Concepts

Mlshell structure based on Pycnfg library. All parameters set from single configuration - Python dictionary. Execution could be controlled in a script/notebook either with pycnfg.run() , or low-level pycnfg.Handler . Sub-configurations in each sub-section describes initial state and steps to produce target object.

ML task configuration usually contains following main sections:

• path - set project path.

• logger - make logger.

• pipeline - create/load model.

• dataset - load from db.

• metric - specify scorers.

• workflow - fit/predict pipelines on datasets and optimize/validate metrics.

Configurations executied in priority:

project path >> logger >> pipeline, dataset, metric >> workflow

MLshell implements unified object interface for main sections and corresponding producer classes for object createion. All classes have high level of abstraction and encapsulation to provide extension with minimal code changes. See detailed description below.

See Examples.

Note

Arbitrary objects could be pre-accommodated in pycnfg.run(), so no need to always specify objects via configurations.

Each produce inherits from pycnfg.Producer , that includes load_cache/dump_cache methods to load/dump intermediate object state. The producers methods can be extended/rewrote either explicit in source or through patch key in configuration. In most cases default implementation will be sufficient.

Pycnfg configuration allows to set decorators for any step. It could be usefull for time/memory profiling and checking parameter consistence. For example built-in mlshell.decorator.checker() print:

• objects hash alteration.

• numpy errors.

Sections

Path

Set string path to project directory or pycnfg.find_path() could be used to auto-detect path.

Section example:

import pycnfg

'path': {
    'default': {
        'init': pycnfg.find_path,
        'producer': pycnfg.Producer,
        'priority': 1,
        'steps': [],
    }
    # Alternative.
    'my_path': {
        'init': '/home/user/project',
        'producer': pycnfg.Producer,
        'priority': 1,
        'steps': [],
    },
}

Logger

Set string name in init and use mlshell.LoggerProducer to create logger.

mlshell.LoggerProducer.make() makes logger via logging.config.dictConfig() . See default logger configuration mlshell.LOGGER_CONFIG for details.

Section example:

import pycnfg
import logging

'logger': {
    'default': {
        'init': 'default',
        'producer': mlshell.LoggerProducer,
        'priority': 2,
        'steps': [
            ('make',),
        ],
    }
    # Alternative.
    'my_logger': {
        'init': logging.getLogger('my_logger') ,
        'producer': pycnfg.Producer,
        'priority': 2,
        'steps': [],
    },
}

Pipeline

mlshell.Pipeline implements interface to access arbitrary pipeline. All sklearn based pipelines supported from the box. For others, adaptation in compliance with interface needed.

mlshell.PipelineProducer executes steps on pipeline.

• mlshell.PipelineProducer.load() loads existed model from IO.

• mlshell.PipelineProducer.info() logs model info.

• mlshell.PipelineProducer.make() creates pipeline from steps via sklearn.pipeline.Pipeline .

  • Any pypeline parameter(hp) can be tuned in GS.

  • Stable cross-validation scheme prevents data leaks.

By default unified steps embedded for broad range of ml tasks mlshell.pipeline.Steps :

default_steps = [
    ('pass_custom',      Pass custom params to scorer while grid search.)
    ('select_rows',      Delete rows (outliers/anomalies).)
    ('process_parallel',
        ('pipeline_categoric',
           ('select_columns',      Select categorical & binary sub-columns.)
           ('encode_onehot',       OneHot encoder.)
        )
        ('pipeline_numeric',
            ('select_columns',     Select numeric sub-columns.)
            ('impute',
                ('indicators',     Impute indicators.)
                ('gaps',           Impute gaps.)
            )
            ('transform_normal',   Yeo-Johnson features transformation.)
            ('scale_row_wise',     Row-wise transformation.)
            ('scale_column_wise',  Column-wise transformation.)
            ('add_polynomial',     Add polynomial features.)
            ('compose_columns',
                ("discretize",     Discretize columns.)
            )
        )
    )
    ('select_columns',   Model-wise feature selection.)
    ('reduce_dimension', Factor analyze feature selection/transformation.)
    ('estimate',         Target estimation.)
]

• For regressor estimate step supports target transformation:

sklearn.compose.TransformedTargetRegressor .

