ssbgm is a python library which enables you to generate synthetic data using a score based generative model with scikit-learn.
You can use ssbgm to predict a target value with some features by generating synthetic data given the features.
- Python (>= 3.10)
- libraries:
- catboost>=1.2.7
- lightgbm>=4.5.0
- numpy>=1.26.4
- scikit-learn>=1.5.2
- tqdm>=4.67.0
- types-tqdm>=4.66.0.20240417
See ./pyproject.toml for more details.
You can install ssbgm via pip:
pip install git+https://github.com/hmasdev/ssbgm.gitor
git clone https://github.com/hmasdev/ssbgm.git
pip install .Here is an example of generating synthetic data using ssbgm:
from sklearn.linear_model import LinearRegression
from ssbgm import ScoreBasedGenerator
# Prepare the dataset which you want to generate synthetic data
# row: sample, column: output dimension
X: np.ndarray = ...
# initialize the generator with LinearRegression
generator = ScoreBasedGenerator(estimator=LinearRegression())
# fit the generator
generator.fit(X)
# generate synthetic data
# Langevin Monte Carlo is used to generate synthetic data
X_syn_lmc = sbmgenerator2.sample(n_samples=128, sampling_method=sbmgenerator2.SamplingMethod.LANGEVIN_MONTECARLO, alpha=0.2).squeeze()
X_syn_euler = sbmgenerator2.sample(n_samples=128, sampling_method=sbmgenerator2.SamplingMethod.EULER).squeeze()
X_syn_em = sbmgenerator2.sample(n_samples=128, sampling_method=sbmgenerator2.SamplingMethod.EULER_MARUYAMA).squeeze()
# The shape of each X_syn_* is (128, X.shape[1])You can use ssbgm to predict a target value with some features by generating synthetic data given the features.
from sklearn.linear_model import LinearRegression
from ssbgm import ScoreBasedGenerator
# Prepare the dataset which you want to generate synthetic data
# row: sample, column: features
X: np.ndarray = ...
# row: sample, column: target value
y: np.ndarray = ...
# initialize the generator with LinearRegression
generator = ScoreBasedGenerator(estimator=LinearRegression())
# fit the generator
generator.fit(X, y)
# predict the target value with on X
y_pred_by_mean, y_pred_std = generator.predict(X, aggregate='mean', return_std=True) # Shape: (X.shape[0], y.shape[1]), (X.shape[0], y.shape[1])
y_pred_by_median = generator.predict(X, aggregate='median') # Shape: (X.shape[0], y.shape[1])
# generate synthetic data conditioned by X
# Langevin Monte Carlo is used to generate synthetic data
X_syn_lmc = sbmgenerator2.sample(X, n_samples=128, sampling_method=sbmgenerator2.SamplingMethod.LANGEVIN_MONTECARLO, alpha=0.2, n_warmup=1000).squeeze()
X_syn_euler = sbmgenerator2.sample(X, n_samples=128, sampling_method=sbmgenerator2.SamplingMethod.EULER).squeeze()
X_syn_em = sbmgenerator2.sample(X, n_samples=128, sampling_method=sbmgenerator2.SamplingMethod.EULER_MARUYAMA).squeeze()
# The shape of each X_syn_* is (128, X.shape[0], X.shape[1])In this section, we will see some examples of using ssbgm.
If you want to know more details, see ./samples directory. Especially, ./samples/cheatsheet.ipynb is a good starting point.
See ./samples/mixed_gaussian_distribution.ipynb for more details.
