting. + +Each line of train.list and test.list is an absolute or relative path (relative +to the PaddePaddle program runtime) of data file. Fascinatingly more, each line +can also be a HDFS file path or a SQL connection string. As long as the user +assures how to access each file in DataProvider. + +Please refer to the following articles for more information about the detail +usages of DataProvider and how to implement a new DataProvider, + +.. toctree:: + + pydataprovider2.rst + write_new_dataprovider.rst diff --git a/doc/ui/data_provider/pydataprovider2.rst b/doc/ui/data_provider/pydataprovider2.rst new file mode 100644 index 0000000000000000000000000000000000000000..c4897a7bfc9a02c85c1fde7b0ee82ff21fead54e --- /dev/null +++ b/doc/ui/data_provider/pydataprovider2.rst @@ -0,0 +1,250 @@ +How to use PyDataProvider2 +========================== + +We highly recommand users to use PyDataProvider2 to provide training or testing +data to PaddlePaddle. The user only needs to focus on how to read a single +sample from the original data file by using PyDataProvider2, leaving all of the +trivial work, including, transfering data into cpu/gpu memory, shuffle, binary +serialization to PyDataProvider2. PyDataProvider2 uses multithreading and a +fanscinating but simple cache strategy to optimize the efficiency of the data +providing process. + +DataProvider for the non-sequential model +----------------------------------------- + +Here we use the MNIST handwriting recognition data as an example to illustrate +how to write a simple PyDataProvider. + +MNIST is a handwriting classification data set. It contains 70,000 digital +grayscale images. Labels of the training sample range from 0 to 9. All the +images have been size-normalized and centered into images with a same size +of 28 x 28 pixels. + +A small part of the original data as an example can be found in the path below: + +.. literalinclude:: ../../../doc_cn/ui/data_provider/mnist_train.txt + +Each line of the data contains two parts, separated by ';'. The first part is +label of an image. The second part contains 28x28 pixel float values. + +Just write path of the above data into train.list. It looks like this: + +.. literalinclude:: ../../../doc_cn/ui/data_provider/train.list + +The corresponding dataprovider can be found in the path below: + +.. literalinclude:: ../../../doc_cn/ui/data_provider/mnist_provider.py + : linenos: + +The first line imports PyDataProvider2 package. +The main function is the process function, that has two parameters. +The first parameter is the settings, which is not used in this example. +The second parameter is the filename, that is exactly each line of train.list. +This parameter is passed to the process function by PaddlePaddle. + +:code:`@provider` is a Python +`Decorator `_ . +It sets some properties to DataProvider, and constructs a real PaddlePaddle +DataProvider from a very sample user implemented python function. It does not +matter if you are not familiar with `Decorator`_. You can keep it sample by +just taking :code:`@provider` as a fixed mark above the provider function you +implemented. + +`input_types`_ defines the data format that a DataProvider returns. +In this example, it is set to a 28x28-dimensional dense vector and an integer +scalar, whose value ranges from 0 to 9. +`input_types`_ can be set to several kinds of input formats, please refer to the +document of `input_types`_ for more details. + + +The process method is the core part to construct a real DataProvider in +PaddlePaddle. It implements how to open the text file, how to read one sample +from the original text file, converted them into `input_types`_, and give them +back to PaddlePaddle process at line 23. +Note that data yields by the process function must follow a same order that +`input_types`_ are defined. + + +With the help of PyDataProvider2, user can focus on how to generate ONE traning +sample by using keywords :code:`yield`. +:code:`yield` is a python keyword, and a concept related to it includes +:code:`generator`. + +Only a few lines of codes need to be added into the training configuration file, +you can take this as an example. + +.. literalinclude:: ../../../doc_cn/ui/data_provider/mnist_config.py + +Here we specify training data by 'train.list', and no testing data is specified. + +Now, this simple example of using PyDataProvider is finished. +The only thing that the user should know is how to generte **one sample** from +**one data file**. +And PaddlePadle will do all of the rest things\: + +* Form a training batch +* Shuffle the training