Auto PyTorch下载 - Auto PyTorch源代码下载

Auto-Pytorch

版权（C）2021 Automl组Freiburg和Hannover

尽管早期的汽车框架着重于优化传统的ML管道及其超参数，但汽车的另一种趋势是专注于神经体系结构搜索。为了将这两个世界中的最好的世界融合在一起，我们开发了自动pytorch ，它们共同且可靠地优化了网络体系结构和培训超参数，以实现完全自动化的深度学习（AUTODL）。

Auto-Pytorch主要是为支持表格数据（分类，回归）和时间序列数据（预测）而开发的。自动图数据中的最新功能用于表格数据中的“自动图形表格：多效金属”，以提高效率且稳健的自动化量”（Bibtex Ref请参见下文）。有关多休性时间序列预测任务的Auto-Pytorch的详细信息，请参见“有效的自动化深度学习时间序列预测”（另请参见Bibtex Ref）。

另外，在此处找到文档。

从v0.1.0开始，使用SMAC作为基础优化软件包以及更改代码结构，已更新AutopyTorch，以进一步提高可用性，鲁棒性和效率。因此，从v0.0.2到v0.1.0将破坏兼容性。如果您想使用旧的API，可以在master_old上找到它。

工作流程

在下图中绘制了自动pytorch工作流程的粗略描述。

在该图中，数据由用户提供，投资组合是神经网络的一组配置，它们在不同的数据集上效果很好。当前版本仅支持纸张自动图表中所述的贪婪投资组合：多效金属学习效率和稳健的自动化量，该投资组合用于温暖SMAC的优化。换句话说，我们将提供的数据作为初始配置评估投资组合。然后API启动以下过程：

验证输入数据：处理每种数据类型，例如编码分类数据，以便可以处理自动数据。
创建数据集：创建一个可以在此API中处理的数据集，可以选择交叉验证或保持拆分。
评估基线
- 表格数据集*1：在预定池中使用固定的超参数配置和sklearn.dummy的虚拟模型训练每个算法，代表了最糟糕的性能。
- 时间序列预测数据集：训练一个虚拟预测指标，该预测指标重复每个系列中的最后观察值
SMAC搜索：
一个。确定超级带的预算和截止规则
b。采样管道超参数配置 *2由SMAC样品
c。通过获得的结果更新观察结果
d。重复a。 - c。直到预算用完为止
从合奏的观测值和模型选择中为提供的数据集建立最佳的合奏。

*1：基准是机器学习算法的预定池，例如LightGBM和支持向量机，用于在提供的数据集中求解回归或分类任务

*2：管道高参数配置指定了每个步骤中的组件的选择，例如目标算法，神经网络的形状，（其中指定每个步骤中的组件选择及其相应的超级标准器）。

安装

PYPI安装

pip install autoPyTorch

时间序列预测的自动播种需要其他依赖关系

pip install autoPyTorch[forecasting]

手动安装

我们建议使用Anaconda进行以下开发：

 # Following commands assume the user is in a cloned directory of Auto-Pytorch

# We also need to initialize the automl_common repository as follows
# You can find more information about this here:
# https://github.com/automl/automl_common/
git submodule update --init --recursive

# Create the environment
conda create -n auto-pytorch python=3.8
conda activate auto-pytorch
conda install swig
python setup.py install

同样，要安装所有依赖关系，以进行自动播放：

git submodule update --init --recursive

conda create -n auto-pytorch python=3.8
conda activate auto-pytorch
conda install swig
pip install -e[forecasting]

例子

简而言之：

 from autoPyTorch . api . tabular_classification import TabularClassificationTask

# data and metric imports
import sklearn . model_selection
import sklearn . datasets
import sklearn . metrics
X , y = sklearn . datasets . load_digits ( return_X_y = True )
X_train , X_test , y_train , y_test = 
        sklearn . model_selection . train_test_split ( X , y , random_state = 1 )

# initialise Auto-PyTorch api
api = TabularClassificationTask ()

# Search for an ensemble of machine learning algorithms
api . search (
    X_train = X_train ,
    y_train = y_train ,
    X_test = X_test ,
    y_test = y_test ,
    optimize_metric = 'accuracy' ,
    total_walltime_limit = 300 ,
    func_eval_time_limit_secs = 50
)

# Calculate test accuracy
y_pred = api . predict ( X_test )
score = api . score ( y_pred , y_test )
print ( "Accuracy score" , score )

时间序列预测任务

 from autoPyTorch . api . time_series_forecasting import TimeSeriesForecastingTask

