Pipeline

개요

• Scikit-Learn의 Pipeline은 강력하다.
• PyCaret, Skorch에도 사용이 가능하다.
• Google Colab에서 시도해보자.

필수 라이브러리 설치

• pycaret을 설치 한 후에는 반드시 런타임 재시작을 클릭한다.

!pip install pycaret

Collecting pycaret
  Downloading pycaret-2.3.5-py3-none-any.whl (288 kB)
.
.
Successfully installed Boruta-0.3 Mako-1.1.6 PyYAML-6.0 alembic-1.4.1 databricks-cli-0.16.2 docker-5.0.3 funcy-1.17 gitdb-4.0.9 gitpython-3.1.24 gunicorn-20.1.0 htmlmin-0.1.12 imagehash-4.2.1 imbalanced-learn-0.7.0 joblib-1.0.1 kmodes-0.11.1 lightgbm-3.3.1 mlflow-1.22.0 mlxtend-0.19.0 multimethod-1.6 pandas-profiling-3.1.0 phik-0.12.0 prometheus-flask-exporter-0.18.7 pyLDAvis-3.2.2 pycaret-2.3.5 pydantic-1.8.2 pynndescent-0.5.5 pyod-0.9.6 python-editor-1.0.4 querystring-parser-1.2.4 requests-2.26.0 scikit-learn-0.23.2 scikit-plot-0.3.7 scipy-1.5.4 smmap-5.0.0 tangled-up-in-unicode-0.1.0 umap-learn-0.5.2 visions-0.7.4 websocket-client-1.2.3

!pip install -U skorch

Requirement already satisfied: skorch in /usr/local/lib/python3.7/dist-packages (0.11.0)
Requirement already satisfied: tabulate>=0.7.7 in /usr/local/lib/python3.7/dist-packages (from skorch) (0.8.9)
Requirement already satisfied: scikit-learn>=0.19.1 in /usr/local/lib/python3.7/dist-packages (from skorch) (0.23.2)
Requirement already satisfied: tqdm>=4.14.0 in /usr/local/lib/python3.7/dist-packages (from skorch) (4.62.3)
Requirement already satisfied: numpy>=1.13.3 in /usr/local/lib/python3.7/dist-packages (from skorch) (1.19.5)
Requirement already satisfied: scipy>=1.1.0 in /usr/local/lib/python3.7/dist-packages (from skorch) (1.5.4)
Requirement already satisfied: joblib>=0.11 in /usr/local/lib/python3.7/dist-packages (from scikit-learn>=0.19.1->skorch) (1.0.1)
Requirement already satisfied: threadpoolctl>=2.0.0 in /usr/local/lib/python3.7/dist-packages (from scikit-learn>=0.19.1->skorch) (3.0.0)

from pycaret.datasets import get_data
data = get_data("electrical_grid")

PyTorchModel

• sktorch 라이브러리는 PyTorch 모델과 함께 작동한다.
• MLP 모델을 작성하는 클래스를 설계한다.

import torch.nn as nn

class Net(nn.Module): 
  def __init__(self, num_inputs=12, num_units_d1 = 200, num_units_d2 = 100):
    super(Net, self).__init__() 

    self.dense0 = nn.Linear(num_inputs, num_units_d1)
    self.nonlin = nn.ReLU()
    self.dropout = nn.Dropout(0.5)
    self.dense1 = nn.Linear(num_units_d1, num_units_d2)
    self.output = nn.Linear(num_units_d2, 2)
    self.softmax = nn.Softmax(dim=-1)

  def forward(self, X, **kwargs):
    X = self.nonlin(self.dense0(X))
    X = self.dropout(X)
    X = self.nonlin(self.dense1(X))
    X = self.softmax(self.output(X))
    return X

Skorch Classifier

• NeuralNetClassifier 클래스를 PyTorch 클래스와 연동한다.
• Optimizer 기본값인 SGD를 사용한다. 만약 다른 Optimizer로 변경을 원하면 다음 링크에서 확인한다.
  • 참고: https://skorch.readthedocs.io/en/latest/user/neuralnet.html#optimizer
• Sktorch 5 폴드 교차검증을 수행한다.
  • 학습 데이터는 80%, 나머지 20%는 검증 데이터로 활용한다.

from skorch import NeuralNetClassifier 

net = NeuralNetClassifier(
    module = Net, 
    max_epochs = 30, 
    lr = 0.1, 
    batch_size = 32, 
    train_split = None
)

PyCaret과 신경망 학습 방법

• SKORCH NN model을 초기화 했다면, 이번에는 PyCaret과 함께 모델을 학습할 수 있다.
• PyCaret은 기본적으로 Pandas DataFrame을 메인 객체로 사용하다.
• 그런데, sktorch model을 사용하기 위해서는 pipeline을 구성할 때는 DataFrameTransformer() 함수를 사용해야 한다.

from skorch.helper import DataFrameTransformer
import numpy as np
from sklearn.pipeline import Pipeline

nn_pipe = Pipeline(
    [("transform", DataFrameTransformer()), 
     ("net", net), ]
)

