def fit(self, features, labels):
# tran standard SWAG model
results = super().fit(features, labels)
mean, var, cov = self.swag_model.get_space()
if self.use_cuda:
mean, cov = mean.cuda(), cov.cuda()
if self.temperature is None:
self.temperature = features.shape[0] / cov.size(0)
print('Temperature:', self.temperature)
# print(cov.size())
subspace = SubspaceModel(mean, cov)
self.ess_model = EllipticalSliceSampling(self.base,
subspace=subspace,
var=var,
loader=self.data_loader,
criterion=self.criterion,
num_samples=self.num_samples,
use_cuda=self.use_cuda,
*self.args,
**self.kwargs)
train_loss = self.ess_model.fit(scale=self.prior_std,
use_cuda=self.use_cuda,
temperature=self.temperature,
minibatch=self.mb)
print(train_loss)
def fit(self, features, labels):
if self.fxn_smpl_pts == 'subset':
if self.smpl_num > 0:
num_meas_pts = self.smpl_num
else:
num_meas_pts = int(self.smpl_frac * len(features))
meas_features = features[:num_meas_pts]
meas_labels = labels[:num_meas_pts]
features = features[num_meas_pts:]
labels = labels[num_meas_pts:]
results = super().fit(features, labels)
else:
# tran standard SWAG model
results = super().fit(features, labels)
raise NotImplementedError("Not implemented...")
if self.use_cuda:
device = torch.device('cuda')
else:
device = torch.device('cpu')
mean, var, cov = self.swag_model.get_space(
torch.FloatTensor(meas_features).to(
device)) # special subspace class requires measurement points.
if self.use_cuda:
mean, cov = mean.cuda(), cov.cuda()
if self.temperature is None:
self.temperature = features.shape[0] / cov.size(0)
print('Temperature:', self.temperature)
# print(cov.size())
subspace = SubspaceModel(mean, cov)
self.ess_model = EllipticalSliceSampling(self.base,
subspace=subspace,
var=var,
loader=self.data_loader,
criterion=self.criterion,
num_samples=self.num_samples,
use_cuda=self.use_cuda,
*self.args,
**self.kwargs)
train_loss = self.ess_model.fit(scale=self.prior_std,
use_cuda=self.use_cuda,
temperature=self.temperature,
minibatch=self.mb)
print(train_loss)
class ESSRegRunner(RegressionRunner):
def __init__(self, base, epochs, criterion,
batch_size = 50, lr_init=1e-2, momentum = 0.9, wd=1e-4,
swag_lr = 1e-3, swag_freq = 1, swag_start = 50, subspace_type='pca', subspace_kwargs={'max_rank': 20},
use_cuda = False, use_swag = False, const_lr = False, double_bias_lr = False,model_variance=True,
num_samples = 30, scale = 0.5, temperature = 1., mb = False, prior_log_sigma = 1.0, *args, **kwargs):
super(ESSRegRunner, self).__init__(base=base, epochs=epochs, criterion=criterion, model_variance=model_variance,
batch_size = batch_size, lr_init=lr_init, momentum=momentum, wd=wd, use_cuda = use_cuda, use_swag = use_swag,
swag_lr = swag_lr, swag_freq = swag_freq, swag_start = swag_start, subspace_type=subspace_type, subspace_kwargs=subspace_kwargs,
num_samples = num_samples, scale = scale, const_lr=const_lr, double_bias_lr=double_bias_lr, *args, **kwargs)
self.prior_std = np.exp(prior_log_sigma)
self.temperature = temperature
self.mb = mb
self.args = args
self.kwargs = kwargs
self.use_cuda = use_cuda
def fit(self, features, labels):
# tran standard SWAG model
results = super().fit(features, labels)
mean, var, cov = self.swag_model.get_space()
if self.use_cuda:
mean, cov = mean.cuda(), cov.cuda()
if self.temperature is None:
self.temperature = features.shape[0] / cov.size(0)
print('Temperature:', self.temperature)
# print(cov.size())
subspace = SubspaceModel(mean, cov)
self.ess_model = EllipticalSliceSampling(self.base, subspace=subspace, var=var, loader=self.data_loader,
criterion=self.criterion, num_samples=self.num_samples, use_cuda = self.use_cuda,
*self.args, **self.kwargs)
train_loss = self.ess_model.fit(scale=self.prior_std, use_cuda=self.use_cuda, temperature = self.temperature, minibatch=self.mb)
print(train_loss)
def predict(self, features):
return super().predict(features, swag_model=self.ess_model)
class ASLESSRegRunner(RegressionRunner):
def __init__(self,
base,
epochs,
criterion,
batch_size=50,
lr_init=1e-2,
momentum=0.9,
wd=1e-4,
fxn_smple_pts='subset',
smpl_frac=0.1,
smpl_num=0,
swag_lr=1e-3,
swag_freq=1,
swag_start=50,
subspace_type='pca',
subspace_kwargs={'max_rank': 20},
use_cuda=False,
use_swag=False,
const_lr=False,
double_bias_lr=False,
model_variance=True,
num_samples=30,
scale=0.5,
temperature=1.,
mb=False,
prior_log_sigma=1.0,
*args,
**kwargs):
super(ASLESSRegRunner, self).__init__(base=base,
epochs=epochs,
criterion=criterion,
model_variance=model_variance,
batch_size=batch_size,
lr_init=lr_init,
momentum=momentum,
wd=wd,
use_cuda=use_cuda,
use_swag=use_swag,
swag_lr=swag_lr,
swag_freq=swag_freq,
swag_start=swag_start,
subspace_type=subspace_type,
subspace_kwargs=subspace_kwargs,
num_samples=num_samples,
scale=scale,
const_lr=const_lr,
double_bias_lr=double_bias_lr,
*args,
**kwargs)
self.prior_std = np.exp(prior_log_sigma)
self.temperature = temperature
self.mb = mb
self.args = args
self.kwargs = kwargs
self.use_cuda = use_cuda
self.fxn_smpl_pts = fxn_smple_pts
self.smpl_frac = smpl_frac
self.smpl_num = smpl_num
def fit(self, features, labels):
if self.fxn_smpl_pts == 'subset':
if self.smpl_num > 0:
num_meas_pts = self.smpl_num
else:
num_meas_pts = int(self.smpl_frac * len(features))
meas_features = features[:num_meas_pts]
meas_labels = labels[:num_meas_pts]
features = features[num_meas_pts:]
labels = labels[num_meas_pts:]
results = super().fit(features, labels)
else:
# tran standard SWAG model
results = super().fit(features, labels)
raise NotImplementedError("Not implemented...")
if self.use_cuda:
device = torch.device('cuda')
else:
device = torch.device('cpu')
mean, var, cov = self.swag_model.get_space(
torch.FloatTensor(meas_features).to(
device)) # special subspace class requires measurement points.
if self.use_cuda:
mean, cov = mean.cuda(), cov.cuda()
if self.temperature is None:
self.temperature = features.shape[0] / cov.size(0)
print('Temperature:', self.temperature)
# print(cov.size())
subspace = SubspaceModel(mean, cov)
self.ess_model = EllipticalSliceSampling(self.base,
subspace=subspace,
var=var,
loader=self.data_loader,
criterion=self.criterion,
num_samples=self.num_samples,
use_cuda=self.use_cuda,
*self.args,
**self.kwargs)
train_loss = self.ess_model.fit(scale=self.prior_std,
use_cuda=self.use_cuda,
temperature=self.temperature,
minibatch=self.mb)
print(train_loss)
def predict(self, features):
return super().predict(features, swag_model=self.ess_model)