def get_dset_split(dataset, split, data_dir, return_means_stds=False):
if dataset == "flights":
X_train, X_test, _, _, y_train, y_test, y_means, y_stds = load_flight(base_dir=data_dir,
k800=(split == "800k"))
trainset = Datafeed(X_train, y_train, transform=None)
testset = Datafeed(X_test, y_test, transform=None)
N_train = X_train.shape[0]
input_dim = X_train.shape[1]
output_dim = y_train.shape[1]
elif dataset in uci_names + uci_gap_names:
gap = False
if dataset in uci_gap_names:
gap = True
dataset = dataset[:-4]
X_train, X_test, _, _, y_train, y_test, y_means, y_stds = \
load_gap_UCI(base_dir=data_dir, dname=dataset, n_split=int(split), gap=gap)
trainset = Datafeed(X_train, y_train, transform=None)
testset = Datafeed(X_test, y_test, transform=None)
N_train = X_train.shape[0]
input_dim = X_train.shape[1]
output_dim = y_train.shape[1]
else:
raise Exception("Dataset not implemented yet.")
if return_means_stds:
return trainset, testset, N_train, input_dim, output_dim, y_means, y_stds
return trainset, testset, N_train, input_dim, output_dim
def __init__(self,
*args,
base_dir='nb_dir/data/',
prop_val=0.05,
k800=False,
early_stop=None,
**kwargs):
super().__init__(*args,
network=network,
width=width,
batch_size=batch_size,
**kwargs)
# setup dataset
X_train, X_test, x_means, x_stds, y_train, y_test, y_means, y_stds = load_flight(
base_dir, k800=k800)
X_train = (X_train * x_stds) + x_means
y_train = (y_train * y_stds) + y_means
# print(X_train.shape)
Ntrain = int(X_train.shape[0] * (1 - prop_val))
X_val = X_train[Ntrain:]
y_val = y_train[Ntrain:]
X_train = X_train[:Ntrain]
y_train = y_train[:Ntrain]
# print(X_train.shape)
x_means, x_stds = X_train.mean(axis=0), X_train.std(axis=0)
y_means, y_stds = y_train.mean(axis=0), y_train.std(axis=0)
x_stds[x_stds < 1e-10] = 1.
X_train = ((X_train - x_means) / x_stds)
y_train = ((y_train - y_means) / y_stds)
X_val = ((X_val - x_means) / x_stds)
y_val = ((y_val - y_means) / y_stds)
self.trainset = Datafeed(X_train, y_train, transform=None)
self.valset = Datafeed(X_val, y_val, transform=None)
self.N_train = X_train.shape[0]
self.input_dim = X_train.shape[1]
self.output_dim = y_train.shape[1]
self.early_stop = early_stop
self.regression = True
def __init__(self, *args, early_stop=None, **kwargs):
super().__init__(*args,
network=network,
width=width,
batch_size=batch_size,
**kwargs)
# setup dataset
X, y = gen_spirals(n_samples=2000,
shuffle=True,
noise=0.2,
random_state=1234,
n_arms=2,
start_angle=0,
stop_angle=720)
X_train, X_test, y_train, y_test = train_test_split(
X, y, test_size=0.8, random_state=1234)
x_means, x_stds = X_train.mean(axis=0), X_train.std(axis=0)
X_train = ((X_train - x_means) / x_stds).astype(np.float32)
X_test = ((X_test - x_means) / x_stds).astype(np.float32)
y_train = y_train.astype(np.float32)
y_test = y_test.astype(np.float32)
self.trainset = Datafeed(X_train, y_train, transform=None)
self.valset = Datafeed(X_test, y_test, transform=None)
self.N_train = X_train.shape[0]
self.input_dim = 2
self.output_dim = 2
self.early_stop = early_stop
self.regression = False
X_train, y_train, X_test, y_test = load_my_1d(args.datadir)
