Checkpointing

Loading

Pretrained ML models can be loaded in APPFL via torch.load().

The model parameters obtained from the pretrained model can be used as an initial guess of the FL model parameters in APPFL. To use this feature, one should revise the corresponding fields in the configuration class Config. For example, suppose that we have a pretrained model stored as model_pretrained.pt in a examples/models directory. Then, one can revise the configurations as follows:

# Loading Configurations
from OmegaConf import OmegaConf
from appfl.config import Config
cfg = OmegaConf.structured(Config)
# Loading Models
cfg.load_model = True
cfg.load_model_dirname = "./models"
cfg.load_model_filename = "model_pretrained"

Saving

After federated learning, the resulting models can be stored via torch.save(). To use this feature, one should revise the configuration accordingly as well. See the followings for an example:

# Saving Models
cfg.save_model = True
cfg.save_model_dirname = "./save_models"
cfg.save_model_filename = "model"
cfg.checkpoints_interval = 2

By setting checkpoints_interval = 2, trained model will be saved for every 2 iteration.

Note

When using docker container, one can download the trained models via docker cp (https://docs.docker.com/engine/reference/commandline/cp/).

For example, if my container ID is aa90d20f96c0d143012d2e6ca7d7820ed9ed8a36b163cddf8bfd6dd0e6228dab then

docker cp aa90d20f96c0d143012d2e6ca7d7820ed9ed8a36b163cddf8bfd6dd0e6228dab:/APPFL/save_models/ .