Add Paddle Inference docs (PaddlePaddle#307)

* Add get_start notebook example

* fix some bugs of tsdataset api cn docs

* Fix some docs bugs in dataset overview

* fix some bugs of tsdataset api cn docs

* fix some bugs of tsdataset api cn docs

* fix some bugs of tsdataset api cn docs

* Fix bugs in metric for index merge

* Fix bugs in dataset zh api doc

* Remove version res print in get_start docs

* Support Paddle Inference

* Support Paddle Inference

* Support Paddle Inference

* Support Paddle Inference

* Add Paddle Inference docs

* Add Paddle Inference docs

docs/index.rst

@@ -91,6 +91,7 @@ Project GitHub: https://github.com/PaddlePaddle/PaddleTS
     Probability Forecasting <source/modules/models/probability_forecasting.rst>
     Representation  <source/modules/models/representation.rst>
     Anomaly Detection <source/modules/models/anomaly.rst>
+    Paddle Inference <source/modules/models/paddle_inference.rst>
 
 
 .. toctree::

docs/locale/zh_CN/LC_MESSAGES/source/modules/models/paddle_inference.po

@@ -0,0 +1,139 @@
+# SOME DESCRIPTIVE TITLE.
+# Copyright (C) 2022, PaddlePaddle
+# This file is distributed under the same license as the   package.
+# FIRST AUTHOR <EMAIL@ADDRESS>, 2022.
+#
+#, fuzzy
+msgid ""
+msgstr ""
+"Project-Id-Version:  \n"
+"Report-Msgid-Bugs-To: \n"
+"POT-Creation-Date: 2022-12-20 10:20+0800\n"
+"PO-Revision-Date: YEAR-MO-DA HO:MI+ZONE\n"
+"Last-Translator: FULL NAME <EMAIL@ADDRESS>\n"
+"Language-Team: LANGUAGE <LL@li.org>\n"
+"MIME-Version: 1.0\n"
+"Content-Type: text/plain; charset=utf-8\n"
+"Content-Transfer-Encoding: 8bit\n"
+"Generated-By: Babel 2.10.3\n"
+
+#: ../../source/modules/models/paddle_inference.rst:3
+#: 23fde0ece7e741678c6531036f267349
+msgid "Paddle Inference Support"
+msgstr ""
+
+#: ../../source/modules/models/paddle_inference.rst:5
+#: 207df382c8e54baba854247cc1c0e391
+msgid ""
+"The vast majority of models in PaddleTS support Paddle Inference. For the"
+" instruction and functions of Paddle Inference, please refer to `Paddle "
+"Inference "
+"<https://paddleinference.paddlepaddle.org.cn/master/product_introduction/inference_intro.html>`_"
+" . PaddleTS supports export of native Paddle network models for the "
+"deployment of Paddle Inference. To simplify the process, PaddleTS "
+"provides the Python tool to build input data, so users can build input "
+"data for Paddle Inference easily."
+msgstr ""
+"当前的PaddleTS已完成了绝大多数模型对Paddle Inference功能的支持, "
+"Paddle Inference相关介绍以及功能可以参考:  `Paddle Inference <https://paddleinference.paddlepaddle.org.cn/master/product_introduction/inference_intro.html>`_ . "
+"PaddleTS 支持了原生paddle network模型的导出, 可用于Paddle Inference进行推理部署, 同时为了简化用户使用流程, 提供了python的前序数据构建的工具,"
+"可以让用户轻松完成Paddle Inference的输入数据构建."
+
+#: ../../source/modules/models/paddle_inference.rst:12
+#: 2a7ff0d1ec5848a3b702e27e3676bc58
+msgid "1. Build and Train Model"
+msgstr "1. 模型准备以及训练"
+
+#: ../../source/modules/models/paddle_inference.rst:39
+#: 57cf77fe16ad4612aa823b21b622aced
+msgid "2 Save model"
+msgstr "2 模型保存"
+
+#: ../../source/modules/models/paddle_inference.rst:41
+#: 672d76f366954239abc13837e4d06765
+msgid ""
+"`network_model` and `dygraph_to_static` parameters are added in `save` "
+"interfaces of all PaddleTS time-series forecasting and anomaly detection "
+"models. The `network_model` parameter controls which objects are dumped. "
+"The default value of network_model is False, which means the dumped files"
+" can be only used by `PaddleTS.predict` interface. If True, additional "