# data and metric imports
from sktime . datasets import load_longley
targets , features = load_longley ()

# define the forecasting horizon
forecasting_horizon = 3

# Dataset optimized by APT-TS can be a list of np.ndarray/ pd.DataFrame where each series represents an element in the 
# list, or a single pd.DataFrame that records the series
# index information: to which series the timestep belongs? This id can be stored as the DataFrame's index or a separate
# column
# Within each series, we take the last forecasting_horizon as test targets. The items before that as training targets
# Normally the value to be forecasted should follow the training sets
y_train = [ targets [: - forecasting_horizon ]]
y_test = [ targets [ - forecasting_horizon :]]

# same for features. For uni-variant models, X_train, X_test can be omitted and set as None
X_train = [ features [: - forecasting_horizon ]]
# Here x_test indicates the 'known future features': they are the features known previously, features that are unknown
# could be replaced with NAN or zeros (which will not be used by our networks). If no feature is known beforehand,
# we could also omit X_test
known_future_features = list ( features . columns )
X_test = [ features [ - forecasting_horizon :]]

start_times = [ targets . index . to_timestamp ()[ 0 ]]
freq = '1Y'

# initialise Auto-PyTorch api
api = TimeSeriesForecastingTask ()

# Search for an ensemble of machine learning algorithms
api . search (
    X_train = X_train ,
    y_train = y_train ,
    X_test = X_test , 
    optimize_metric = 'mean_MAPE_forecasting' ,
    n_prediction_steps = forecasting_horizon ,
    memory_limit = 16 * 1024 ,  # Currently, forecasting models use much more memories
    freq = freq ,
    start_times = start_times ,
    func_eval_time_limit_secs = 50 ,
    total_walltime_limit = 60 ,
    min_num_test_instances = 1000 ,  # proxy validation sets. This only works for the tasks with more than 1000 series
    known_future_features = known_future_features ,
)

# our dataset could directly generate sequences for new datasets
test_sets = api . dataset . generate_test_seqs ()

# Calculate test accuracy
y_pred = api . predict ( test_sets )
score = api . score ( y_pred , y_test )
print ( "Forecasting score" , score )

有关更多示例，包括自定义搜索空间，分解代码等，请查看examples文件夹

$ cd examples/

该论文的代码可在TPAMI.2021.3067763分支中的examples/ensemble下找到。

贡献

如果您想为Auto-Pytorch做出贡献，请克隆存储库并结帐我们当前的开发分支机构

$ git checkout development

执照

该程序是免费的软件：您可以根据Apache许可证2.0的条款对其进行重新分配和/或对其进行修改（请参阅许可证文件）。

该程序的分布是希望它将有用的，但没有任何保修；即使没有对特定目的的适销性或适合性的隐含保证。

您应该已经收到了Apache许可证2.0的副本以及此程序（请参阅许可证文件）。

参考

请参阅分支TPAMI.2021.3067763以复制纸张自动式表格：多效金属学习，以提高自动级。

  @article { zimmer-tpami21a ,
  author = { Lucas Zimmer and Marius Lindauer and Frank Hutter } ,
  title = { Auto-PyTorch Tabular: Multi-Fidelity MetaLearning for Efficient and Robust AutoDL } ,
  journal = { IEEE Transactions on Pattern Analysis and Machine Intelligence } ,
  year = { 2021 } ,
  note = { also available under https://arxiv.org/abs/2006.13799 } ,
  pages = { 3079 - 3090 }
}

 @incollection { mendoza-automlbook18a ,
  author    = { Hector Mendoza and Aaron Klein and Matthias Feurer and Jost Tobias Springenberg and Matthias Urban and Michael Burkart and Max Dippel and Marius Lindauer and Frank Hutter } ,
  title     = { Towards Automatically-Tuned Deep Neural Networks } ,
  year      = { 2018 } ,
  month     = dec,
  editor    = { Hutter, Frank and Kotthoff, Lars and Vanschoren, Joaquin } ,
  booktitle = { AutoML: Methods, Sytems, Challenges } ,
  publisher = { Springer } ,
  chapter   = { 7 } ,
  pages     = { 141--156 }
}

 @article { deng-ecml22 ,
  author    = { Difan Deng and Florian Karl and Frank Hutter and Bernd Bischl and Marius Lindauer } ,
  title     = { Efficient Automated Deep Learning for Time Series Forecasting } ,
  year      = { 2022 } ,
  booktitle = { Machine Learning and Knowledge Discovery in Databases. Research Track
               - European Conference, {ECML} {PKDD} 2022 } ,
  url       = { https://doi.org/10.48550/arXiv.2205.05511 } ,
}