• For classifier estimate step supports two-stage meta-estimator to predict

  probabilities and apply classification threshold.

  • (‘predict_proba’, mlshell.model_selection.PredictionTransformer ),

  • (‘apply_threshold’, mlshell.model_selection.ThresholdClassifier ),

Section example:

import pycnfg
import lightgbm

'pipeline': {
    'my_pipe': {
        'init': mlshell.Pipeline,
        'producer': mlshell.PipelineProducer,
        'priority': 3,
        'steps': [
            ('make', {
                'estimator': lightgbm.LGBMRegressor()
                }),
        ],
    },
}

Dataset

mlshell.Dataset implements interface to access arbitrary dataset. Implemented as dictionary with some additional attributes and methods.

mlshell.DatasetProducer executes steps on dataset.

• mlshell.DatasetProducer.load() loads raw data from IO (csv imlemented).

• mlshell.DatasetProducer.info() logs dataset info.

• mlshell.DatasetProducer.split() splits dataset on test and train.

• mlshell.DatasetProducer.preprocess() preprocesses raw data to final state.

  • Categorical features are Ordinal encoded.

  • Categorical gaps are imputed with fillna(value=’unknown’).

  • Mumeric gaps are imputed with fillna(value= numpy.nan ).

  • Features casted to numpy.float64 .

Section example:

import pycnfg

'dataset': {
    'my_data': {
        'init': mlshell.Dataset,
        'producer': mlshell.DatasetProducer,
        'priority': 3,
        'steps': [
            ('load', {'filepath': 'data/train.csv',}),
            ('preprocess', {
                'targets_names': ['loss'],
                'categor_names': [f'cat{i}' for i in range(1, 117)],
                }),
            ('split', {'train_size': 0.7, 'shuffle': False}),
        ],
    },
}

Metric

mlshell.Metric implements interface to access arbitrary scorer. Extended scorer .

mlshell.MetricProducer executes steps on scorer.

mlshell.MetricProducer.make() make scorer from callable. Extended sklearn.metrics.make_scorer() .

Section example:

import pycnfg

'metric': {
    'my_score': {
        'init': mlshell.Metric,
        'producer': mlshell.MetricProducer,
        'priority': 3,
        'steps': [
            ('make', {
                'score_func': sklearn.metrics.r2_score,
                'greater_is_better': True
                }),
        ],
    }
}

Workflow

Initial section object set to dict .

mlshell.Workflow executes steps and writes results to dictionary.

• mlshell.Workflow.fit() fit pipeline.

• mlshell.Workflow.optimize() optimize pipeline.

• mlshell.Workflow.validate() make and score prediction.

• mlshell.Workflow.predict() make and dump prediction.

• mlshell.Workflow.plot() plot results.

• mlshell.Workflow.dump() dump pipeline.

Each method (except dump) includes subset argument to specify part of dataset to apply on. Typically:

• fit/optimize pipeline with CV on train/train subset.

• validate on postponed train subset.

• predict on test.

Section example:

import pycnfg

'workflow': {
    'my_flow': {
        'init': {},
        'producer': mlshell.Workflow,
        'priority': 4,
        'global': {
            'pipeline_id': 'pipeline__my_pipe',
            'dataset_id': 'dataset__my_data',
            'metric_id': ['metric__my_score'],
            'hp_grid': {
                'estimate__regressor__n_estimators':
                    np.linspace(50, 1000, 10, dtype=int),
            },
        },
        'steps': [
            ('optimize', ),
            ('validate', ),
            ('predict', ),
            ('dump', ),
        ],
    },
}

Hyperparameters

Any pipeline parameter hp can be optimized in grid search.