+"files which can be used by `paddle inference` will be dumped. "
+"`dygraph_to_static` converts the dumped model from a dynamic graph to a "
+"static one, and it works only when network_model is True. For more "
+"information, please refer to `dygraph_to_static "
+"<https://www.paddlepaddle.org.cn/documentation/docs/zh/guides/jit/index_cn.html>`_."
+msgstr ""
+"PaddleTS所有时序预测以及异常检测模型的save接口都新增了 `network_model` 以及 `dygraph_to_static` 的参数设置;"
+"其中, network_model默认是False, 表示仅导出只支持PaddleTS.predict推理的模型文件, 当network_model设置为True的时候, 再次基础上，会"
+"新增对paddle 原始network 的模型以及参数的导出, 可用于 Paddle Inference进行推理; dygraph_to_static参数仅当当network_model为True的"
+"时候起作用，表示将导出的模型从动态图转换成静态图, 参考 `动转静 <https://www.paddlepaddle.org.cn/documentation/docs/zh/guides/jit/index_cn.html>`_."
+
+#: ../../source/modules/models/paddle_inference.rst:55
+#: c5817d208321450db32a297b3dbb11df
+msgid ""
+"The preceding code snippet is an example to dump model files after "
+"saving. `rnn.pdmodel` and `rnn.pdiparams` are native paddle models and "
+"parameters, which can be used for `Paddle Inference`. At the meanwhile, "
+"PaddleTS generates `rnn_model_meta` file for model description with input"
+" data types and metadata of shape. Users can deploy the app correctly in "
+"an easy way."
+msgstr ""
+"上述代码展示了save后的模型文件, 其中rnn.pdmodel以及rnn.pdiparams作为paddle 原生模型以及模型参数, 可用于Paddle Inference的应用;"
+"同时PaddleTS生成了rnn_model_meta文件用于模型的描述, 里面包含了模型的输入数据类型以及shape的各种元信息, 便于用户对模型进行正确的部署应用."
+
+#: ../../source/modules/models/paddle_inference.rst:59
+#: c238de92d88e4c20982c34727b726d4f
+msgid "3. Paddle Inference"
+msgstr ""
+
+#: ../../source/modules/models/paddle_inference.rst:61
+#: 74a54f80f97748ebab0dd9ec8b9cbdfe
+msgid ""
+"With the dumped model in step 2, users can deploy models by `Paddle "
+"Inference`. Here's the example."
+msgstr ""
+"有了第二步导出的模型, 用户即可利用Paddle Inference 进行模型推理部署了，下面给出了简单的示例"
+
+#: ../../source/modules/models/paddle_inference.rst:64
+#: 98eaee212cfa484fa52669f0f99f54be
+msgid "3.1 Load model"
+msgstr "3.1 模型导入"
+
+#: ../../source/modules/models/paddle_inference.rst:83
+#: 0a71194595794d498615c78a64d5f2d1
+msgid ""
+"As the above code snippet shows, we can build the input based on "
+"input_name, which contains attributes of the data "
+"(target、known_cov、observed_cov). In addition to input_name, "
+"rnn_model_meta contains input types, the shape format of data, original "
+"`in_chunk_len` and `out_chunk_len` and so on. With these information, "
+"users can build input data correctly and easily."
+msgstr ""
+"通过上述代码我们可以看到, 我们可以基于input_name去构建我们的输入, 同时input_name本身也包含了数据的属性(target、known_cov、observed_cov), "
+"并且rnn_model_meta文件中除了保函input_name之外, 还包含了其输入类型、数据的shape格式、原始的in_chunk_len、out_chunk_len等信息; 基于这些信息，"
+
+#: ../../source/modules/models/paddle_inference.rst:88
+#: d8ae8acaf28c4360978621407f3313b7
+msgid "3.2 Build Input Tensor"
+msgstr "3.2 构建输入的tensor"
+
+#: ../../source/modules/models/paddle_inference.rst:90
+#: 7883630a56b348f8b6a8c8c5d15dd374
+msgid ""
+"PaddleTS also has built-in functions to build Paddle Inference input "
+"automatically."
+msgstr ""
+"当然, 为了简化用户的使用, PaddleTS也内置了基于TSDataset自动构建Paddle Inference输入的功能"
+
+#: ../../source/modules/models/paddle_inference.rst:107
+#: 5ce56fe707a645a8b5399ae4aa1ed5c6
+msgid "3.3 Inference"
+msgstr "3.3 推理"
+
+#: ../../source/modules/models/paddle_inference.rst:120
+#: a0d58aac48e44a218464b3a189356d86
+msgid ""
+"The results of above code snippet are basically consistent with the "
+"results of predict in chapter 1."
+msgstr ""
+"我们可以看到, 上述程序输出的结果和章节1中predict的结果基本一致"