elif args.dataset == 'matern_1d':
X_train, y_train = load_matern_1d(args.datadir)
elif args.dataset == 'agw_1d':
X_train, y_train = load_agw_1d(args.datadir, get_feats=False)
elif args.dataset == 'andrew_1d':
X_train, y_train = load_andrew_1d(args.datadir)
elif args.dataset == 'axis':
X_train, y_train = load_axis(args.datadir)
elif args.dataset == 'origin':
X_train, y_train = load_origin(args.datadir)
elif args.dataset == 'wiggle':
X_train, y_train = load_wiggle()
trainset = Datafeed(torch.Tensor(X_train),
torch.Tensor(y_train),
transform=None)
if args.dataset == 'my_1d':
valset = Datafeed(torch.Tensor(X_test),
torch.Tensor(y_test),
transform=None)
print(X_train.shape, y_train.shape, X_test.shape, y_test.shape)
else:
valset = Datafeed(torch.Tensor(X_train),
torch.Tensor(y_train),
transform=None)
print(X_train.shape, y_train.shape)
trainloader = torch.utils.data.DataLoader(trainset,
batch_size=batch_size,
shuffle=True,
random_state=1234)
X_train, X_test, y_train, y_test = train_test_split(X,
y,
test_size=0.9,
random_state=1234)
x_means, x_stds = X_train.mean(axis=0), X_train.std(axis=0)
X_train = ((X_train - x_means) / x_stds).astype(np.float32)
X_test = ((X_test - x_means) / x_stds).astype(np.float32)
y_train = y_train.astype(np.float32)
y_test = y_test.astype(np.float32)
trainset = Datafeed(X_train, y_train, transform=None)
valset = Datafeed(X_test, y_test, transform=None)
# Declare model
for d in np.arange(0, 101):
cuda = torch.cuda.is_available()
input_dim = 2
width = 20
n_layers = d
output_dim = 2
# prior_probs = 0.85 ** (1 + np.arange(n_layers + 1))
# prior_probs = prior_probs / prior_probs.sum()
def __init__(self,
dname,
*args,
base_dir='nb_dir/data/',
prop_val=0.15,
n_split=0,
early_stop=None,
**kwargs):
super().__init__(*args,
network=network,
width=width,
batch_size=batch_size,
**kwargs)
gap = False
if dname in [
'boston', 'concrete', 'energy', 'power', 'wine', 'yacht',
'kin8nm', 'naval', 'protein'
]:
pass
elif dname in [
'boston_gap', 'concrete_gap', 'energy_gap', 'power_gap',
'wine_gap', 'yacht_gap', 'kin8nm_gap', 'naval_gap',
'protein_gap'
]:
gap = True
dname = dname[:-4]
X_train, X_test, x_means, x_stds, y_train, y_test, y_means, y_stds = \
load_gap_UCI(base_dir=base_dir, dname=dname, n_split=n_split, gap=gap)
X_train = (X_train * x_stds) + x_means
y_train = (y_train * y_stds) + y_means
# print(X_train.shape)
Ntrain = int(X_train.shape[0] * (1 - prop_val))
X_val = X_train[Ntrain:]
y_val = y_train[Ntrain:]
X_train = X_train[:Ntrain]
y_train = y_train[:Ntrain]
# print(X_train.shape)
x_means, x_stds = X_train.mean(axis=0), X_train.std(axis=0)
y_means, y_stds = y_train.mean(axis=0), y_train.std(axis=0)
x_stds[x_stds < 1e-10] = 1.
X_train = ((X_train - x_means) / x_stds)
y_train = ((y_train - y_means) / y_stds)
X_val = ((X_val - x_means) / x_stds)
y_val = ((y_val - y_means) / y_stds)
self.trainset = Datafeed(X_train, y_train, transform=None)
self.valset = Datafeed(X_val, y_val, transform=None)
self.N_train = X_train.shape[0]
self.input_dim = X_train.shape[1]
self.output_dim = y_train.shape[1]
self.early_stop = early_stop
self.regression = True