• mlshell.Workflow.fit() awaits for hp argument to update pipeline

  parameters before fitting. If parameters range provided, zero position will

  be used.
• mlshell.Workflow.optimize() awaits for hp_grid argument to set

  pipeline parameters range before optimizing.

hp_grid example for pipeline created from mlshell.pipeline.Steps :

hp_grid = {
    # custom metric(s) param
    'pass_custom__kw_args': [{'metric__id': {'param_a': 1, 'param_b': 'c'}},
                             {'metric__id': {'param_a': 2, 'param_b': 'd'}}],
    # transformers param
    'select_rows__kw_args': [{}],
    'process_parallel__pipeline_numeric__impute__gaps__strategy': ['constant'],
    'process_parallel__pipeline_numeric__transform_normal__skip': [True],
    'process_parallel__pipeline_numeric__scale_column_wise__quantile_range': [(1, 100)],
    'process_parallel__pipeline_numeric__add_polynomial__degree': [3],
    'process_parallel__pipeline_numeric__compose_columns__discretize__n_bins': [5],
    'select_columns__estimator__skip': [True],
    'reduce_dimension__skip': [True],
    # regressor only
    'estimate__transformer': [None, sklearn.preprocessing.FunctionTransformer(func=np.log, inverse_func=np.exp)],
    # estimator params
    'estimate__regressor__n_estimators': np.linspace(50, 500, 10, dtype=int),
    'estimate__regressor__num_leaves' :[2 ** i for i in range(1, 6 + 1)],
    'estimate__regressor__max_depth': np.linspace(1, 10, 10, dtype=int),
    # classifier only
    'estimate__apply_threshold__threshold': [0.1],
    'estimate__predict_proba__classifier__min_child_samples': np.linspace(1, 100, 10, dtype=int),
}

Note

Probaility distribution for hp_grid params are also possible, it should supports .rvs() sampling method.

Resolver

Some parameters like categoric/numeric indices depend on specific dataset, they need resolving before pipeline fitting. For that case, ‘auto’/[‘auto’] can be set in parameters (hp or hp_grid), that will trigger hp value resolution according to mlshell.model_selection.Resolver . If [‘auto’], parameter will be substituted with [‘resolved value’].

Note

Default resolver supports classification threshold range resolution from ROC curve on OOF prediction mlshell.model_selection.Resolver.th_resolver() . See also Classification threshold below.

Optimizer

mlshell.model_selection.Optimizer proposes unified interface to use arbitrary optimizers. Intended to be used in . For new optimizer formats no need to edit mlshell.Workflow class, just adapt Optimizer in compliance to interface.

mlshell.model_selection.RandomizedSearchOptimizer contains implementation for sklearn.model_selection.RandomizedSearchCV .

mlshell.model_selection.MockOptimizer contains implementation to efficient brute force prediction-related parameters as separate optimize step. For example: classification threshold or score function kwargs don`t need whole pipeline refit to probe.

mlshell.Workflow supports multi-stage optimization, for any pipeline-dataset pair hyper-parameters could be tuned step-wise with different optimizers. Pipeline in objects auto updated to the best estimator find before for the current dataset. See mlshell.model_selection.Optimizer.update_best() for stage merging logic.

scorer kwargs

Pipeline step pass_custom__kw_args allows to pass arbitrary parameters to scorer function while grid search (as if there was additional nested loop). Scorer should be made with needs_custom_kw_args=True flag and supports corresponding kwargs. Efficient mlshell.model_selection.MockOptimizer could be used.

def custom_score_metric(y_true, y_pred, **kwargs):
    """Custom precision metric."""
    if kwargs:
        # some logic depends on kwargs.
        pass
    tp = np.count_nonzero((y_true == 1) & (y_pred == 1))
    fp = np.count_nonzero((y_true == 0) & (y_pred == 1))
    score = tp/(fp+tp) if tp+fp != 0 else 0
return score

classifier threshold

For classification task it is possible to tune classification threshold th_ on CV. If positive class probability P(positive label) = 1 - sum(P(negative label)) > th_ for some sample, classifier set positive label for this sample, otherwise negative label with max probability.

