def main():
str_acq = 'ei'
num_iter = 10
X_train = np.array([
[-5],
[-1],
[1],
[2],
])
num_init = X_train.shape[0]
model_bo = bo.BO(np.array([[-6., 6.]]), str_acq=str_acq)
X_test = np.linspace(-6, 6, 400)
X_test = np.reshape(X_test, (400, 1))
for ind_ in range(1, num_iter + 1):
Y_train = fun_target(X_train)
next_x, dict_info = model_bo.optimize(X_train,
fun_target(X_train),
str_initial_method_ao='uniform')
cov_X_X = dict_info['cov_X_X']
inv_cov_X_X = dict_info['inv_cov_X_X']
hyps = dict_info['hyps']
mu_test, sigma_test, Sigma_test = gp.predict_test_(
X_train, Y_train, X_test, cov_X_X, inv_cov_X_X, hyps)
acq_test = acquisition.ei(mu_test.flatten(), sigma_test.flatten(),
Y_train)
acq_test = np.expand_dims(acq_test, axis=1)
X_train = np.vstack((X_train, next_x))
Y_train = fun_target(X_train)
utils_plotting.plot_bo_step(X_train,
Y_train,
X_test,
fun_target(X_test),
mu_test,
sigma_test,
path_save=PATH_SAVE,
str_postfix='bo_{}_'.format(str_acq) +
str(ind_),
int_init=num_init)
utils_plotting.plot_bo_step_acq(X_train,
Y_train,
X_test,
fun_target(X_test),
mu_test,
sigma_test,
acq_test,
path_save=PATH_SAVE,
str_postfix='bo_{}_'.format(str_acq) +
str(ind_),
int_init=num_init)
def main():
num_iter = 10
X_train = np.array([
[-5],
[-1],
[1],
[2],
])
num_init = X_train.shape[0]
model_bo = bo.BO(np.array([[-6., 6.]]))
X_test = np.linspace(-6, 6, 400)
X_test = np.reshape(X_test, (400, 1))
for ind_ in range(1, num_iter + 1):
Y_train = fun_target(X_train)
next_x, _, _, cov_X_X, inv_cov_X_X, hyps = model_bo.optimize(
X_train, fun_target(X_train), str_initial_method='uniform')
mu_test, sigma_test = gp.predict_test_(X_train, Y_train, X_test,
cov_X_X, inv_cov_X_X, hyps)
acq_test = acquisition.ei(mu_test.flatten(), sigma_test.flatten(),
Y_train)
acq_test = np.expand_dims(acq_test, axis=1)
X_train = np.vstack((X_train, next_x))
Y_train = fun_target(X_train)
utils_plotting.plot_bo_step(X_train,
Y_train,
X_test,
fun_target(X_test),
mu_test,
sigma_test,
path_save=PATH_SAVE,
str_postfix='basics_bo_' + str(ind_),
int_init=num_init)
utils_plotting.plot_bo_step_acq(X_train,
Y_train,
X_test,
fun_target(X_test),
mu_test,
sigma_test,
acq_test,
path_save=PATH_SAVE,
str_postfix='basics_bo_' + str(ind_),
int_init=num_init,
is_acq_axis_small=True)
def test_ei():
with pytest.raises(AssertionError) as error:
package_target.ei('abc', np.ones(10), np.zeros((5, 1)))
with pytest.raises(AssertionError) as error:
package_target.ei(np.ones(10), 'abc', np.zeros((5, 1)))
with pytest.raises(AssertionError) as error:
package_target.ei(np.ones(10), np.ones(10), 'abc')
with pytest.raises(AssertionError) as error:
package_target.ei(np.ones(10), np.ones(10), np.zeros((5, 1)), 1)
with pytest.raises(AssertionError) as error:
package_target.ei(np.ones(5), np.ones(10), np.zeros((5, 1)))
with pytest.raises(AssertionError) as error:
package_target.ei(np.ones(10), np.ones(10), np.zeros(5))
with pytest.raises(AssertionError) as error:
package_target.ei(np.ones(10), np.ones((10, 1)), np.zeros((5, 1)))
with pytest.raises(AssertionError) as error:
package_target.ei(np.ones((10, 1)), np.ones(10), np.zeros((5, 1)))
val_acq = package_target.ei(np.arange(0, 10), np.ones(10), np.zeros((5, 1)))
truth_val_acq = np.array([3.98942280e-01, 8.33154706e-02, 8.49070261e-03, 3.82154315e-04, 7.14525833e-06, 5.34616535e-08, 1.56356969e-10, 1.76032579e-13, 7.55026079e-17, 1.22477876e-20])
assert (np.abs(val_acq - truth_val_acq) < TEST_EPSILON).all()
def oracle(args, model):
seed = 42
num_all = 100
num_iter = 50
num_init = args.bo_num_init
str_cov = 'se'
