API Reference: utils

torchlight.utils stores and restores models.

from torchlight.utils import save_model, load_model, save_state, load_state
from torchlight.utils import state_dict, load_state_dict

Whole-model save/load

def save_model(model: Any, path: str) -> None
def load_model(path: str) -> Any

Pickle / unpickle an entire Module.

save_model(model, "model.pkl")
loaded = load_model("model.pkl")

State dicts (portable, numpy)

def save_state(module: Module, path: str) -> None
def load_state(module: Module, path: str) -> None

Write/read a model's named state as a .npz file (no pickling).

save_state(model, "state.npz")
load_state(new_model, "state.npz")   # same weights, no pickle

Dict helpers

def state_dict(module: Module) -> Dict[str, np.ndarray]
def load_state_dict(module: Module, state, strict=True) -> None

state_dict() returns the numpy arrays behind every named parameter. With strict=True (default) load_state_dict raises if the names don't match exactly.

state = state_dict(model)              # {'linear.weight': np.ndarray, ...}
load_state_dict(model, state)          # restore in place