Neuraxle AutoML - 为啥会出错?
Posted
技术标签:
【中文标题】Neuraxle AutoML - 为啥会出错?【英文标题】:Neuraxle AutoML - Why is it erroring?Neuraxle AutoML - 为什么会出错? 【发布时间】:2021-05-08 03:16:20 【问题描述】:我正在按照 AutoML 示例对 Neuraxle 进行试验。
未修改的示例工作正常。
当我修改它以在ChooseOneStepOf(classifiers)
之前包含我自己的管道组件时,它失败了,我不明白为什么。
from neuraxle.base import BaseTransformer
from neuraxle.pipeline import Pipeline
from neuraxle.hyperparams.space import HyperparameterSpace
from neuraxle.steps.numpy import NumpyRavel
from neuraxle.steps.output_handlers import OutputTransformerWrapper
from typing import List
from sklearn.preprocessing import OneHotEncoder
from neuraxle.pipeline import Pipeline
from neuraxle.union import FeatureUnion
from sklearn.impute import SimpleImputer
# sklearn classifiers, and sklearn wrapper for neuraxle
from neuraxle.steps.sklearn import SKLearnWrapper
from sklearn.tree import DecisionTreeClassifier, ExtraTreeClassifier
from sklearn.linear_model import RidgeClassifier, LogisticRegression
# neuraxle distributions
from neuraxle.hyperparams.distributions import Choice, RandInt, Boolean, LogUniform
from neuraxle.steps.flow import ChooseOneStepOf
from neuraxle.base import BaseTransformer, ForceHandleMixin
from neuraxle.metaopt.auto_ml import ValidationSplitter
from neuraxle.metaopt.callbacks import ScoringCallback
from sklearn.metrics import accuracy_score
from neuraxle.metaopt.callbacks import MetricCallback
from sklearn.metrics import f1_score, precision_score, recall_score
from neuraxle.metaopt.auto_ml import InMemoryHyperparamsRepository
from neuraxle.plotting import TrialMetricsPlottingObserver
from neuraxle.metaopt.tpe import TreeParzenEstimatorHyperparameterSelectionStrategy
from sklearn.model_selection import train_test_split
import numpy as np
import pandas as pd
from neuraxle.metaopt.auto_ml import AutoML
import os
classifiers: List[BaseTransformer] = [
SKLearnWrapper(DecisionTreeClassifier(), HyperparameterSpace(
'criterion': Choice(['gini', 'entropy']),
'splitter': Choice(['best', 'random']),
'min_samples_leaf': RandInt(2, 5),
'min_samples_split': RandInt(1, 3)
)).set_name('DecisionTreeClassifier'),
Pipeline([
OutputTransformerWrapper(NumpyRavel()),
SKLearnWrapper(RidgeClassifier(), HyperparameterSpace(
'alpha': Choice([(0.0, 1.0, 10.0), (0.0, 10.0, 100.0)]),
'fit_intercept': Boolean(),
'normalize': Boolean()
))
]).set_name('RidgeClassifier'),
Pipeline([
OutputTransformerWrapper(NumpyRavel()),
SKLearnWrapper(LogisticRegression(), HyperparameterSpace(
'C': LogUniform(0.01, 10.0),
'fit_intercept': Boolean(),
'dual': Boolean(),
'penalty': Choice(['l1', 'l2']),
'max_iter': RandInt(20, 200)
))
]).set_name('LogisticRegression')
]
class ColumnSelectTransformer(BaseTransformer, ForceHandleMixin):
def __init__(self, required_columns):
BaseTransformer.__init__(self)
ForceHandleMixin.__init__(self)
self.required_columns = required_columns
def inverse_transform(self, processed_outputs):
pass
def fit(self, X, y=None):
return self
def transform(self, X):
if not isinstance(X, pd.DataFrame):
X = pd.DataFrame(X)
return X[self.required_columns]
columns = ['BEDCERT', 'RESTOT', 'INHOSP', 'CCRC_FACIL',
'SFF', 'CHOW_LAST_12MOS', 'SPRINKLER_STATUS',
'EXP_TOTAL', 'ADJ_TOTAL']
