dynamodelx 0.1.3

Dynamic Models for ML with Uncertainty

DynaModelX

---

Introduction

DynaModelX stands for Dynamic model extension, it is built as a side project for quick modeling with minimal UI. This package can dynamically build, train and return model performance according to the user given specifications.

To install DynaModelX, run

pip install dynamodelx

DynaModelX contains:

1. UFA - Universal Function Approximator for tabular data.
2. BNNRegressor - Bayesian Neural Network Regressor for tabular data.

Universal Function Approximator (UFA)

---

UFA can mainly perform 6 tasks on tabular data (atleast for now):

1. Univariate Regression
2. Multivariate Regression
3. Heteroscedastic Regression (Single Target)
4. Heteroscedastic Regression (Multi-Target)
5. Binary Classification (Bernoulli output)
6. Multi-Class Classification (Categorical Output)

UFA is built on top of Pytorch, we can alternate with these tasks by just changing some couple of parameters.

Importing The Estimator

To use UFA, import the estimator from dynamodelx

from dynamodelx import UFA

Using UFA

For Univariate Regression,

from sklearn.datasets import fetch_california_housing
from dynamodelx import UFA
from dynamodelx.plots import draw_plots

data = fetch_california_housing()
X, y = data.data, data.target

ufa = UFA(  
            task = 'regression', 
            model_size = 'small', 
            input_dim = X.shape[1],
            output_dim = 1,
            loss = 'mean_square_loss',
            device = 'cuda',
            weights_init = 'he', 
            hidden_activation = 'relu',
            optimizer = 'adam',
            uncertainty = False,
            multiclass = False,
            custom_architecture = None, 
            return_metrics = True,
            auto_build = True
        )

performance = ufa.train(X, y, epochs=50, batch_size=32, learning_rate=0.01, val_size=0.3, test_size=0.2)

draw_plots(performance=performance)

prediction = ufa.predict(X[:5])

ufa.save(parameters_path='ufa_model.pt', arguments_path='ufa_args.pt')
ufa = UFA.load(parameters_path='ufa_model.pt', arguments_path='ufa_args.pt')

Functions in UFA:

• build - to build the model architecture according to the user specifications.
• train - to train the model on the given dataset.
• predict - to make predictions on new data. Returns prediction, or prediction and standard-deviation (if uncertainty=True).
• save - to save the trained model to the disk.
• load - to load a saved model from the disk.

The hyperparameters for this estimator to initialize UFA are as follows:

• task represents the kind of task we are performing, it takes a string. Either 'regression' or 'classification'.
• model_size represents the size of the model, it takes one of these values:
  • 'small' - contains 2 hidden layers. 64 and 32 neurons respectively.
  • 'medium' - contains 3 hidden layers. 128, 64 and 32 neurons respectively.
  • 'large' - contains 4 layers. 256, 128, 64 and 32 neurons respectively.
  • None - when we want to use our very own custom architecture, note the model_size should be None if custom_architecture is given.(By default, custom_architecture is None. More about it discussed below)
• input_dim - takes an integer, the feature-space of data.
• output_dim - takes an integer, depends on the task we are solving. For example, 1 if it's univariate regression or heteroscedastic regression (single target) or binary classification, d (required ouput dimension)if it's multivariate regression or heteroscedastic regression (multi-target) or multi-class classification.
• loss represents the loss function we'll use to optimize the model, it takes one of these strings:
  • 'mean_square_loss' - a classic loss function for regression.
  • 'gaussian_nll_loss' - for regression with uncertainty (heteroscedastic regression).
  • 'binary_cross_entropy' - for binary classification.
  • 'cross_entropy_loss' - for multi-class classification.
• device takes any of these strings:
  • 'cpu' - model uses cpu to train.
  • 'cuda' or 'cuda:n' - model uses cuda to train if it's available, where n is the gpu index.
  • By default, it's set to 'cuda'.
• weights_init takes any of these values:
  • 'xavier' - for xavier/glorot uniform weight initialization. Usually used with sigmoid/tanh activations. Acts with gain=1.0.
  • 'he' - for he uniform weight initialization. Usually used with relu/leaky relu activations. Acts with a=0, mode='fan_in', nonlinearity='leaky_relu'.
  • 'uniform' - for uniform weight initialization. Acts with a=0.0, b=1.0.
  • 'normal' - for normal weight initialization. Acts with mean=0.0, std=1.0.
  • None - pytorch default weight initialization.
  • By default, it's set to None.
• hidden_activation takes any of these strings:
  • 'relu' - Rectified Linear Unit activation.
  • 'leaky_relu' - Leaky Rectified Linear Unit activation.
  • 'prelu' - Parametric Rectified Linear Unit activation.
  • 'elu' - Exponential Linear Unit activation.
  • 'sigmoid' - Sigmoid activation.
  • 'tanh' - Hyperbolic Tangent activation.
  • 'gelu' - Gaussian Error Linear Unit activation.
  • 'mish' - Mish activation.
  • By default, it's set to 'relu'.
• optimizer takes any of these strings:
  • 'adam' - Adam optimizer.
  • 'adamw' - AdamW optimizer.
  • 'sgd' - Stochastic Gradient Descent optimizer.
  • By default, it's set to 'adam'.
• uncertainty takes a boolean value, True if we want to perform regression with uncertainty (heteroscedastic regression), False otherwise. By default, it's set to False. If it's set to True, the loss function should be 'gaussian_nll_loss' and the model will output both prediction and standard-deviation for each input sample. Note,
  • Making uncertainty True will work only when the task is set to 'regression'.
  • Model also shows PICP and MPIW metrics for test data
  • For multi-target heteroscedastic regression, the model will output both prediction and standard-deviation for each target variable.
  • We don't have to include standard-deviation as a separate output dimension, the model will automatically takes care of it. If the output_dim is set to d, the model will output d predictions and d standard-deviation values.
• multiclass takes a boolean value, True if we want to perform multi-class classification, False otherwise. By default, it's set to False. If it's set to True, the output_dim should be greater than 1 and the loss function should be 'cross_entropy_loss'. Note, making multiclass True will work only when the task is set to 'classification'.
• custom_architecture takes a list of integers, representing the number of neurons in each hidden layer. For example, [128, 64, 32] represents a model with 3 hidden layers with 128, 64 and 32 neurons respectively. By default, it's set to None. If it's set to None, the model will be built according to the model_size parameter. If we want to use our own custom architecture, we need to set the model_size parameter to None and provide the architecture through this parameter. Note this is only for hidden layers and the length of the list should be at least 1 and all the values should be positive integers greater than 0. The output layer neurons are defined by the output_dim parameter, no need to mention it here.
• return_metrics takes a boolean value, True if we want to return performance metrics after training, False otherwise. By default, it's set to True. If it's set to True, the train method will return a dictionary of type TrainingHistory containing training loss, validation and test losses and their metrics.
• auto_build takes a boolean value, True if we want to automatically build the model after initialization, False otherwise. By default, it's set to True. If it's set to False, we need to manually call the build methods after initialization to build the model.