data +* Read data with multithreading +* Cache the training data (Optional) +* CPU-> GPU double buffering. + +Is this cool? + +DataProvider for the sequential model +------------------------------------- +A sequence model takes sequences as its input. A sequence is made up of several +timesteps. The so-called timestep, is not necessary to have something to do +with 'time'. It can also be explained to that the order of data are taken into +consideration into model design and training. +For example, the sentence can be interpreted as a kind of sequence data in NLP +tasks. + +Here is an example on data proivider for English sentiment classification data. +The original input data are simple English text, labeled into positive or +negative sentiment (marked by 0 and 1 respectively). + +A small part of the original data as an example can be found in the path below: + +.. literalinclude:: ../../../doc_cn/ui/data_provider/sentimental_train.txt + +The corresponding data provider can be found in the path below: + +.. literalinclude:: ../../../doc_cn/ui/data_provider/sentimental_provider.py + +This data provider for sequential model is a little bit complex than that +for MINST dataset. +A new initialization method is introduced here. +The method :code:`on_init` is configured to DataProvider by :code:`@provider`'s +:code:`init_hook` parameter, and it will be invoked once DataProvider is +initialized. The :code:`on_init` function has the following parameters: + +* The first parameter is the settings object. +* The rest parameters are passed by key word arguments. Some of them are passed + by PaddlePaddle, see reference for `init_hook`_. + The :code:`dictionary` object is a python dict object passed from the trainer + configuration file, and it maps word string to word id. + +To pass these parameters into DataProvider, the following lines should be added +into trainer configuration file. + +.. literalinclude:: ../../../doc_cn/ui/data_provider/sentimental_config.py + +The definition is basically same as MNIST example, except: +* Load dictionary in this configuration +* Pass it as a parameter to the DataProvider + +The `input_types` is configured in method :code:`on_init`. It has the same +effect to configure them by :code:`@provider`'s :code:`input_types` parameter. +However, the :code:`input_types` is set at runtime, so we can set it to +different types according to the input data. Input of the neural network is a +sequence of word id, so set :code:`seq_type` to :code:`integer_value_sequence`. + +Durning :code:`on_init`, we save :code:`dictionary` variable to +:code:`settings`, and it will be used in :code:`process`. Note the settings +parameter for the process function and for the on_init's function are a same +object. + +The basic processing logic is the same as MNIST's :code:`process` method. Each +sample in the data file is given back to PaddlePaddle process. + +Thus, the basic usage of PyDataProvider is here. +Please refer to the following section reference for details. + +Reference +--------- + +.. _@provider:: +@provider ++++++++++ + +'@provider' is a Python `Decorator`_, it can construct a PyDataProvider in +PaddlePaddle from a user defined function. Its parameters are: + +* `input_types`_ defines format of the data input. +* should_shuffle defines whether to shuffle data or not. By default, it is set + true during training, and false during testing. +* pool_size is the memory pool size (in sample number) in DataProvider. + -1 means no limit. +* can_over_batch_size defines whether PaddlePaddle can store little more + samples than pool_size. It is better to set True to avoid some deadlocks. +* calc_batch_size is a function define how to calculate batch size. This is + usefull in sequential model, that defines batch size is counted upon sequence + or token. By default, each sample or sequence counts to 1 when calculating + batch size. +* cache is a data cache strategy, see `cache`_ +* Init_hook function is invoked once the data provider is initialized, + see `init_hook`_ + +.. _input_types:: +input_types ++++++++++++ + +PaddlePaddle has four data types, and three sequence types. +The four data types are: + +* dense_vector represents dense float vector. +* sparse_binary_vector sparse binary vector, most of the value is 0, and + the non zero elements are fixed to 1. +* sparse_float_vector sparse float vector, most of the value is 0, and some + non zero elements that can be any float value. They are given by the user. +* integer represents an integer scalar, that is especially used for label or + word index. + + +The