def nn_experiments(X_train, Y_train, X_test, Y_test, weight_variance,
bias_variance, noise_variance):
H = defaults.hidden_units
num_layers = defaults.shared_depth
D_IN = X_train.shape[1]
D_OUT = 1
X_train_var = Variable(torch.from_numpy(X_train).type(defaults.tdtype),
requires_grad=False)
Y_train_var = Variable(torch.from_numpy(Y_train).type(defaults.tdtype),
requires_grad=False)
X_test_var = Variable(torch.from_numpy(X_test).type(defaults.tdtype),
requires_grad=False)
H = defaults.hidden_units
model = shared.get_nn_model(D_IN, H, D_OUT, num_layers, 'identity',
weight_variance, bias_variance)
burn_in = 200
nthin = 50
results_manager = shared.ResultsManager(Y_test, burn_in, nthin, True, True,
True, True, True)
num_samples = 1000000
shared.nn_model_regression(X_train_var,
Y_train_var,
X_test_var,
model,
num_samples=num_samples,
epsilon=0.0005,
beta=0.1,
leap_frog_iters=10,
noise_variance=noise_variance,
results_manager=results_manager)
return results_manager
def main():
torch.manual_seed(2)
torch.set_num_threads(1)
num_dim = 4
num_train = 10
num_test = num_train
#randomly generate X and Y
rng = np.random.RandomState(3)
X_train = rng.randn(num_train, num_dim)
X_test = rng.randn(num_test, num_dim)
#create a random network.
H = defaults.hidden_units
num_layers = defaults.shared_depth
D_IN = X_train.shape[1]
D_OUT = 1
model = shared.get_nn_model(D_IN, H, D_OUT, num_layers)
#create torch version of inputs
X_train_t = Variable(torch.from_numpy(X_train).type(defaults.tdtype),
requires_grad=False)
X_test_t = Variable(torch.from_numpy(X_test).type(defaults.tdtype),
requires_grad=False)
#create train and test data.
f_train = model(X_train_t)
f_test = model(X_test_t)
#corrupt with correct level of noise.
Y_train = (f_train + Variable(torch.randn(*f_train.size()).type(
defaults.tdtype).mul(math.sqrt(defaults.noise_variance)),
requires_grad=False)).detach()
Y_test = (f_test + Variable(torch.randn(*f_test.size()).type(
defaults.tdtype).mul(math.sqrt(defaults.noise_variance)),
requires_grad=False)).detach()
densities_gp = gp_experiments(X_train, Y_train.data.numpy(), X_test,
Y_test.data.numpy(), num_layers)
#print('hold_out_gp ', hold_out_gp )
results = nn_experiments(X_train_t, Y_train, X_test_t, Y_test, H,
num_layers)
results['log_densities_GP'] = densities_gp
results['X_train'] = X_train
results['X_test'] = X_test
results['Y_train'] = Y_train
results['Y_test'] = Y_train
pickle.dump(results, open(results_file_name, 'wb'))
embed()
plt.show()
def nn_experiments(X,Y,grid_points):
H = defaults.hidden_units
num_layers = defaults.shared_depth
D_IN = X.shape[1]
D_OUT = 1
X_var = Variable( torch.from_numpy(X).type(defaults.tdtype), requires_grad=False)
Y_var = Variable( torch.from_numpy(Y).type(defaults.tdtype), requires_grad=False)
grid_var = Variable( torch.from_numpy(grid_points).type(defaults.tdtype), requires_grad=False)
model = shared.get_nn_model(D_IN,H,D_OUT, num_layers)
#model.cuda()
#3000
nthin = 1
burn_in = 50
results_manager = shared.ResultsManager(None, burn_in, nthin, False, False, True, False, True)
shared.nn_model_regression(X_var,Y_var,grid_var, model, num_samples = 30000, epsilon = 0.05, beta = 1., leap_frog_iters=10, results_manager = results_manager )
return results_manager