The hyperparameters for the train method of UFA are as follows:

• X and y are the feature matrix and target vector respectively. Note that y should be in integer format (0, 1, 2, ..., n-1) for multi-class classification or 0 and 1 for binary classification.
• epochs takes an integer, number of epochs to train the model.
• learning_rate takes a float value, learning rate for the optimizer.
• momentum takes a float value between 0 and 1, momentum for the SGD optimizer. By default, it's set to None. This parameter is used only when the optimizer is set to 'sgd'.
• val_size takes a float value between 0 and 1, representing the proportion of validation data from the whole dataset. By default, it's set to 0.2.
• test_size takes a float value between 0 and 1, representing the proportion of test data from the whole dataset. By default, it's set to 0.1.
• batch_size takes an integer, number of samples per batch. By default, it's set to 32.

metrics returned for different tasks are as follows:

• For Regression:

  • Mean Absolute Error (MAE)
  • R2 Score

• For Classification:

  • Accuracy
  • Precision
  • Recall
  • F1 Score

Bayesian Neural Network Regressor (BNNRegressor)

---

BNNRegressor can perform regression tasks on tabular data with uncertainty estimation. It is built on top of pytorch to provide a simple interface for training Bayesian Neural Networks using Variational Inference. It handles both aleatoric and epistemic uncertainties.

Importing The Estimator

To use BNNRegressor, import the estimator from dynamodelx

from dynamodelx import BnnRegressor

Using BNNRegressor

For Univariate Regression with Uncertainty,

from sklearn.datasets import fetch_california_housing
from dynamodelx import BNNRegressor
from dynamodelx.plots import draw_plots
data = fetch_california_housing()
X, y = data.data, data.target

bnn = BnnRegressor(
    model_size='small',
    input_dim=X.shape[1],
    output_dim = 1,
    device = 'cuda',
    hidden_activation = 'gelu',
    optimizer = 'adam',
    mc_samples = 20,
    custom_architecture = None,
    return_metrics = True,
    auto_build = True
)

performance = bnn.train(X=X, y=y, epochs=50, learning_rate = 0.01, momentum=None, val_size=0.2, test_size=0.1, batch_size=120)
draw_plots(performance)
prediction, std_dev = bnn.predict(X[:5])
bnn.save(parameters_path='bnn_model.pt', arguments_path='bnn_args.pt')
bnn = BnnRegressor.load(parameters_path='bnn_model.pt', arguments_path='bnn_args.pt')

Functions in BNNRegressor:

• build - to build the model architecture according to the user specifications.
• train - to train the model on the given dataset.
• predict - to make predictions on new data. Returns prediction and standard-deviation.
• save - to save the trained model to the disk.
• load - to load a saved model from the disk.