simple_features = Pipeline([ColumnSelectTransformer(columns),
SimpleImputer(missing_values=np.nan,
strategy='mean')])
categorical_features = Pipeline([ColumnSelectTransformer(['OWNERSHIP', 'CERTIFICATION']),
OneHotEncoder(sparse=False)
])
business_features = FeatureUnion([simple_features,
categorical_features])
p: Pipeline = Pipeline([
business_features,
ChooseOneStepOf(classifiers)
])
validation_splitter = ValidationSplitter(test_size=0.20)
scoring_callback = ScoringCallback(
metric_function=accuracy_score,
name='accuracy',
higher_score_is_better=False,
print_metrics=False
)
callbacks = [
MetricCallback('f1', metric_function=f1_score, higher_score_is_better=True, print_metrics=False),
MetricCallback('precision', metric_function=precision_score, higher_score_is_better=True, print_metrics=False),
MetricCallback('recall', metric_function=recall_score, higher_score_is_better=True, print_metrics=False)
]
hyperparams_repository = InMemoryHyperparamsRepository(cache_folder='cache')
hyperparams_repository.subscribe(TrialMetricsPlottingObserver(
plotting_folder_name='metric_results',
save_plots=False,
plot_trial_on_next=False,
plot_all_trials_on_complete=True,
plot_individual_trials_on_complete=False
))
hyperparams_optimizer = TreeParzenEstimatorHyperparameterSelectionStrategy(
number_of_initial_random_step=10,
quantile_threshold=0.3,
number_good_trials_max_cap=25,
number_possible_hyperparams_candidates=100,
prior_weight=0.,
use_linear_forgetting_weights=False,
number_recent_trial_at_full_weights=25
)
tmpdir = 'cache'
if not os.path.exists(tmpdir):
os.makedirs(tmpdir)
n_trials = 10
n_epochs = 10
auto_ml = AutoML(
pipeline=p,
validation_splitter=validation_splitter,
refit_trial=True,
n_trials=n_trials,
epochs=n_epochs,
cache_folder_when_no_handle=str(tmpdir),
scoring_callback=scoring_callback,
callbacks=callbacks,
hyperparams_repository=hyperparams_repository
)
def generate_classification_data():
# data_inputs, expected_outputs = make_classification(
# n_samples=10000,
# n_repeated=0,
# n_classes=3,
# n_features=4,
# n_clusters_per_class=1,
# class_sep=1.5,
# flip_y=0,
# weights=[0.5, 0.5, 0.5]
# )
data = pd.read_csv('./ml-data/providers-train.csv', encoding='latin1')
fine_counts = data.pop('FINE_CNT')
fine_totals = data.pop('FINE_TOT')
cycle_2_score = data.pop('CYCLE_2_TOTAL_SCORE')
X_train, X_test, y_train, y_test = train_test_split(
data,
fine_counts > 1,
test_size=0.20
)
return X_train, y_train, X_test, y_test
X_train, y_train, X_test, y_test = generate_classification_data()
auto_ml = auto_ml.fit(X_train, y_train)
Output as follows:-
/Users/simon/venvs/wqu_q4/bin/python /Users/simon/Dev/wqu_q4/main.py 新试验:“ChooseOneStepOf”:“choice”:“RidgeClassifier”
trial 1/10 Traceback(大多数 最近通话最后):文件 “/Users/simon/venvs/wqu_q4/lib/python3.9/site-packages/neuraxle/metaopt/auto_ml.py”, 第 785 行,在 _fit_data_container repo_trial_split = self.trainer.execute_trial( 文件 “/Users/simon/venvs/wqu_q4/lib/python3.9/site-packages/neuraxle/metaopt/trial.py”, 第 290 行,在 exit 提出 exc_val 文件 “/Users/simon/venvs/wqu_q4/lib/python3.9/site-packages/neuraxle/metaopt/auto_ml.py”, 第 785 行,在 _fit_data_container repo_trial_split = self.trainer.execute_trial( 文件 “/Users/simon/venvs/wqu_q4/lib/python3.9/site-packages/neuraxle/metaopt/auto_ml.py”, 第 595 行,在 execute_trial self.print_func('success trial score: '.format( 文件 “/Users/simon/venvs/wqu_q4/lib/python3.9/site-packages/neuraxle/metaopt/trial.py”, 第 570 行,在 exit 中提出 exc_val 文件 “/Users/simon/venvs/wqu_q4/lib/python3.9/site-packages/neuraxle/metaopt/auto_ml.py”, 第 574 行,在 execute_trial trial_split_description = _get_trial_split_description(文件“/Users/simon/venvs/wqu_q4/lib/python3.9/site-packages/neuraxle/metaopt/auto_ml.py”, 第 876 行,在 _get_trial_split_description json.dumps(repo_trial.hyperparams, sort_keys=True, indent=4) 文件 "/usr/local/Cellar/python@3.9/3.9.0_1/Frameworks/Python.framework/Versions/3.9/lib/python3.9/json/init.py", 第 234 行,在转储中返回 cls(文件 "/usr/local/Cellar/python@3.9/3.9.0_1/Frameworks/Python.framework/Versions/3.9/lib/python3.9/json/encoder.py", 第 201 行,在编码块 = 列表(块)文件中 "/usr/local/Cellar/python@3.9/3.9.0_1/Frameworks/Python.framework/Versions/3.9/lib/python3.9/json/encoder.py", 第 431 行,在 _iterencode 中的 _iterencode_dict(o, _current_indent_level)文件“/usr/local/Cellar/python@3.9/3.9.0_1/Frameworks/Python.framework/Versions/3.9/lib/python3.9/json/encoder.py”, 第 405 行,在块文件中的 _iterencode_dict 产量 "/usr/local/Cellar/python@3.9/3.9.0_1/Frameworks/Python.framework/Versions/3.9/lib/python3.9/json/encoder.py", 第 438 行,在 _iterencode o = _default(o) 文件中 "/usr/local/Cellar/python@3.9/3.9.0_1/Frameworks/Python.framework/Versions/3.9/lib/python3.9/json/encoder.py", 第 179 行,默认提高 TypeError(f'Object of type o.class.name ' TypeError: 类型类型的对象不是 JSON serializable 在处理上述异常的过程中,另一个异常 发生:回溯(最近一次通话最后一次):文件 “/Users/simon/Dev/wqu_q4/main.py”,第 210 行,在 auto_ml = auto_ml.fit(X_train, y_train) 文件 “/Users/simon/venvs/wqu_q4/lib/python3.9/site-packages/neuraxle/base.py”, 第 3475 行,适合 new_self = self.handle_fit(data_container, context) 文件 “/Users/simon/venvs/wqu_q4/lib/python3.9/site-packages/neuraxle/base.py”, 第 980 行,在 handle_fit new_self = self._fit_data_container(data_container, context) 文件 “/Users/simon/venvs/wqu_q4/lib/python3.9/site-packages/neuraxle/metaopt/auto_ml.py”, 第 802 行,在 _fit_data_container repo_trial_split=repo_trial_split, UnboundLocalError:之前引用的局部变量“repo_trial_split” 分配过程以退出代码 1 结束
【问题讨论】:
Neuraxle 的版本是多少?您可以尝试更新它吗?我看不出这会失败的原因。似乎它试图使用一个不存在的变量,这很奇怪。也许重置你的 .pyc 预编译文件,或者尝试重新安装你的 venv ? 0.5.6 在 pycharm 中,这很奇怪,因为 Github 表明您只有 0.5.5 我想知道它是否与我的 ColumnTransformer 没有以正确的格式返回数据是否需要在一个 numpy 数组中或者我可以返回一个熊猫数据框吗?我只是为此目的创建了虚拟环境,但可以重试。 【参考方案1】:一些可以帮助您解决当前问题的注意事项:
“UnboundLocalError:分配前引用的局部变量 'repo_trial_split'”是在 AutoML 循环中的流水线内发生崩溃时发生的错误。您应该将真正的错误记录在您在此处发布的错误之上。此外,Neuraxle 版本 0.5.7(尚未发布,但在 github 上可用)通过添加一个名为“continue_loop_on_error”的参数来解决此问题,您应该将其设置为 False。
您似乎在为您的 ColumnSelectTransformer 实例使用 ForceHandleMixin。使用 ForceHandleMixin 意味着您应该定义以下函数 _fit_data_container、_transform_data_container 和 _fit_transform_data_container 而不是 fit/fit_transform/transform。
您可能需要编写一个 Neuraxle 类来包装 scikit 的 SimpleImputer。
希望这对您有所帮助。完成这些更改后,请随时在此处发布更新,我很乐意为您提供帮助。您也可以在 Neuraxle 的 Slack 上发帖,我可能会在那里更快地回答。
干杯!
附言另一方面,我将在接下来的几天内发布 0.5.7 版本。
【讨论】:
感谢您的回复,我会进一步调查以上是关于Neuraxle AutoML - 为啥会出错?的主要内容,如果未能解决你的问题,请参考以下文章
在 Cloud AutoML Vision 中将图像导入 Google 存储时出错
如何最好地处理 Neuraxle 管道中的错误和/或丢失数据?
Pandas DataFrame 中的 Neuraxle 选择列