Set last pipeline step to:

step = sklearn.pipeline.Pipeline([
    ('estimate', sklearn.pipeline.Pipeline([
        ('predict_proba', :class:`mlshell.model_selection.PredictionTransformer` (estimator)),
        ('apply_threshold', :class:`mlshell.model_selection.ThresholdClassifier` (threshold)),
    ]))
])

Note

Mlshell supports multiple strategy for th_ tuning:

0. Not use th_.

   • Not all classifiers provide predict_proba (SVM).

   • Use f1, logloss metrics.

1. One-stage optimization.

   • Tune th_ on a par with other hps.

2. Multi-stage optimization.

   First optimizer hps, then th_ separately.

   • For hps stage use for example auc-roc score.

   • For th_ stage use original score function.

   Efficient mlshell.model_selection.MockOptimizer could be used.

For (1)/(2) th_ search range should be known in advance:

• Either set arbitrary range in hp_grid.

{‘estimate__apply_threshold__threshold’: [val,] }

• Or set ‘auto’/[‘auto’] in hp_grid to resolve typical values from ROC curve on OOF probabilities predictions.

Note, will be used default hps in case (1), best on previous stage hps in case (2). {‘estimate__apply_threshold__threshold’: ‘auto’}

th_ range extracted from ROC curve example:

Validator

mlshell.model_selection.Validator proposes unified interface to validate pipeline on scorers. Intended to be used in mlshell.Workflow. For new pipeline/metric interface, just adapt ‘Validator’ in compliance to interface, so no need to edit Workflow class

Plotter

mlshell.plot.Plotter proposes unified interface to plot results.

gui

Mlshell provides gui based on matplotlib.widgets .

For small dataset it is reasonable to visualize vectorized score per samples. Sliders for hyper-parameters are available. Model retrained and make predict on slider change.

Note

Scorer score_func_vector should be available to plot vectorized score.

Results

pycnfg.run() returns objects dictionary, containing resulted objects for all executed configurations. Each pipeline is fitted corresponding to last fit/optimize method.

Each optimization stage produces <timestamp>_runs.csv file. See mlshell.model_selection.Optimizer.dump_runs() for details.

*_runs.csv files could be merged in dataframe for further analyse:

import glob

files = sorted(glob.glob(f"project/results/runs/*_runs.csv"))
df_lis = list(range(len(files)))
for i, f in enumerate(files):
    try:
        df_lis[i] = pd.read_csv(f, sep=",", header=0)
        print('Read runs.csv\n', f, df_lis[i].shape,
              df_lis[i]['dataset__id'][0], df_lis[i]['pipeline__id'][0])
    except Exception as e:
        print(e)
        continue
df = pd.concat(df_lis, axis=0, sort=False).reset_index()

# groupby data hash
df.groupby('dataset__hash').size()
# groupby estimator
df.groupby('pipeline__hash').size()

Project structure

It is convenient to follow the structure:

|_project/
    ** input **
    |__ conf.py
    |__ run.py
    |__ EDA.ipynb
    |__ data/
        ~~ could be remote db ~~
        |__ train.csv
        |__ test.csv
    ** output autogenerated **
    |__ results/
        |__ models/
            ~~ dump fitted models and predictions ~~
            |__ <id>.model
            |__ <id>_pred.csv
        |__ runs/
            ~~ dump all GS runs result ~~
            |__ <timestamp>_runs.csv
        |__ <logger_name>_logs/
            ~~ script logs ~~
            |__ <logger_name>_<logger_level>.log
        |__ ipython_logs/
            ~~ notebook logs ~~
            |__ <logger_name>_<logger_level>.log
        |__ .temp/
            ~~ cache ~~
            |__ objects/
                ~~ cache for producer objects ~~
            |__ pipeline/
                ~~ cache for sklearn.pipeline.Pipeline(memory=<./.temp/pipeline>) ~~