The hyperparameters for this estimator to initialize BNNRegressor are as follows:

• model_size represents the size of the model, it takes one of these values:
  • 'small' - contains 2 hidden layers. 64 and 32 neurons respectively.
  • 'medium' - contains 3 hidden layers. 128, 64 and 32 neurons respectively.
  • 'large' - contains 4 layers. 256, 128, 64 and 32 neurons respectively.
  • None - when we want to use our very own custom architecture, note the model_size should be None if custom_architecture is given.(By default, custom_architecture is None. More about it discussed below)
• input_dim - takes an integer, the feature-space of data.
• output_dim - takes an integer, number of output dimensions.
• device takes any of these strings:
  • 'cpu' - model uses cpu to train.
  • 'cuda or 'cuda:n' - model uses cuda to train if it's available, where n is the gpu index.
  • By default, it's set to 'cuda'.
• hidden_activation takes any of these strings:
  • 'relu' - Rectified Linear Unit activation.
  • 'leaky_relu' - Leaky Rectified Linear Unit activation.
  • 'prelu' - Parametric Rectified Linear Unit activation.
  • 'elu' - Exponential Linear Unit activation.
  • 'sigmoid' - Sigmoid activation.
  • 'tanh' - Hyperbolic Tangent activation.
  • 'gelu' - Gaussian Error Linear Unit activation.
  • 'mish' - Mish activation.
  • By default, it's set to 'relu'.
• optimizer takes any of these strings:
  • 'adam' - Adam optimizer.
  • 'adamw' - AdamW optimizer.
  • 'sgd' - Stochastic Gradient Descent optimizer.
  • By default, it's set to 'adam'.
• mc_samples takes an integer, number of Monte Carlo samples to estimate the predictive distribution on validation data. By default, it's set to 10.
• custom_architecture takes a list of integers, representing the number of neurons in each hidden layer. For example, [128, 64, 32] represents a model with 3 hidden layers with 128, 64 and 32 neurons respectively. By default, it's set to None. If it's set to None, the model will be built according to the model_size parameter. If we want to use our own custom architecture, we need to set the model_size parameter to None and provide the architecture through this parameter. Note this is only for hidden layers and the length of the list should be at least 1 and all the values should be positive integers greater than 0. The output layer neurons are defined by the output_dim parameter, no need to mention it here.
• return_metrics takes a boolean value, True if we want to return performance metrics after training, False otherwise. By default, it's set to True. If it's set to True, the train method will return a dictionary of type TrainingHistory containing training loss, validation and test losses and their metrics.
• auto_build takes a boolean value, True if we want to automatically build the model after initialization, False otherwise. By default, it's set to True. If it's set to False, we need to manually call the build methods after initialization to build the model.

The hyperparameters for the train method of BNNRegressor are as follows:

• X and y are the feature matrix and target vector respectively.
• epochs takes an integer, number of epochs to train the model.
• learning_rate takes a float value, learning rate for the optimizer.
• momentum takes a float value between 0 and 1, momentum for the SGD optimizer. By default, it's set to None. This parameter is used only when the optimizer is set to 'sgd'. Model ignores this parameter if the optimizer is not 'sgd'.
• val_size takes a float value between 0 and 1, representing the proportion of validation data from the whole dataset. By default, it's set to 0.2.
• test_size takes a float value between 0 and 1, representing the proportion of test data from the whole dataset. By default, it's set to 0.1.
• batch_size takes an integer, number of samples per batch. By default, it's set to 32.

metrics returned after training for regression task are as follows:

• Mean Absolute Error (MAE) on validation and test data.
• R2 Score on validation and test data.

Saving The Model

To save the trained model, use the save method of UFA and BNNRegressor.

ufa.save(parameters_path='ufa_model.pt', arguments_path='ufa_args.pt')
bnn.save(parameters_path='bnn_model.pt', arguments_path='bnn_args.pt')

This method takes the file names (with .pth, .pt, .ckpt, .bin extensions) as input and saves the trained model in the current working directory. parameters_path is the file name to save the model parameters (state_dict) and arguments_path is the file name to save the model initialization arguments.

Loading The Model

To load a saved model, use the load method of UFA and BNNRegressor.

from dynamodelx import UFA, BNNRegressor

ufa = UFA.load(parameters_path='ufa_model.pt', arguments_path='ufa_args.pt')
bnn = BNNRegressor.load(parameters_path='bnn_model.pt', arguments_path='bnn_args.pt')

This method takes the file names (with .pth, .pt, .ckpt, .bin extensions) as input and loads the saved model from the current working directory. parameters_path is the file name to load the model parameters (state_dict) and arguments_path is the file name to load the model initialization arguments.

Plots

To get training and validation loss plots, use the function draw_plots from dynamodelx.plots

from dynamodelx.plots import draw_plots
draw_plots(performance)

This function takes the performance dictionary of type TrainingHistory returned by the train method of UFA, BNNRegressor as input and plots the training, validation and test losses and their metrics.

Examples

More examples can be found in the examples folder of the repository.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Acknowledgements

• Built on top of PyTorch
• Inspired by various machine learning libraries and frameworks.