three sequence types are + +* SequenceType.NO_SEQUENCE means the sample is not a sequence +* SequenceType.SEQUENCE means the sample is a sequence +* SequenceType.SUB_SEQUENCE means it is a nested sequence, that each timestep of + the input sequence is also a sequence. + +Different input type has a defferenct input format. Their formats are shown +in the above table. + ++----------------------+---------------------+-----------------------------------+------------------------------------------------+ +| | NO_SEQUENCE | SEQUENCE | SUB_SEQUENCE | ++======================+=====================+===================================+================================================+ +| dense_vector | [f, f, ...] | [[f, ...], [f, ...], ...] | [[[f, ...], ...], [[f, ...], ...],...] | ++----------------------+---------------------+-----------------------------------+------------------------------------------------+ +| sparse_binary_vector | [i, i, ...] | [[i, ...], [i, ...], ...] | [[[i, ...], ...], [[i, ...], ...],...] | ++----------------------+---------------------+-----------------------------------+------------------------------------------------+ +| sparse_float_vector | [(i,f), (i,f), ...] | [[(i,f), ...], [(i,f), ...], ...] | [[[(i,f), ...], ...], [[(i,f), ...], ...],...] | ++----------------------+---------------------+-----------------------------------+------------------------------------------------+ +| integer_value | i | [i, i, ...] | [[i, ...], [i, ...], ...] | ++----------------------+---------------------+-----------------------------------+------------------------------------------------+ + +where f represents a float value, i represents an integer value. + +.. _init_hook:: +.. _settings:: +init_hook ++++++++++ + +init_hook is a function that is invoked once the data provoder is initialized. +Its parameters lists as follows: + +* The first parameter is a settings object, which is the same to :code:'settings' + in :code:`process` method. The object contains several attributes, including: + * settings.input_types the input types. Reference `input_types`_ + * settings.logger a logging object +* The rest parameters are the key word arguments. It is made up of PaddpePaddle + pre-defined parameters and user defined parameters. + * PaddlePaddle defines parameters including: + * is_train is a bool parameter that indicates the DataProvider is used in + training or testing + * file_list is the list of all files. + * User-defined parameters args can be set in training configuration. + +Note, PaddlePaddle reserves the right to add pre-defined parameter, so please +use :code:`**kwargs` in init_hook to ensure compatibility by accepting the +parameters which your init_hook does not use. + +.. _cache :: +cache ++++++ +DataProvider provides two simple cache strategy. They are +* CacheType.NO_CACHE means do not cache any data, then data is read runtime by + the user implemented python module every pass. +* CacheType.CACHE_PASS_IN_MEM means the first pass reads data by the user + implemented python module, and the rest passes will directly read data from + memory. diff --git a/doc/ui/data_provider/python_case.md b/doc/ui/data_provider/python_case.md deleted file mode 100644 index b3ce70522318a249f8063808f47555ca75d5d338..0000000000000000000000000000000000000000 --- a/doc/ui/data_provider/python_case.md +++ /dev/null @@ -1,112 +0,0 @@ -# Python Use Case # - -This tutorial guides you into using python script that converts user input data into PaddlePaddle Data Format. - -## Quick Start ## - -We use a custom data to show the quick usage. This data consists of two parts with semicolon-delimited `';'`: a) label with 2 dimensions, b) continuous features with 9 dimensions: - - 1;0 0 0 0 0.192157 0.070588 0.215686 0.533333 0 - 0;0 0 0 0.988235 0.913725 0.329412 0.376471 0 0 - -The `simple_provider.py` defines a python data provider: - -```python -from trainer.PyDataProviderWrapper import DenseSlot, IndexSlot, provider - -@provider([DenseSlot(9), IndexSlot(2)]) -def process(obj, file_name): - with open(file_name, 'r') as f: - for line in f: - line = line.split(";") - label = int(line[0]) - image = [float(x) for x in line[1].split()[1:]] - yield label, image -``` - -- `@provider`: specify the SlotType and its dimension. Here, we have 2 Slots, DenseSlot(9) stores continuous features with 9 dimensions, and IndexSlot(2) stores label with 2 dimensions. -- `process`: a generator using **yield** keyword to return results one by one. Here, the return format is 1 Discrete Feature and a list of 9 float Continuous Features. - -The