Update Docs.

DataCanvasIO · Jul 19, 2022 · 784a6fb · 784a6fb
1 parent 13c8229
commit 784a6fb
Show file tree

Hide file tree

Showing 8 changed files with 101 additions and 33 deletions.
diff --git a/docs/en_US/source/conf.py b/docs/en_US/source/conf.py
@@ -23,7 +23,7 @@
 author = 'DataCanvas.com'
 
 # The full version, including alpha/beta/rc tags
-release = '0.1.0'
+release = '0.1.3'
 
 
 # -- General configuration ---------------------------------------------------

diff --git a/docs/en_US/source/contents/0600_custom_functions.rst b/docs/en_US/source/contents/0600_custom_functions.rst
@@ -91,9 +91,6 @@ HyperTS provides various algorithms with default search space for every mode. Mo
 
 By setting the argument ``search_space``, users could define their own search space. The instructions and an example are given below to modify the ``StatsForecastSearchSpace``. 
 
-- Mandatory! Describe the exact name of the task. For instance, ``task='univariate-forecast'``;
-- Mandatory! Assign the name of the ``timestamp`` column;
-- Mandatory! If have the covariables, describe them clearly. For instance, ``covariables={xxx: xxx, ...}``;
 - Set the argument as false to disable a certain algorithm. For instance, ``enable_arima=False``;
 - Change the initial parameters of a certain algorithm by function ``prophet_init_kwargs={xxx:xxx, ...}``;
 - Import the argument ``Choice``, ``Int`` ``Real`` from ``hypernets.core.search_space`` could define the parameters with specific options. For instance, ``Choice`` supports the boolean data type. ``Real`` supports the floating data type.
@@ -106,11 +103,8 @@ Code example:
     from hypernets.core.search_space import Choice, Int, Real
     from hyperts.framework.search_space.macro_search_space import StatsForecastSearchSpace
 
-    custom_search_space = StatsForecastSearchSpace(task='univariate-forecast', 
-                                                timestamp='TimeStamp',
-                                                covariables=['HourSin', 'WeekCos', 'CBWD'],
-                                                enable_arima=False,
-                                                prophet_init_kwargs={
+    custom_search_space = StatsForecastSearchSpace(enable_arima=False,
+                                                   prophet_init_kwargs={
                                                     'seasonality_mode': 'multiplicative',
                                                     'daily_seasonality': Choice([True, False]),
                                                     'n_changepoints': Int(10, 50, step=10),
@@ -374,10 +368,7 @@ Add the estimator to the search space, in which the hyperparameters also could b
     df = load_network_traffic(univariate=True)
     train_data, test_data = train_test_split(df, test_size=168, shuffle=False)
 
-    custom_search_space = DLForecastSearchSpacePlusTransformer(task='univariate-forecast', 
-                                                            timestamp='TimeStamp',
-                                                            covariables=['HourSin', 'WeekCos', 'CBWD'],
-                                                            metrics=['mape'])
+    custom_search_space = DLForecastSearchSpacePlusTransformer()
 
     experiment = make_experiment(train_data, 
                                 task='univariate-forecast',

diff --git a/docs/en_US/source/contents/0700_models.rst b/docs/en_US/source/contents/0700_models.rst
@@ -155,4 +155,11 @@ For more information, please refer to the paper `Modeling Long- and Short-Term T
 
 Neural Architecture Search
 *************
-...
+Since AlexNet won the 2012 ImageNet competition, deep learning has made breakthroughs in many challenging tasks and fields. In addition to AlexNet,
+e.g., VGG, Inception, ResNet, Transformer, GPT and so on have been proposed and widely used in industry and academia. And then, behind all these great networks,
+it is the crystallization of the experience of countless human experts. Consequently, neural architecture search (NAS) has emerged as a promising tool to alleviate human efforts in this trial-and-error design process.
+
+HyperTS relies on the basic capabilities provided by Hypernets (``Hpyer Model`` + ``Search Strategy`` + ``Estimation Strategy``), to build ``Search Space`` for time series tasks based on NAS powerful expression capabilities.
+
+.. tip::
+    NAS is applied to uni/multi-variate forecasting and regression.
diff --git a/docs/en_US/source/contents/0800_release_note.rst b/docs/en_US/source/contents/0800_release_note.rst
@@ -79,4 +79,40 @@ Details of the HypertTS update are as follows:
 
 - Optimizes search space and architecture.
 
