def get_full_data(dataset, scaling=False, n_train=None):
data_path = os.path.dirname(os.path.abspath(__file__))
data_cfg = dict(scaling=scaling)
if n_train is not None:
data_cfg.update(n_train=n_train)
file_path = os.path.join(
data_path, 'datasets', dataset,
var_to_str(data_cfg) + '_full_data') + format_files
if os.path.exists(file_path):
full_data = load(open(file_path, 'rb'))
else:
train_data, test_data = get_data(dataset, scaling=scaling)
train_loader = DataLoader(train_data, batch_size=128)
test_loader = DataLoader(test_data, batch_size=128)
full_data = dict(train=dict(X=[], y=[]), test=dict(X=[], y=[]))
for datasplit, loader in zip(full_data.values(),
[train_loader, test_loader]):
for x, y in loader:
datasplit['X'].append(x)
datasplit['y'].append(y)
datasplit['X'] = torch.cat(datasplit['X'])
datasplit['y'] = torch.cat(datasplit['y'])
if n_train is not None:
full_data['train']['X'] = full_data['train']['X'][:n_train]
full_data['train']['y'] = full_data['train']['y'][:n_train]
full_data.update(n_class=n_class_datasets[dataset])
save(full_data, open(file_path, 'wb'))
return full_data
def get_hyper_params_ref(dataset, scaling=True):
subsample = 5
dataset_path = os.path.join(os.path.dirname(os.path.abspath(__file__)),
'datasets', dataset)
data_cfg = dict(scaling=scaling)
file_path = os.path.join(
dataset_path,
var_to_str(data_cfg) + '_hyper_params_ref') + format_files
if os.path.exists(file_path):
hyperparams_ref = load(open(file_path, 'rb'))
else:
full_data = get_full_data(dataset, scaling=scaling)
X_train = full_data['train']['X']
X_train = X_train[::subsample]
del full_data
cov = (X_train.t().mm(X_train) / len(X_train)).numpy()
lambda_max = np.real(splinalg.eigs(cov, 1)[0])
reg_ref = (lambda_max / np.trace(cov)).item()
reg_ref = float('{:1.2e}'.format(reg_ref))
print(f'reg ref: {reg_ref}')
del cov
dists = torch.cdist(X_train, X_train)**2
del X_train
dists = torch.triu(dists)
sigma_ref = torch.sqrt(torch.median(dists[dists != 0]) / 2).item()
sigma_ref = float('{:1.2e}'.format(sigma_ref))
print(f'sigma ref: {sigma_ref}')
os.makedirs(os.path.dirname(file_path), exist_ok=True)
hyperparams_ref = dict(reg_ref=reg_ref, sigma_ref=sigma_ref)
save(hyperparams_ref, open(file_path, 'wb'))
return hyperparams_ref
def get_reg_ref_rbf(dataset, scaling, nb_train, sigma):
subsample = 5
dataset_path = os.path.join(os.path.dirname(os.path.abspath(__file__)),
'datasets', dataset)
data_cfg = dict(scaling=scaling, nb_train=nb_train, sigma=sigma)
file_path = os.path.join(
dataset_path,
var_to_str(data_cfg) + '_reg_rbf_ref') + format_files
if os.path.exists(file_path):
reg_ref_rbf = load(open(file_path, 'rb'))
else:
gram_data = preprocess_gram(dataset,
scaling=scaling,
n_train=nb_train,
kernel='rbf',
sigma=sigma)
gram_train = gram_data['train']['gram']
del gram_data
gram_train = gram_train[::subsample, ::subsample] / gram_train.size(0)
lambda_max = np.real(splinalg.eigs(gram_train, 1)[0])
reg_ref_rbf = (lambda_max / np.trace(gram_train)).item()
reg_ref_rbf = float('{:1.2e}'.format(reg_ref_rbf))
print(f'reg_ref_rbf: {reg_ref_rbf}')
os.makedirs(os.path.dirname(file_path), exist_ok=True)
save(reg_ref_rbf, open(file_path, 'wb'))
return reg_ref_rbf
def preprocess_gram(dataset,
scaling,
n_train=None,
kernel='linear',
reg=None,
sigma=None):
data_path = os.path.dirname(os.path.abspath(__file__))
data_cfg = dict(scaling=scaling, kernel=kernel)
if reg is not None:
data_cfg.update(reg=reg)
if sigma is not None:
data_cfg.update(sigma=sigma)
if n_train is not None:
data_cfg.update(n_train=n_train)
file_path = os.path.join(data_path, 'datasets', dataset,
var_to_str(data_cfg) + 'gram_data') + format_files
if os.path.exists(file_path):
gram_data = load(open(file_path, 'rb'))
else:
full_data = get_full_data(dataset, scaling, n_train)
gram_train = compute_gram(full_data['train']['X'],
full_data['train']['X'], kernel, reg, sigma)
gram_test = compute_gram(full_data['test']['X'],
full_data['train']['X'], kernel, reg, sigma)
gram_data = dict(train=dict(gram=gram_train,
y=full_data['train']['y']),
test=dict(gram=gram_test, y=full_data['test']['y']),
n_class=full_data['n_class'])
save(gram_data, open(file_path, 'wb'))
return gram_data
def wrapped_exp_core(data_cfg, optim_cfg):
exp_cfg = dict(data_cfg=data_cfg, optim_cfg=optim_cfg)
print(*['{0}:{1}'.format(key, value) for key, value in exp_cfg.items()],
sep='\n')
exp_path = os.path.dirname(os.path.abspath(__file__))
file_path = os.path.join(exp_path, 'results', var_to_str(data_cfg),
var_to_str(optim_cfg)) + format_files
if os.path.exists(file_path):
_, test_acc = load(open(file_path, 'rb'))
else:
test_acc = exp_core(**data_cfg, **optim_cfg)
os.makedirs(os.path.dirname(file_path), exist_ok=True)
save([exp_cfg, test_acc], open(file_path, 'wb'))
result = {**data_cfg, **optim_cfg}
result.update(test_acc=test_acc.item())
result = DataFrame(result, index=[0])
return result