What is Pyemits

PyEmits, a python package for easy manipulation in time-series data.

The ultimate goal:

Keep it simple and stupid

Make everything configurable

Uniform API for machine learning and deep learning

Why need Pyemits?

Time-series data is very common in real life.

• Engineering
• FSI industry (Financial Services Industry)
• FMCG (Fast Moving Consumer Good)

Data scientist's work consists of:

• forecasting
• prediction/simulation
• data preparation
• cleansing
• anomaly detection
• descriptive data analysis/exploratory data analysis/data profile
• data processing and ETL pipeline scripts

each new business unit shall build the following wheels again and again

if you are facing these problems, then Pyemits is fit for you

1. data processing and ETL pipeline
   1. extraction
   2. transformation
      1. cleansing
      2. feature engineering
      3. remove outliers
      4. AI landing for prediction, forecasting
   3. write it back to database
2. ml framework
   1. multiple model training
   2. multiple model prediction
   3. kfold validation
   4. anomaly detection
   5. forecasting
   6. develop deep learning model (regression)
   7. ensemble modelling
3. exploratory data analysis
   1. descriptive data analysis
   2. data profile
   3. data set comparison

data scientist need to write different code to develop their model is there a package integrate all ml lib with single simple api? That's why I create this project.

This project is under active development, free to use (Apache 2.0) I am happy to see anyone can contribute for more advancement on features

New feature:

data processing pipeline

db connection and manipulation

Development Progress

Release Update

Install

pip install pyemits

Features highlight

scikit-learn API style

inherit the design concept of pyecharts, "everything is configurable"

highly flexible configuration items, can easily integrate with existing model

easily integrate to SaaS product for product proof of concept

Easy training

import numpy as np

from pyemits.core.ml.regression.trainer import RegTrainer, RegressionDataModel

X = np.random.randint(1, 100, size=(1000, 10))
y = np.random.randint(1, 100, size=(1000, 1))

raw_data_model = RegressionDataModel(X, y)
trainer = RegTrainer(['XGBoost'], [None], raw_data_model)
trainer.fit()

Accept neural network as model

import numpy as np

from pyemits.core.ml.regression.trainer import RegTrainer, RegressionDataModel
from pyemits.core.ml.regression.nn import KerasWrapper

X = np.random.randint(1, 100, size=(1000, 10, 10))
y = np.random.randint(1, 100, size=(1000, 4))

keras_lstm_model = KerasWrapper.from_simple_lstm_model((10, 10), 4)
raw_data_model = RegressionDataModel(X, y)
trainer = RegTrainer([keras_lstm_model], [None], raw_data_model)
trainer.fit()

also keep flexibility on customized model

import numpy as np

from pyemits.core.ml.regression.trainer import RegTrainer, RegressionDataModel
from pyemits.core.ml.regression.nn import KerasWrapper

X = np.random.randint(1, 100, size=(1000, 10, 10))
y = np.random.randint(1, 100, size=(1000, 4))

from keras.layers import Dense, Dropout, LSTM
from keras import Sequential

model = Sequential()
model.add(LSTM(128,
               activation='softmax',
               input_shape=(10, 10),
               ))
model.add(Dropout(0.1))
model.add(Dense(4))
model.compile(loss='mse', optimizer='adam', metrics=['mse'])

keras_lstm_model = KerasWrapper(model, nickname='LSTM')
raw_data_model = RegressionDataModel(X, y)
trainer = RegTrainer([keras_lstm_model], [None], raw_data_model)
trainer.fit()

or attach it in algo config

import numpy as np

from pyemits.core.ml.regression.trainer import RegTrainer, RegressionDataModel
from pyemits.core.ml.regression.nn import KerasWrapper
from pyemits.common.config_model import KerasSequentialConfig