docs/source/modules/models/paddle_inference.rst

@@ -0,0 +1,122 @@
+========================
+Paddle Inference Support
+========================
+
+The vast majority of models in PaddleTS support Paddle Inference. For the instruction and functions of Paddle Inference, please refer to
+`Paddle Inference <https://paddleinference.paddlepaddle.org.cn/master/product_introduction/inference_intro.html>`_ . 
+PaddleTS supports export of native Paddle network models for the deployment of Paddle Inference. 
+To simplify the process, PaddleTS provides the Python tool to build input data, so users can build input data for Paddle Inference easily.
+
+
+1. Build and Train Model
+========================
+
+.. code-block:: python
+
+    from paddlets.datasets.repository import get_dataset
+    from paddlets.models.forecasting.dl.rnn import RNNBlockRegressor
+
+    # prepare data
+    dataset = get_dataset("WTH")
+
+    rnn = RNNBlockRegressor(
+        in_chunk_len=4, 
+        out_chunk_len=2,
+        max_epochs=10
+    )
+
+    #fit
+    rnn.fit(dataset)
+
+    #predict
+    rnn.predict(dataset)
+
+    #                      WetBulbCelsius
+    # 2014-01-01 00:00:00       -1.436116
+    # 2014-01-01 01:00:00       -2.057547
+
+2 Save model
+============
+
+`network_model` and `dygraph_to_static` parameters are added in `save` interfaces of all PaddleTS time-series forecasting and anomaly detection models.
+The `network_model` parameter controls which objects are dumped. The default value of network_model is False, which means the dumped files can be only used by `PaddleTS.predict` interface. If True, additional files which can be used by `paddle inference` will be dumped. 
+`dygraph_to_static` converts the dumped model from a dynamic graph to a static one, and it works only when network_model is True. 
+For more information, please refer to `dygraph_to_static <https://www.paddlepaddle.org.cn/documentation/docs/zh/guides/jit/index_cn.html>`_. 
+
+
+.. code-block:: python
+    
+    rnn.save("./rnn", network_model=True, dygraph_to_static=True)
+
+    # dump file names
+    # ./rnn.pdmodel
+    # ./rnn.pdiparams
+    # ./rnn_model_meta
+
+The preceding code snippet is an example to dump model files after saving. `rnn.pdmodel` and `rnn.pdiparams` are native paddle models and parameters, which can be used for `Paddle Inference`.
+At the meanwhile, PaddleTS generates `rnn_model_meta` file for model description with input data types and metadata of shape. Users can deploy the app correctly in an easy way.
+
+3. Paddle Inference
+===================
+
+With the dumped model in step 2, users can deploy models by `Paddle Inference`. Here's the example.
+
+3.1 Load model
+---------------
+
+.. code-block:: python
+
+    import paddle.inference as paddle_infer
+    config = paddle_infer.Config("./rnn.pdmodel", "./rnn.pdiparams")
+    predictor = paddle_infer.create_predictor(config)
+    input_names = predictor.get_input_names()
+    print(f"input_name: f{input_names}")
+
+    # input_name: f['observed_cov_numeric', 'past_target']
+
+    import json
+    with open("./rnn_model_meta") as f:
+        json_data = json.load(f)
+        print(json_data)
+    
+    # {'model_type': 'forecasting', 'ancestor_classname_set': ['RNNBlockRegressor', 'PaddleBaseModelImpl', 'PaddleBaseModel', 'BaseModel', 'Trainable', 'ABC', 'object'], 'modulename': 'bts.models.forecasting.dl.rnn', 'size': {'in_chunk_len': 4, 'out_chunk_len': 2, 'skip_chunk_len': 0}, 'input_data': {'past_target': [None, 4, 1], 'observed_cov_numeric': [None, 4, 11]}}
+
+As the above code snippet shows, we can build the input based on input_name, which contains attributes of the data (target、known_cov、observed_cov).
+In addition to input_name, rnn_model_meta contains input types, the shape format of data, original `in_chunk_len` and `out_chunk_len` and so on.
+With these information, users can build input data correctly and easily.
+
+3.2 Build Input Tensor
+----------------------
+
+PaddleTS also has built-in functions to build Paddle Inference input automatically.
+
+.. code-block:: python
+
+    from paddlets.utils.utils import build_ts_infer_input
+    
+    input_data = build_ts_infer_input(dataset, "./rnn_model_meta")
+
+    for key, value in json_data['input_data'].items():
+        input_handle1 = predictor.get_input_handle(key)
+        #set batch_size=1
+        value[0] = 1
+        input_handle1.reshape(value)
+        input_handle1.copy_from_cpu(input_data[key])
+
+
+3.3 Inference
+-------------
+
+.. code-block:: python
+
+    predictor.run()
+    output_names = predictor.get_output_names()
+    output_handle = predictor.get_output_handle(output_names[0])
+    output_data = output_handle.copy_to_cpu() 
+    print(output_data)
+
+    # [[[-1.436116 ]
+    #   [-2.0575469]]]
+
+The results of above code snippet are basically consistent with the results of predict in chapter 1.
+