Synthetic Data Pipeline

To enable scalable and easy-to-run machine learning experiments on time-series data, SimbaML offers multiple pipelines covering data pre-processing, training, and evaluation of ML models. The synthetic data piplines and mixed data pipelines can be run by passing them the corresponding config files.

Configure Synthetic Data Pipeline

All provided machine learning pipelines of SimbaML can be configured based on config files. This way, users can change the models that are going to be trained, their hyperparameters, specify details for preprocessing and much more.

# Define the models to be used for training
[[models]]
id = "KerasDenseNeuralNetworkTransferLearning"
normalizer = true
[models.training_params]
epochs = 10
patience = 5
batch_size = 32
validation_split = 0.2
verbose = 0
[models.archictecture_params]
units = [32, 32]
activation = "relu"

[[models]]
id = "PytorchLightningDenseNeuralNetwork"
normalizer = true
[models.training_params]
epochs = 10
patience = 5
batch_size = 32
validation_split = 0.2
verbose = 0
[models.archictecture_params]
units = [32, 32]
activation = "relu"

[[models]]
id = "DecisionTreeRegressor"
[models.model_params]
criterion = "squared_error" # "squared_error", "absolute_error"
splitter ="best" # "best", "random"

[[models]]
id = "LinearRegressor"
[models.model_params]
fit_intercept= true
n_jobs = 1 # TODO: What is good value here
positive = false

[[models]]
id = "NearestNeighborsRegressor"
[models.model_params]
n_neighbors = 5
weights = "uniform" # "uniform", "distance"

[[models]]
id = "RandomForestRegressor"
[models.model_params]
n_estimators = 100
max_depth = 10
min_samples_split = 2
min_samples_leaf = 1
min_weight_fraction_leaf = 0.0

[[models]]
id = "SVMRegressor"
kernel = "linear" # "linear", "poly", "rbf", "sigmoid", "precomputed"

# Define the metrics to be used for evaluating the model
metrics = [
  "r_square",
  "mean_absolute_error",
  "mean_squared_error",
  "mean_absolute_percentage_error",
  "root_mean_squared_error",
  "normalized_root_mean_squared_error",
]

# Define data configurations
[data]
synthetic = "path/to/synthetic/data"
test_split = 0.2 # [0, 1]
split_axis = "vertical" # "vertical", "horizontal"

# Define the time series configurations
[data.time_series]
input_features = ["Infected", "Recovered"]
output_features = ["Infected", "Recovered"]
input_length = 1
output_length = 1

# Define the plugins you wrote to use your own models
plugins = [
  "tests.example_plugin",
  "simba_ml.prediction.time_series.models.keras",
  "simba_ml.prediction.time_series.models.pytorch_lightning"
]

# If you include this section, the results will be logged to wandb
# make sure to specify the right project and entity
# [logging]
# project = "your-wandb-project"
# entity = "your-wandb-entity"

Start Synthetic Data Pipeline

$ simba_ml start-prediction synthetic_data –config-path synthetic_data_pipeline.toml