-- Fixes some bugs.
+- Fixes some bugs.
+
+
+Version 0.1.3
+******************
+
+Details of the HypertTS update are as follows:
+
+- Tuning search space hyperparameters;
+
+- Added report_best_trial_params;
+
+- Fixed ARIMA to be compatible with version 0.12.1 and above;
+
+- Fixed the pt issue of LSTNet;
+
+- Simplified custom search space, task, timestamp, covariables and metircs can not be passed;
+
+- Added OutliersTransformer, supported dynamic handling of outliers;
+
+- Adjusted final train processing - lr, batch_size, epcochs and so on;
+
+- Added time series meta-feature extractor;
+
+- Added Time2Vec, RevIN, etc. layers;
+
+- Added N-Beats time series forecasting model;
+
+- Added InceptionTime time series classification model;
+
+- Supported dynamic downsampling for time series forecasting;
+
+- Refactored positive label inference method;
+
+- Added neural architecture search mode;
+
+- Fixed some known issues.
diff --git a/docs/zh_CN/source/conf.py b/docs/zh_CN/source/conf.py
@@ -23,7 +23,7 @@
 author = 'DataCanvas.com'
 
 # The full version, including alpha/beta/rc tags
-release = '0.1.0'
+release = '0.1.3'
 
 
 # -- General configuration ---------------------------------------------------

diff --git a/docs/zh_CN/source/contents/0600_user_defined.rst b/docs/zh_CN/source/contents/0600_user_defined.rst
@@ -93,10 +93,6 @@ HyperTS针对不同的模式内置了丰富的建模算法, 例如:
 
 假如现在我们想修改预测任务下的统计模式的搜索空间, 即 ``StatsForecastSearchSpace``, 您可以做如下操作:
 
-- 详细指定任务类型 ``task``, 否则无法判断是单变量预测还是多变量预测任务;
-- 指定 ``timestamp`` 列名;
-- 如果数据存在协变量, 请设置参数 ``covariables={xxx: xxx, ...}``;
-- **以上三步请严格遵守, 否则自定义失败!**
 - 如果想禁止某个算法, 不进行搜索, 可以设置参数为False, 例如 ``enable_arima=False``;
 - 如果想更改某个算法的搜索空间参数初始化,可以传递参数 ``xxx_init_kwargs={xxx:xxx, ...}``;
 - 如果希望自定义的参数是可搜索的, 您可以使用 ``hypernets.core.search_space`` 中的 ``Choice``, ``Int`` 及 ``Real``。其中, ``Choice`` 支持离散值, ``Int`` 支持整数连续值, ``Real`` 支持浮点数连续值。详情可参考 `Search Space <https://github.com/DataCanvasIO/Hypernets/blob/master/hypernets/core/search_space.py>`_。
@@ -108,11 +104,8 @@ HyperTS针对不同的模式内置了丰富的建模算法, 例如:
     from hypernets.core.search_space import Choice, Int, Real
     from hyperts.framework.search_space.macro_search_space import StatsForecastSearchSpace
 
