Skip to content

Save & Load Models

All estimators in Sapsan depend either on torch_backend or sklearn_backend, depending on the model architecture. Both backends have save and load functions. Thus, no matter whether you are using included estimators or designing your own, both methods will be available. Loaded models can be used either for evaluation or to continue training. In the case of the latter either old or new config parameters can be set.

Saving the Model

To save the model, call:

1
estimator.save(path = {save_path})

For PyTorch, the states of the model and optimizer will be saved, along with the last epoch and loss in {save_path}/model.pt. The config parameters will be saved in {save_path}/params.json

For Sklearn, only the model itself will be saved, in {save_path}/model.pt

Loading the Model

Even though all Sapsan estimators have load method, you can use a dummy estimator to load your model.

PyTorch

Import load_estimator() to load your PyTorch model. You can pass new ModelConfig() parameters as well if you intend to continue training your model.

1
2
3
4
5
6
7
from sapsan.lib.estimator import load_estimator

estimator = CNN3d(config = CNN3dConfig(n_epoch=100),
                  loaders = loaders)

loaded_estimator = load_estimator.load({path_to_model}, 
                                       estimator = estimator)

Sklearn

Sklearn uses a different interface, so you will need to call load_sklearn_estimator()

1
2
3
4
5
6
7
from sapsan.lib.estimator import load_sklearn_estimator

estimator = KRR(config = KRRConfig(),
                loaders = loaders)

loaded_estimator = load_sklearn.estimator.load({path_to_model}, 
                                               estimator = estimator)