Imbrium
Standard and Hybrid Deep Learning Multivariate-Multi-Step & Univariate-Multi-Step Time Series Forecasting.
Install / Use
/learn @maxmekiska/ImbriumREADME
imbrium 
Status
| Build | Status|
|---|---|
| MAIN BUILD | 
|
| DEV BUILD | 
|
Pip install
pip install imbrium
Standard and Hybrid Deep Learning Multivariate-Multi-Step & Univariate-Multi-Step Time Series Forecasting.
██╗███╗░░░███╗██████╗░██████╗░██╗██╗░░░██╗███╗░░░███╗
║████╗░████║██╔══██╗██╔══██╗██║██║░░░██║████╗░████║
██║██╔████╔██║██████╦╝██████╔╝██║██║░░░██║██╔████╔██║
██║██║╚██╔╝██║██╔══██╗██╔══██╗██║██║░░░██║██║╚██╔╝██║
██║██║░╚═╝░██║██████╦╝██║░░██║██║╚██████╔╝██║░╚═╝░██║
╚═╝╚═╝░░░░░╚═╝╚═════╝░╚═╝░░╚═╝╚═╝░╚═════╝░╚═╝░░░░░╚═╝
Introduction to imbrium
imbrium is a deep learning library that specializes in time series forecasting. Its primary objective is to provide a user-friendly repository of deep learning architectures for this purpose. The focus is on simplifying the process of creating and applying these architectures, with the goal of allowing users to create complex architectures without having to build them from scratch. Instead, the emphasis shifts to high-level configuration of the architectures.
imbrium Summary
imbrium is designed to simplify the application of deep learning models for time series forecasting. The library offers a variety of pre-built architectures. The user retains full control over the configuration of each layer, including the number of neurons, the type of activation function, loss function, optimizer, and metrics applied. This allows for the flexibility to adapt the architecture to the specific needs of the forecast task at hand. Imbrium also offers a user-friendly interface for training and evaluating these models, making it easy to quickly iterate and test different configurations.
imbrium uses the sliding window approach to generate forecasts. The sliding window approach in time series forecasting involves moving a fixed-size window - steps_past through historical data, using the data within the window as input features. The next data points outside the window are used as the target variables - steps_future. This method allows the model to learn sequential patterns and trends in the data, enabling accurate predictions for future points in the time series.
Get started with imbrium
<details> <summary>Expand</summary> <br>Note, if you choose to use TensorBoard, run the following command to display the results:
tensorboard --logdir logs/
from imbrium import PureUni
# create a PureUni object (numpy array expected)
predictor = PureUni(target = target_numpy_array, evaluation_split = 0.1, validation_split = 0.2)
# the following models are available for a PureUni objects;
# create and fit a muti-layer perceptron model
predictor.create_fit_mlp(
steps_past,
steps_future,
optimizer = "adam",
optimizer_args = None,
loss = "mean_squared_error",
metrics = "mean_squared_error",
dense_block_one = 1,
dense_block_two = 1,
dense_block_three = 1,
layer_config = {
"layer0": {
"config": {
"neurons": 50,
"activation": "relu",
"regularization": 0.0,
"dropout": 0.0,
}
},
"layer1": {
"config": {
"neurons": 50,
"activation": "relu",
"regularization": 0.0,
"dropout": 0.0,
}
},
"layer2": {
"config": {"neurons": 50, "activation": "relu", "regularization": 0.0}
},
},
epochs = 100,
batch_size = None,
show_progress = 1,
board = False,
)
# create and fit a recurrent neural network model
predictor.create_fit_rnn(
steps_past,
steps_future,
optimizer = "adam",
optimizer_args = None,
loss = "mean_squared_error",
rnn_block_one = 1,
rnn_block_two = 1,
rnn_block_three = 1,
metrics = "mean_squared_error",
layer_config = {
"layer0": {
"config": {
"neurons": 50,
"activation": "relu",
"regularization": 0.0,
"dropout": 0.0,
}
},
"layer1": {
"config": {
"neurons": 50,
"activation": "relu",
"regularization": 0.0,
"dropout": 0.0,
}
},
"layer2": {
"config": {"neurons": 50, "activation": "relu", "regularization": 0.0}
},
},
epochs = 100,
batch_size = None,
show_progress = 1,
board = False,
)
# create and fit a long short-term neural network model
predictor.create_fit_lstm(
steps_past,
steps_future,
optimizer = "adam",
optimizer_args = None,
loss = "mean_squared_error",
metrics = "mean_squared_error",
lstm_block_one = 1,
lstm_block_two = 1,
lstm_block_three = 1,
layer_config = {
"layer0": {
"config": {
"neurons": 50,
"activation": "relu",
"regularization": 0.0,
"dropout": 0.0,
}
},
"layer1": {
"config": {
"neurons": 50,
"activation": "relu",
"regularization": 0.0,
"dropout": 0.0,
}
},
"layer2": {
"config": {"neurons": 50, "activation": "relu", "regularization": 0.0}
},
},
epochs = 100,
batch_size = None,
show_progress = 1,
board = False,
)
# create and fit a convolutional neural network
predictor = create_fit_cnn(
steps_past,
steps_future,
optimizer = "adam",
optimizer_args = None,
loss = "mean_squared_error",
metrics = "mean_squared_error",
conv_block_one = 1,
conv_block_two = 1,
dense_block_one = 1,
layer_config = {
"layer0": {
"config": {
"filters": 64,
"kernel_size": 1,
"activation": "relu",
"regularization": 0.0,
"dropout": 0.0,
}
},
"layer1": {
"config": {
"filters": 32,
"kernel_size": 1,
"activation": "relu",
"regularization": 0.0,
"dropout": 0.0,
}
},
"layer2": {
"config": {
"pool_size": 2,
}
},
"layer3": {
"config": {
"neurons": 32,
"activation": "relu",
"regularization": 0.0,
}
},
},
epochs = 100,
batch_size = None,
show_progress = 1,
board = False,
)
# create and fit a gated recurrent unit neural network
predictor.create_fit_gru(
steps_past,
steps_future,
optimizer = "adam",
optimizer_args = None,
loss = "mean_squared_error",
metrics = "mean_squared_error",
gru_block_one = 1,
gru_block_two = 1,
gru_block_three = 1,
layer_config = {
"layer0": {
"config": {
"neurons": 50,
"activation": "relu",
"regularization": 0.0,
"dropout": 0.0,
}
},
"layer1": {
"config": {
"neurons": 50,
"activation": "relu",
"regularization": 0.0,
"dropout": 0.0,
}
},
"layer2": {
"config": {
"neurons": 50,
"activation": "relu",
"regularization": 0.0,
}
},
},
epochs = 100,
batch_size = None,
show_progress = 1,
board = False,
)
# create and fit a bidirectional recurrent neural network
predictor.create_fit_birnn(
steps_past,
steps_future,
optimizer = "adam",
optimizer_args = None,
loss = "mean_squared_error",
metrics = "mean_squared_error",
birnn_block_one = 1,
rnn_block_one = 1,
layer_config = {
"layer0": {
"config": {
"neurons": 50,
"activation": "relu",
"regularization": 0.0,
"dropout": 0.0,
}
},
"layer1": {
"config": {
"neurons": 50,