corresponding python **Train** data source `define_py_data_source` is: - -```python -define_py_data_sources('train.list', None, 'simple_provider', 'process') -``` -See here for detail API reference of `define_py_data_sources`. - -## Sequence Example ## - -In some tasks such as Natural Language Processing (NLP), the dimension of Slot is related to the dictionary size, and the dictionary should be dynamically loaded during training or generating. PyDataProviderWrapper can satisfy all these demands easily. - -### Sequence has no sub-sequence ### -Following is an example of data provider when using LSTM network to do sentiment analysis (If you want to understand the whole details of this task, please refer to [Sentiment Analysis Tutorial](../demo/sentiment_analysis/index.md)). - -The input data consists of two parts with two-tabs-delimited: a) label with 2 dimensions, b) sequence with dictionary length dimensions: - - 0 I saw this movie at the AFI Dallas festival . It all takes place at a lake house and it looks wonderful . - 1 This documentary makes you travel all around the globe . It contains rare and stunning sequels from the wilderness . - ... - -The `dataprovider.py` in `demo/sentiment` is: - -```python -from trainer.PyDataProviderWrapper import * - -@init_hook_wrapper -def hook(obj, dictionary, **kwargs): - obj.word_dict = dictionary - obj.slots = [IndexSlot(len(obj.word_dict)), IndexSlot(2)] - obj.logger.info('dict len : %d' % (len(obj.word_dict))) - -@provider(use_seq=True, init_hook=hook) -# @provider(use_seq=True, init_hook=hook, pool_size=PoolSize(5000)) -def process(obj, file_name): - with open(file_name, 'r') as fdata: - for line_count, line in enumerate(fdata): - label, comment = line.strip().split('\t\t') - label = int(''.join(label.split(' '))) - words = comment.split() - word_slot = [obj.word_dict[w] for w in words if w in obj.word_dict] - yield word_slot, [label] -``` - -- `hook`: Initialization hook of data provider. Here, it reads the dictionary, sets the obj.slots based on the dictionary length, and uses obj.logger to output some logs. -- `process`: Here, as the Sequence Mode of input is **Seq** and SlotType is IndexSlot, use_seq is set to True, and the yield format is `[int, int, ....]`. -- `PoolSize`: If there are a lot of data, you may need this argument to increase loading speed and reduce memory footprint. Here, PoolSize(5000) means read at most 5000 samples to memory. - -The corresponding python **Train/Test** data sources `define_py_data_sources` is: - -```python -train_list = train_list if not is_test else None -word_dict = dict() -with open(dict_file, 'r') as f: - for i, line in enumerate(open(dict_file, 'r')): - word_dict[line.split('\t')[0]] = i - -define_py_data_sources(train_list, test_list, module = "dataprovider", obj = "processData", - args = {'dictionary': word_dict}, train_async = True) -``` - -### Sequence has sub-sequence ### - -If the sequence of above input data is considered as several sub-sequences joint by dot `'.'`, quesion mark `'?'` or exclamation mark `'!'`, see `processData2` in `demo/sentiment/dataprovider.py` as follows: - -```python -import re - -@provider(use_seq=True, init_hook=hook) -def process2(obj, file_name): - with open(file_name, 'r') as fdata: - pat = re.compile(r'[^.?!]+[.?!]') - for line_count, line in enumerate(fdata): - label, comment = line.strip().split('\t\t') - label = int(''.join(label.split(' '))) - words_list = pat.findall(comment) - word_slot_list = [[obj.word_dict[w] for w in words.split() \ - if w in obj.word_dict] for words in words_list] - yield word_slot_list, [[label]] -``` - -- `hook`: the same as above. Note that as **SubSeq Slot must put before Seq Slot** in PaddlePaddle, we could not reverse the yield order in this case. -- `process2`: Here, as the Sequence Mode of input is **SubSeq**, and the SlotType is IndexSlot, use_seq is set to True, and the yield format is `[[int, int, ...], [int, int, ...], ... ]`. -- `define_py_data_sources`: the same as above. diff --git a/doc/ui/index.md b/doc/ui/index.md index 03699b6832018f15da8287124d2b29bc5ae89258..829994d56bb65414c6bdd9d4ce2a0878ee7c21d7 100644 --- a/doc/ui/index.md +++ b/doc/ui/index.md @@ -2,12 +2,11 @@ ## Data Provider -* [Introduction](data_provider/index.md) -* [Python Use Case](data_provider/python_case.md) +* [Introduction](data_provider/index.rst) +* [PyDataProvider2](data_provider/pydataprovider2.rst) ## API Reference -* [PyDataProviderWrapper](api/py_data_provider_wrapper.rst) * [Trainer Config Helpers](api/trainer_config_helpers/index.md) ## Command Line Argument