list_dict = []
if args.bo_kernel == 'rbf':
kernel = RBFKernel()
elif args.bo_kernel == 'matern':
kernel = Matern52Kernel()
elif args.bo_kernel == 'periodic':
kernel = PeriodicKernel()
else:
raise ValueError(f'Invalid kernel {args.bo_kernel}')
for ind_seed in range(1, num_all + 1):
seed_ = seed * ind_seed
if seed_ is not None:
torch.manual_seed(seed_)
torch.cuda.manual_seed(seed_)
if os.path.exists(
get_str_file('./results',
args.bo_kernel,
'oracle',
args.t_noise,
seed=ind_seed)):
dict_exp = np.load(get_str_file('./results',
args.bo_kernel,
'oracle',
args.t_noise,
seed=ind_seed),
allow_pickle=True)
dict_exp = dict_exp[()]
list_dict.append(dict_exp)
print(dict_exp)
print(dict_exp['global'])
print(np.array2string(dict_exp['minima'], separator=','))
print(np.array2string(dict_exp['regrets'], separator=','))
continue
sampler = GPPriorSampler(kernel, t_noise=args.t_noise)
xp = torch.linspace(-2, 2, 1000).cuda()
xp_ = xp.unsqueeze(0).unsqueeze(2)
yp = sampler.sample(xp_)
min_yp = yp.min()
print(min_yp.cpu().numpy())
model.eval()
batch = AttrDict()
indices_permuted = torch.randperm(yp.shape[1])
batch.x = xp_[:, indices_permuted[:2 * num_init], :]
batch.y = yp[:, indices_permuted[:2 * num_init], :]
batch.xc = xp_[:, indices_permuted[:num_init], :]
batch.yc = yp[:, indices_permuted[:num_init], :]
batch.xt = xp_[:, indices_permuted[num_init:2 * num_init], :]
batch.yt = yp[:, indices_permuted[num_init:2 * num_init], :]
X_train = batch.xc.squeeze(0).cpu().numpy()
Y_train = batch.yc.squeeze(0).cpu().numpy()
X_test = xp_.squeeze(0).cpu().numpy()
list_min = []
list_min.append(batch.yc.min().cpu().numpy())
for ind_iter in range(0, num_iter):
print('ind_seed {} seed {} iter {}'.format(ind_seed, seed_,
ind_iter + 1))
cov_X_X, inv_cov_X_X, hyps = bayesogp.get_optimized_kernel(
X_train,
Y_train,
None,
str_cov,
is_fixed_noise=False,
debug=False)
prior_mu_train = bayesogp.get_prior_mu(None, X_train)
prior_mu_test = bayesogp.get_prior_mu(None, X_test)
cov_X_Xs = covariance.cov_main(str_cov, X_train, X_test, hyps,
False)
cov_Xs_Xs = covariance.cov_main(str_cov, X_test, X_test, hyps,
True)
cov_Xs_Xs = (cov_Xs_Xs + cov_Xs_Xs.T) / 2.0
mu_ = np.dot(np.dot(cov_X_Xs.T, inv_cov_X_X),
Y_train - prior_mu_train) + prior_mu_test
Sigma_ = cov_Xs_Xs - np.dot(np.dot(cov_X_Xs.T, inv_cov_X_X),
cov_X_Xs)
sigma_ = np.expand_dims(np.sqrt(np.maximum(np.diag(Sigma_), 0.0)),
axis=1)
acq_vals = -1.0 * acquisition.ei(np.ravel(mu_), np.ravel(sigma_),
Y_train)
ind_ = np.argmin(acq_vals)
x_new = xp[ind_, None, None, None]
y_new = yp[:, ind_, None, :]
batch.x = torch.cat([batch.x, x_new], axis=1)
batch.y = torch.cat([batch.y, y_new], axis=1)
batch.xc = torch.cat([batch.xc, x_new], axis=1)
batch.yc = torch.cat([batch.yc, y_new], axis=1)
X_train = batch.xc.squeeze(0).cpu().numpy()
Y_train = batch.yc.squeeze(0).cpu().numpy()
min_cur = batch.yc.min()
list_min.append(min_cur.cpu().numpy())
print(min_yp.cpu().numpy())
print(np.array2string(np.array(list_min), separator=','))
print(
np.array2string(np.array(list_min) - min_yp.cpu().numpy(),
separator=','))
dict_exp = {
'seed': seed_,
'str_cov': str_cov,
'global': min_yp.cpu().numpy(),
'minima': np.array(list_min),
'regrets': np.array(list_min) - min_yp.cpu().numpy(),
'xc': X_train,
'yc': Y_train,
'model': 'oracle',
}
np.save(
get_str_file('./results',
args.bo_kernel,
'oracle',
args.t_noise,
seed=ind_seed), dict_exp)
list_dict.append(dict_exp)
np.save(
get_str_file('./figures/results', args.bo_kernel, 'oracle',
args.t_noise), list_dict)
def bo(args, model):
if args.mode == 'bo':
