Esempio n. 1
0
def train_tempering(model: nn.Module = None, dataset_name="MNIST"):
    if model is None:
        model = NetBinary(fc_sizes=linear_features[dataset_name], batch_norm=False)
    model = binarize_model(model)
    trainer = ParallelTempering(model, criterion=nn.CrossEntropyLoss(), dataset_name=dataset_name,
                                trainer_cls=TrainerMCMCGibbs, n_chains=5, monitor_kwargs=dict(watch_parameters=False))
    trainer.train(n_epoch=100, save=False, with_mutual_info=False, epoch_update_step=1)
    return model
Esempio n. 2
0
def train_mcmc(model: nn.Module = None, dataset_name="MNIST"):
    if model is None:
        model = NetBinary(fc_sizes=linear_features[dataset_name], batch_norm=False)
    model = binarize_model(model)
    trainer = TrainerMCMC(model,
                          criterion=nn.CrossEntropyLoss(),
                          dataset_name=dataset_name,
                          flip_ratio=0.1)
    trainer.train(n_epoch=500, save=False, with_mutual_info=False, epoch_update_step=3)
Esempio n. 3
0
def train_gradient(model: nn.Module = None, is_binary=True, dataset_name="MNIST"):
    if model is None:
        model = NetBinary(fc_sizes=linear_features[dataset_name])
    optimizer = AdamCustomDecay(model.parameters(), lr=1e-2, weight_decay=1e-4)
    scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, mode='min', factor=0.5, patience=10,
                                                           threshold=1e-3, min_lr=1e-4)
    if is_binary:
        model = binarize_model(model, keep_data=False)
        trainer_cls = TrainerGradBinary
    else:
        trainer_cls = TrainerGradFullPrecision
    trainer = trainer_cls(model,
                          criterion=nn.CrossEntropyLoss(),
                          dataset_name=dataset_name,
                          optimizer=optimizer,
                          scheduler=scheduler)
    trainer.train(n_epoch=500, save=False, with_mutual_info=True)