X = np.random.randint(1, 100, size=(1000, 10, 10))
y = np.random.randint(1, 100, size=(1000, 4))

from keras.layers import Dense, Dropout, LSTM
from keras import Sequential

keras_lstm_model = KerasWrapper(nickname='LSTM')
config = KerasSequentialConfig(layer=[LSTM(128,
                                           activation='softmax',
                                           input_shape=(10, 10),
                                           ),
                                      Dropout(0.1),
                                      Dense(4)],
                               compile=dict(loss='mse', optimizer='adam', metrics=['mse']))

raw_data_model = RegressionDataModel(X, y)
trainer = RegTrainer([keras_lstm_model],
                     [config],
                     raw_data_model,
                     {'fit_config': [dict(epochs=10, batch_size=32)]})
trainer.fit()

PyTorch, MXNet under development you can leave me a message if you want to contribute

MultiOutput training

import numpy as np

from pyemits.core.ml.regression.trainer import RegressionDataModel, MultiOutputRegTrainer
from pyemits.core.preprocessing.splitting import SlidingWindowSplitter

X = np.random.randint(1, 100, size=(10000, 1))
y = np.random.randint(1, 100, size=(10000, 1))

# when use auto-regressive like MultiOutput, pls set ravel = True
# ravel = False, when you are using LSTM which support multiple dimension
splitter = SlidingWindowSplitter(24, 24, ravel=True)
X, y = splitter.split(X, y)

raw_data_model = RegressionDataModel(X, y)
trainer = MultiOutputRegTrainer(['XGBoost'], [None], raw_data_model)
trainer.fit()

Parallel training

• provide fast training using parallel job
• use RegTrainer as base, but add Parallel running

import numpy as np

from pyemits.core.ml.regression.trainer import RegressionDataModel, ParallelRegTrainer

X = np.random.randint(1, 100, size=(10000, 1))
y = np.random.randint(1, 100, size=(10000, 1))

raw_data_model = RegressionDataModel(X, y)
trainer = ParallelRegTrainer(['XGBoost', 'LightGBM'], [None, None], raw_data_model)
trainer.fit()

or you can use RegTrainer for multiple model, but it is not in Parallel job

import numpy as np

from pyemits.core.ml.regression.trainer import RegressionDataModel, RegTrainer

X = np.random.randint(1, 100, size=(10000, 1))
y = np.random.randint(1, 100, size=(10000, 1))

raw_data_model = RegressionDataModel(X, y)
trainer = RegTrainer(['XGBoost', 'LightGBM'], [None, None], raw_data_model)
trainer.fit()

KFold training

• KFoldConfig is global config, will apply to all

import numpy as np

from pyemits.core.ml.regression.trainer import RegressionDataModel, KFoldCVTrainer
from pyemits.common.config_model import KFoldConfig

X = np.random.randint(1, 100, size=(10000, 1))
y = np.random.randint(1, 100, size=(10000, 1))

raw_data_model = RegressionDataModel(X, y)
trainer = KFoldCVTrainer(['XGBoost', 'LightGBM'], [None, None], raw_data_model,
                         {'kfold_config': KFoldConfig(n_splits=10)})
trainer.fit()

Easy prediction

import numpy as np
from pyemits.core.ml.regression.trainer import RegressionDataModel, RegTrainer
from pyemits.core.ml.regression.predictor import RegPredictor

X = np.random.randint(1, 100, size=(10000, 1))
y = np.random.randint(1, 100, size=(10000, 1))

raw_data_model = RegressionDataModel(X, y)
trainer = RegTrainer(['XGBoost', 'LightGBM'], [None, None], raw_data_model)
trainer.fit()

predictor = RegPredictor(trainer.clf_models, 'RegTrainer')
predictor.predict(RegressionDataModel(X))

Forecast at scale

• see examples: forecast at scale.ipynb

Data Model

from pyemits.common.data_model import RegressionDataModel
import numpy as np

X = np.random.randint(1, 100, size=(1000, 10, 10))
y = np.random.randint(1, 100, size=(1000, 1))

data_model = RegressionDataModel(X, y)

directly write an attribute to the data model

data_model._update_attributes('X_shape', (1000, 10, 10))
data_model.X_shape
>> > (1000, 10, 10)

write something to the meta data

data_model.add_meta_data('dimension', (1000, 10, 10))
data_model.meta_data
>> > {'dimension': (1000, 10, 10)}

