def fit(self, features, labels):
self.features, self.labels = torch.FloatTensor(
features), torch.FloatTensor(labels)
# construct data loader
self.data_loader = DataLoader(TensorDataset(self.features,
self.labels),
batch_size=self.batch_size,
shuffle=True)
# now train with pre-specified options
result = self.train(model=self.model,
loader=self.data_loader,
optimizer=self.optimizer,
criterion=self.criterion,
lr_init=self.lr_init,
use_cuda=self.use_cuda,
epochs=self.epochs,
const_lr=self.const_lr)
if self.criterion.noise_var is not None:
# another forwards pass through network to estimate noise variance
preds, targets = utils.predictions(model=self.model,
test_loader=self.data_loader,
regression=True,
cuda=self.use_cuda)
self.var = np.power(np.linalg.norm(preds - targets),
2.0) / targets.shape[0]
print(self.var)
if self.use_cuda:
self.features = self.features.cuda()
self.net_features = self.model(self.features, output_features=True)
return result
def fit(self, features, labels):
self.features, self.labels = torch.FloatTensor(features), torch.FloatTensor(labels)
# construct data loader
# may want to turn shuffle = False for the very smallest datasets (e.g. uci small)
self.data_loader = DataLoader(TensorDataset(self.features, self.labels), batch_size = self.batch_size, shuffle = True)
# now train with pre-specified options
result = self.train(model=self.model, loader=self.data_loader, optimizer=self.optimizer, criterion=self.criterion,
lr_init=self.lr_init, swag_model=self.swag_model, swag=self.use_swag, swag_start=self.swag_start,
swag_freq=self.swag_freq, swag_lr=self.swag_lr, use_cuda=self.use_cuda, epochs=self.epochs, const_lr=self.const_lr)
if self.criterion.noise_var is not None:
# another forwards pass through network to estimate noise variance
preds, targets = utils.predictions(model=self.model, test_loader=self.data_loader, regression=True,cuda=self.use_cuda)
self.var = np.power(np.linalg.norm(preds - targets), 2.0) / targets.shape[0]
print(self.var)
return result
lr = args.lr_init
if (args.swag and (epoch + 1) > args.swag_start) and args.cov_mat:
train_res = utils.train_epoch(train_loader, model, criterion, optimizer, epoch)
else:
train_res = utils.train_epoch(train_loader, model, criterion, optimizer, epoch)
if epoch == 0 or epoch % args.eval_freq == args.eval_freq - 1 or epoch == args.epochs - 1:
test_res = utils.eval(val_loader, model, criterion)
utils.mata_eval(model, meta_valloader, meta_testloader, 'BASELINE')
else:
test_res = {'loss': None, 'accuracy': None}
if args.swag and (epoch + 1) > args.swag_start and (epoch + 1 - args.swag_start) % args.swag_c_epochs == 0:
sgd_preds, sgd_targets = utils.predictions(val_loader, model)
# sgd_res = utils.predict(val_loader, model)
# sgd_preds = sgd_res["predictions"]
# sgd_targets = sgd_res["targets"]
# print("updating sgd_ens")
if sgd_ens_preds is None:
sgd_ens_preds = sgd_preds.copy()
else:
#TODO: rewrite in a numerically stable way
sgd_ens_preds = sgd_ens_preds * n_ensembled / (n_ensembled + 1) + sgd_preds/ (n_ensembled + 1)
n_ensembled += 1
swag_model.collect_model(model)
if epoch == 0 or epoch % args.eval_freq == args.eval_freq - 1 or epoch == args.epochs - 1:
# swag_model.set_swa()
swag_model.sample(0.0)
utils.bn_update(train_loader, swag_model)