# import libraries
from catboost import CatBoostRegressor
from lightgbm import LGBMRegressor
import matplotlib.pyplot as plt # Installing matplotlib is required
import numpy as np
import sys
sys.path.append('../')
from ssbgm import ScoreBasedGenerator
np.random.seed(0)
N = 10000# Case: 1d mixed gaussian
# generate a training dataset
x_train = np.random.randn(N) + (2*(np.random.rand(N) > 0.5) - 1) * 1.6
# train a generative model with score-based model
generative_model_1d_mixed_gaussian = ScoreBasedGenerator(LGBMRegressor(random_state=42)).fit(x_train, noise_strengths=np.sqrt(np.logspace(-3, np.log(x_train.var()), 101)))
# generate samples from the trained model
x_gen = generative_model_1d_mixed_gaussian.sample(n_samples=N, sampling_method=ScoreBasedGenerator.SamplingMethod.EULER).squeeze()
# plot the results
true_pdf = lambda x: 0.5*np.exp(-0.5*(x-1.6)**2)/np.sqrt(2*np.pi) + 0.5*np.exp(-0.5*(x+1.6)**2)/np.sqrt(2*np.pi)
plt.hist(x_train, bins=30, label='train data', color='blue', alpha=0.5, density=True)
plt.hist(x_gen, bins=30, label='generated data', color='red', alpha=0.5, density=True)
plt.plot(np.linspace(x_train.min(), x_train.max()), true_pdf(np.linspace(x_train.min(), x_train.max())), 'k-', label='true pdf')
plt.legend(loc='upper left')
plt.show()[LightGBM] [Info] Auto-choosing row-wise multi-threading, the overhead of testing was 0.004026 seconds.
You can set force_row_wise=true to remove the overhead.
And if memory is not enough, you can set force_col_wise=true.
[LightGBM] [Info] Total Bins 357
[LightGBM] [Info] Number of data points in the train set: 1010000, number of used features: 2
[LightGBM] [Info] Start training from score 0.000087
# Case: 2d mixed gaussian
# generate a training dataset
X_train = np.random.randn(N, 2)
label = 2*(np.random.rand(N) > 0.5) - 1
X_train[:, 0] = X_train[:, 0] + label * 1.6
X_train[:, 1] = X_train[:, 1] + label * 1.6
# train a generative model with score-based model
generative_model_2d_mixed_gaussian = ScoreBasedGenerator(
estimator=CatBoostRegressor(
verbose=0,
loss_function='MultiRMSE',
random_state=42,
)
)
generative_model_2d_mixed_gaussian.fit(
X_train,
noise_strengths=np.sqrt(np.logspace(-3, np.log(max(np.var(X_train, axis=0))), 11)),
)
# generate samples from the trained model
X_gen = generative_model_2d_mixed_gaussian.sample(n_samples=N, sampling_method=ScoreBasedGenerator.SamplingMethod.EULER).squeeze()
# plot the results
true_pdf = lambda X: 0.5*np.exp(-0.5*(X[:, 0]-1.6)**2 - 0.5*(X[:, 1]-1.6)**2)/2/np.pi + 0.5*np.exp(-0.5*(X[:, 0]+1.6)**2 - 0.5*(X[:, 1]+1.6)**2)/2/np.pi
XX_, YY_ = np.meshgrid(np.linspace(X_train[:, 0].min(), X_train[:, 0].max()), np.linspace(X_train[:, 1].min(), X_train[:, 1].max()))
plt.scatter(X_train[:, 0], X_train[:, 1], label='train data', color='blue', alpha=0.2, marker='x')
plt.scatter(X_gen[:, 0], X_gen[:, 1], label='generated data', color='red', alpha=0.2, marker='o')
plt.contourf(XX_, YY_, true_pdf(np.c_[XX_.ravel(), YY_.ravel()]).reshape(XX_.shape), alpha=0.5)
plt.legend(loc='upper left')
plt.xlim(X_train[:, 0].min(), X_train[:, 0].max())
plt.ylim(X_train[:, 1].min(), X_train[:, 1].max())
plt.show()
-
Fork the repository: https://github.com/hmasdev/ssbgm
-
Clone the repository
git clone https://github.com/{YOURE_NAME}/ssbgm cd ssbgm -
Create a virtual environment
python -m venv venv source venv/bin/activate -
Install the required packages
pip install -e .[dev]
-
Checkout your working branch
git checkout -b your-working-branch
-
Make your changes
-
Test your changes
pytest flake8 ssbgm tests mypy ssbgm tests
-
Commit your changes
git add . git commit -m "Your commit message"
-
Push your changes
git push origin your-working-branch
-
Create a pull request: https://github.com/hmasdev/ssbgm/compare