def _predictor(truths: List[Truth], preds: List[Prediction], loss_fn, fill_fn,
merging_fn, update_fn):
output = []
epiweeks = preds[0].coords["epiweek"]
# Extract init_weights from update_fn
init_weights = signature(update_fn).parameters["init_weights"].default
weights_history = []
for idx, ew in enumerate(epiweeks):
if idx == 0:
if init_weights is None:
weights = uniform_weights(
[pred.attrs["model"] for pred in preds], ones=False)
else:
weights = init_weights
else:
truth = merging_fn(fill_fn(mask_truths(truths, ew)))
losses = [loss_fn(pred.loc[:(ew - 1)], truth) for pred in preds]
weights = update_fn(losses)
weights = weights / weights.sum()
weights_history.append(weights)
output.append(
weighted_prediction([pred.loc[ew] for pred in preds], weights))
return xr.concat(output, dim="epiweek"), weights_history
def noop(losses: List[Loss], init_weights=None) -> Weight:
"""
Do nothing.
"""
models = [loss.attrs["model"] for loss in losses]
if init_weights is None:
return uniform_weights(models, ones=False)
else:
return init_weights
def hedge(losses: List[Loss], eta: float, init_weights=None) -> Weight:
r"""
Exponential weight update. :math:`w_i(t + 1) = w_i(t) e^{- \eta m_i(t)}`
"""
models = [loss.attrs["model"] for loss in losses]
if init_weights is None:
init_weights = uniform_weights(models)
updates = [np.exp(-eta * np.sum(loss)) for loss in losses]
return init_weights * updates
def mw(losses: List[Loss], eta: float, init_weights=None) -> Weight:
r"""
Multiplicative weight update. :math:`w_i(t + 1) = w_i(t) (1 - \eta m_i(t))`
"""
models = [loss.attrs["model"] for loss in losses]
if init_weights is None:
init_weights = uniform_weights(models)
updates = [np.prod(1 - eta * loss) for loss in losses]
return init_weights * updates
def fixed_share(losses: List[Loss], eta: float, alpha: float, init_weights=None) -> Weight:
r"""
Fixed share update.
"""
models = [loss.attrs["model"] for loss in losses]
M = len(models)
T = len(losses[0])
if init_weights is None:
weights = uniform_weights(models, ones=False)
else:
weights = init_weights
# Vectorize this
for t in range(T):
vs = weights * np.exp([-eta * loss[t] for loss in losses])
weights = (alpha * np.sum(vs) / M) + (1 - alpha) * vs
return weights