-    custom_search_space = StatsForecastSearchSpace(task='univariate-forecast', 
-                                                timestamp='TimeStamp',
-                                                covariables=['HourSin', 'WeekCos', 'CBWD'],
-                                                enable_arima=False,
-                                                prophet_init_kwargs={
+    custom_search_space = StatsForecastSearchSpace(enable_arima=False,
+                                                   prophet_init_kwargs={
                                                     'seasonality_mode': 'multiplicative',
                                                     'daily_seasonality': Choice([True, False]),
                                                     'n_changepoints': Int(10, 50, step=10),
@@ -378,10 +371,7 @@ HyperTS针对不同的模式内置了丰富的建模算法, 例如:
     df = load_network_traffic(univariate=True)
     train_data, test_data = train_test_split(df, test_size=168, shuffle=False)
 
-    custom_search_space = DLForecastSearchSpacePlusTransformer(task='univariate-forecast', 
-                                                            timestamp='TimeStamp',
-                                                            covariables=['HourSin', 'WeekCos', 'CBWD'],
-                                                            metrics=['mape'])
+    custom_search_space = DLForecastSearchSpacePlusTransformer()
 
     experiment = make_experiment(train_data, 
                                 task='univariate-forecast',

diff --git a/docs/zh_CN/source/contents/0700_models.rst b/docs/zh_CN/source/contents/0700_models.rst
@@ -134,10 +134,18 @@ LSTNet全称为长短时序网络(Long-and Short-term Time-series network, LSTNe
 详情可参看: `Modeling Long- and Short-Term Temporal Patterns with Deep Neural Networks <https://arxiv.org/abs/1703.07015>`_
 
 .. tip::
-    适用范围: 单/多变量时序预测,回归。
+    适用范围: 单/多变量时序预测, 回归。
 
 --------
 
 神经架构搜索
 *************
-...
+自AlexNet在2012年的ImageNet竞赛中获得冠军, 深度学习在许多具有挑战性的任务及领域取得了突破性的进展。除了AlexNet网络外, 
+比如VGG,Inception, ResNet, Transformer, GPT等也先后被提出并得到工业界及学术界广泛的应用。然后, 这众多优秀网络的背后,
+凝聚的是无数人类专家的经验结晶。因此, 神经体系结构搜索(NAS)已经成为一种有希望的工具，以减轻人类在这种试错设计过程中的努力。
+
+HyperTS依托Hypernets提供的基础能力(``Hpyer Model`` + ``Search Strategy`` + ``Estimation Strategy``), 构建了基于时间
+序列的 ``Search Space``, 为时间序列任务赋予NAS强大的表达能力。
+
+.. tip::
+    适用范围: 单/多变量时序预测, 回归。
diff --git a/docs/zh_CN/source/contents/0800_release_note.rst b/docs/zh_CN/source/contents/0800_release_note.rst
@@ -36,7 +36,7 @@ HyperTS是一款涵盖自动机器学习(AutoML)与自动深度学习(AutoDL)的
 版本 0.1.2.1
 **************
 
-HyperTS本次更新细节如下：
+HyperTS本次更新细节如下:
 
 - 新增交叉验证策略；
 
@@ -52,4 +52,40 @@ HyperTS本次更新细节如下：
 
 - 支持傅里叶时序周期推断；
 
-- 优化搜索空间及参数。
+- 优化搜索空间及参数。
+
+
+版本 0.1.3
+**************
+
+HyperTS本次更新细节如下:
+
+- 调整搜索空间超参数;
+
+- 增加实验执行结束报告最佳试验的参数;
+
+- 修复ARIMA以兼容0.12.1及以上版本;
+
+- 修复LSTNet的pt的问题;
+
+- 简化自定义搜索空间, 可不传参task, timestamp, covariables和metircs;
+
+- 新增 OutliersTransformer, 支持动态处理异常值;
+
+- 调整finaltrain - lr, batch_size, epcochs等;
+
+- 新增时间序列元特征提取器;
+
+- 新增Time2Vec、RevIN等时间处理层;
+
+- 新增N-Beats时序预测模型;
+
+- 新增InceptionTime时序分类模型;
+
+- 支持时序预测动态下采样;
+
+- 重构正标签推断方法;
+
+- 新增神经架构搜索模型;
+
+- 修复了一些已知问题。