Anomaly detection (partial finished)

• see: anomaly detection
• root cause analyzer (under development)
• Kalman filter (under development)

from pyemits.core.ml.anomaly_detection.predictor import AnomalyPredictor
from pyemits.core.ml.anomaly_detection.trainer import AnomalyTrainer, PyodWrapper
from pyemits.common.data_model import AnomalyDataModel
from pyemits.common.config_model import PyodIforestConfig
from pyod.models.iforest import IForest
from pyod.utils import generate_data

contamination = 0.1  # percentage of outliers
n_train = 1000  # number of training points
n_test = 200  # number of testing points

X_train, y_train, X_test, y_test = generate_data(
    n_train=n_train, n_test=n_test, contamination=contamination)

# highly flexible model config, accept str, PyodWrapper with/without initialized model
# either one
trainer = AnomalyTrainer(['IForest', PyodWrapper(IForest()), PyodWrapper(IForest), 'IForest', 'IForest', 'IForest'],
                         None, AnomalyDataModel(X_train))
trainer = AnomalyTrainer([PyodWrapper(IForest(contamination=0.05)), PyodWrapper(IForest)],
                         [None, PyodIforestConfig(contamination=0.05)], AnomalyDataModel(X_train))
trainer.fit()

# option 1
predictor = AnomalyPredictor(trainer.clf_models)

# option 2
predictor = AnomalyPredictor(trainer.clf_models,
                             other_config={'standard_scaler': predictor.misc_container['standard_scaler']})

# option 3
predictor = AnomalyPredictor(trainer.clf_models,
                             other_config={'standard_scaler': predictor.misc_container['standard_scaler'],
                                           'combination_config': {'n_buckets': 5}}, combination_method='moa')

predictor.predict(AnomalyDataModel(X_test))

Data processing pipeline

it features in the following:

• easy configuration
  • register steps, tasks in data processing pipeline
• log data result in each tasks, each steps
• record the flow of pipeline, from steps to work (from marco to micro)

you can embed other function features in the task, but parameter: "data" is required to be passed in

e.g. add email notification, add log, upload to database etc...

from pyemits.core.preprocessing.pipeline import DataNode, NumpyDataNode, PandasDataFrameDataNode, PandasSeriesDataNode,

Pipeline, Step, Task
import pandas as pd
import numpy as np

df = pd.DataFrame(np.random.random(size=(20, 20)))

dn = PandasDataFrameDataNode.from_pandas(df)

def sum_each_col(data, a=1, b=2):
    return data.sum()

def sum_series(data):
    return np.array([data.sum()])

task registration and arguments registration

task_a = Task(sum_each_col)
task_a.register_args(a=10, b=10)
task_b = Task(sum_series)

pipeline register step and execute

pipeline = Pipeline()

step_a = Step('step_a', [task_a], '')
step_b = Step('step_b', [task_b], '')

pipeline.register_step(step_a)
pipeline.register_step(step_b)
pipeline.execute(dn)

know the steps and its tasks from start to end

pipeline.get_step_task_mapping()
>> > {0: ('test', ['sum_each_col']), 1: ('test1', ['sum_series'])}

know the snapshot result in each steps, each tasks, friendly to data scientist for debugging

pipeline.get_pipeline_snapshot_res(step_id=1,tasks_id=0)
> > > array([197.70351007])

Evaluation (under development)

• see module: evaluation
• backtesting
• model evaluation

Ensemble (under development)

• blending
• stacking
• voting
• by combo package
  • moa
  • aom
  • average
  • median
  • maximization

IO

• refer to the following package

dashboard ???

other miscellaneous feature

• continuous evaluation
• aggregation
• dimensional reduction
• data profile (intensive data overview)

to be confirmed

....

References

the following libraries gave me some idea/insight

1. greykit
   1. changepoint detection
   2. model summary
   3. seaonality
2. pytorch-forecasting
3. darts
4. pyaf
5. orbit
6. kats/prophets by facebook
7. sktime
8. gluon ts
9. tslearn
10. pyts
11. luminaries
12. tods
13. autots
14. pyodds
15. scikit-hts