ckpt = torch.load(os.path.join(args.root, 'ckpt.tar'))
model.load_state_dict(ckpt.model)
if args.bo_kernel == 'rbf':
kernel = RBFKernel()
elif args.bo_kernel == 'matern':
kernel = Matern52Kernel()
elif args.bo_kernel == 'periodic':
kernel = PeriodicKernel()
else:
raise ValueError(f'Invalid kernel {args.bo_kernel}')
seed = 42
str_cov = 'se'
num_all = 100
num_iter = 50
num_init = args.bo_num_init
list_dict = []
for ind_seed in range(1, num_all + 1):
seed_ = seed * ind_seed
if seed_ is not None:
torch.manual_seed(seed_)
torch.cuda.manual_seed(seed_)
obj_prior = GPPriorSampler(kernel, t_noise=args.t_noise)
xp = torch.linspace(-2, 2, 1000).cuda()
xp_ = xp.unsqueeze(0).unsqueeze(2)
yp = obj_prior.sample(xp_)
min_yp = yp.min()
print(min_yp.cpu().numpy())
model.eval()
batch = AttrDict()
indices_permuted = torch.randperm(yp.shape[1])
batch.x = xp_[:, indices_permuted[:2 * num_init], :]
batch.y = yp[:, indices_permuted[:2 * num_init], :]
batch.xc = xp_[:, indices_permuted[:num_init], :]
batch.yc = yp[:, indices_permuted[:num_init], :]
batch.xt = xp_[:, indices_permuted[num_init:2 * num_init], :]
batch.yt = yp[:, indices_permuted[num_init:2 * num_init], :]
X_train = batch.xc.squeeze(0).cpu().numpy()
Y_train = batch.yc.squeeze(0).cpu().numpy()
X_test = xp_.squeeze(0).cpu().numpy()
list_min = []
list_min.append(batch.yc.min().cpu().numpy())
for ind_iter in range(0, num_iter):
print('ind_seed {} seed {} iter {}'.format(ind_seed, seed_,
ind_iter + 1))
with torch.no_grad():
outs = model(batch, num_samples=args.bo_num_samples)
print('ctx_ll {:.4f} tar ll {:.4f}'.format(
outs.ctx_ll.item(), outs.tar_ll.item()))
py = model.predict(batch.xc,
batch.yc,
xp[None, :, None].repeat(1, 1, 1),
num_samples=args.bo_num_samples)
mu, sigma = py.mean.squeeze(0), py.scale.squeeze(0)
if mu.dim() == 4:
print(mu.shape, sigma.shape)
var = sigma.pow(2).mean(0) + mu.pow(2).mean(0) - mu.mean(
0).pow(2)
sigma = var.sqrt().squeeze(0)
mu = mu.mean(0).squeeze(0)
mu_ = mu.cpu().numpy()
sigma_ = sigma.cpu().numpy()
acq_vals = -1.0 * acquisition.ei(np.ravel(mu_),
np.ravel(sigma_), Y_train)
# acq_vals = []
# for ind_mu in range(0, mu.shape[0]):
# acq_vals_ = -1.0 * acquisition.ei(np.ravel(mu[ind_mu].cpu().numpy()), np.ravel(sigma[ind_mu].cpu().numpy()), Y_train)
# acq_vals.append(acq_vals_)
# acq_vals = np.mean(acq_vals, axis=0)
else:
mu_ = mu.cpu().numpy()
sigma_ = sigma.cpu().numpy()
acq_vals = -1.0 * acquisition.ei(np.ravel(mu_),
np.ravel(sigma_), Y_train)
# var = sigma.pow(2).mean(0) + mu.pow(2).mean(0) - mu.mean(0).pow(2)
# sigma = var.sqrt().squeeze(0)
# mu = mu.mean(0).squeeze(0)
ind_ = np.argmin(acq_vals)
x_new = xp[ind_, None, None, None]
y_new = yp[:, ind_, None, :]
batch.x = torch.cat([batch.x, x_new], axis=1)
batch.y = torch.cat([batch.y, y_new], axis=1)
batch.xc = torch.cat([batch.xc, x_new], axis=1)
batch.yc = torch.cat([batch.yc, y_new], axis=1)
X_train = batch.xc.squeeze(0).cpu().numpy()
Y_train = batch.yc.squeeze(0).cpu().numpy()
min_cur = batch.yc.min()
list_min.append(min_cur.cpu().numpy())
print(min_yp.cpu().numpy())
print(np.array2string(np.array(list_min), separator=','))
print(
np.array2string(np.array(list_min) - min_yp.cpu().numpy(),
separator=','))
dict_exp = {
'seed': seed_,
'global': min_yp.cpu().numpy(),
'minima': np.array(list_min),
'regrets': np.array(list_min) - min_yp.cpu().numpy(),
'xc': X_train,
'yc': Y_train,
'model': args.model,
'cov': str_cov,
}
list_dict.append(dict_exp)
np.save(
get_str_file('./figures/results', args.bo_kernel, args.model,
args.t_noise), list_dict)