PyCaret Setup

• Skorch API 대신 PyCaret 모델을 사용해본다.
• log_experiment가 True를 사용하게 되면 MLFlow를 사용할 수 있다.
• silent가 True인 경우 중간에 발생하는 press enter to continue 입력 단계를 피할 수 있다.

from pycaret.classification import *
target = "stabf"
clf1 = setup(data = data, 
            target = target,
            train_size = 0.8,
            fold = 5,
            session_id = 123,
            log_experiment = True, 
            experiment_name = 'electrical_grid_1', 
            silent = True)

PyCaret Train Model

• Random Forest 모델을 사용해본다.

model = create_model("rf")

PyCaret Train Skorch Model

• 이번에는 Skorch Model을 Pycaret 함수에 넣어서 확인해본다.

skorch_model = create_model(nn_pipe)

Comparing Models

• 두 모델 중 어떤 모델이 더 좋은지 확인해본다.

best_model = compare_models(include=[skorch_model, model], sort = "AUC")

Hyperparameter Grid

• Hyperparameter 튜닝을 적용하도록 한다.
• 모형 튜닝을 위한 parameter 값은 다음 명령어를 통해서 확인할 수 있다.

skorch_model.get_params().keys()

dict_keys(['memory', 'steps', 'verbose', 'transform', 'net', 'transform__float_dtype', 'transform__int_dtype', 'transform__treat_int_as_categorical', 'net__module', 'net__criterion', 'net__optimizer', 'net__lr', 'net__max_epochs', 'net__batch_size', 'net__iterator_train', 'net__iterator_valid', 'net__dataset', 'net__train_split', 'net__callbacks', 'net__predict_nonlinearity', 'net__warm_start', 'net__verbose', 'net__device', 'net___kwargs', 'net__classes', 'net__callbacks__epoch_timer', 'net__callbacks__train_loss', 'net__callbacks__train_loss__name', 'net__callbacks__train_loss__lower_is_better', 'net__callbacks__train_loss__on_train', 'net__callbacks__valid_loss', 'net__callbacks__valid_loss__name', 'net__callbacks__valid_loss__lower_is_better', 'net__callbacks__valid_loss__on_train', 'net__callbacks__valid_acc', 'net__callbacks__valid_acc__scoring', 'net__callbacks__valid_acc__lower_is_better', 'net__callbacks__valid_acc__on_train', 'net__callbacks__valid_acc__name', 'net__callbacks__valid_acc__target_extractor', 'net__callbacks__valid_acc__use_caching', 'net__callbacks__print_log', 'net__callbacks__print_log__keys_ignored', 'net__callbacks__print_log__sink', 'net__callbacks__print_log__tablefmt', 'net__callbacks__print_log__floatfmt', 'net__callbacks__print_log__stralign'])

net.get_params().keys()

dict_keys(['module', 'criterion', 'optimizer', 'lr', 'max_epochs', 'batch_size', 'iterator_train', 'iterator_valid', 'dataset', 'train_split', 'callbacks', 'predict_nonlinearity', 'warm_start', 'verbose', 'device', '_kwargs', 'classes', 'callbacks__epoch_timer', 'callbacks__train_loss', 'callbacks__train_loss__name', 'callbacks__train_loss__lower_is_better', 'callbacks__train_loss__on_train', 'callbacks__valid_loss', 'callbacks__valid_loss__name', 'callbacks__valid_loss__lower_is_better', 'callbacks__valid_loss__on_train', 'callbacks__valid_acc', 'callbacks__valid_acc__scoring', 'callbacks__valid_acc__lower_is_better', 'callbacks__valid_acc__on_train', 'callbacks__valid_acc__name', 'callbacks__valid_acc__target_extractor', 'callbacks__valid_acc__use_caching', 'callbacks__print_log', 'callbacks__print_log__keys_ignored', 'callbacks__print_log__sink', 'callbacks__print_log__tablefmt', 'callbacks__print_log__floatfmt', 'callbacks__print_log__stralign'])

import torch.optim as optim

custom_grid = {
	'net__max_epochs':[20, 30],
	'net__lr': [0.01, 0.05, 0.1],
	'net__module__num_units_d1': [50, 100, 150],
	'net__module__num_units_d2': [50, 100, 150],
	'net__optimizer': [optim.Adam, optim.SGD, optim.RMSprop]
	}

• 이번에는 hyperparameter 모델을 적용하여 모델을 빠르게 만들어 본다.

tuned_skorch_model = tune_model(skorch_model, custom_grid = custom_grid)

References

1. https://pycaret.org/
2. https://www.analyticsvidhya.com/blog/2020/05/pycaret-machine-learning-model-seconds/
3. https://github.com/skorch-dev/skorch
4. https://towardsdatascience.com/skorch-pytorch-models-trained-with-a-scikit-learn-wrapper-62b9a154623e
5. https://towardsdatascience.com/pycaret-skorch-build-pytorch-neural-networks-using-minimal-code-57079e197f33