def clone_model(model: nn.Module) -> nn.Module:
names_binary = set(name for name, param in named_parameters_binary(model))
model_copied = copy.deepcopy(model)
for name, param in model_copied.named_parameters():
if name in names_binary:
param.is_binary = True
return model_copied
def __init__(self, trainer):
super().__init__(trainer)
self.autocorrelation = Autocorrelation(
n_lags=self.timer.batches_in_epoch,
with_autocorrelation=isinstance(trainer.train_loader.dataset,
MNISTSmall))
self.graph_mcmc = GraphMCMC(named_params=named_parameters_binary(
self.model),
timer=self.timer,
history_heatmap=True)
def __init__(self, trainer, is_active=True, watch_parameters=False):
super().__init__(trainer,
is_active=is_active,
watch_parameters=watch_parameters)
self.autocorrelation = Autocorrelation(
n_lags=self.timer.batches_in_epoch,
with_autocorrelation=isinstance(trainer.train_loader.dataset,
MNISTSmall))
named_param_shapes = iter(
(name, param.shape)
for name, param in named_parameters_binary(trainer.model))
self.graph_mcmc = GraphMCMC(named_param_shapes=named_param_shapes,
timer=self.timer,
history_heatmap=True)
def train_batch(self, images, labels):
return self.train_batch_mcmc(images, labels, named_params=named_parameters_binary(self.model))