def run(data_name, num_train, num_test, Phi,
depth, widths, lc_w_range, shift_w_range,
optim_name, optim_args,
num_epochs, batch_size, chkpt_freq):
id = get_info()
identifier_id = '%s%s' % (identifier, id)
train_data, test_data = load_data(data_name, num_train, num_test)
train_ll, test_ll = load_log_ll(data_name, num_train, num_test)
print('Computing ground truth manually because tagged log likelihood is wrong')
from phi_listing import ClaytonPhi
gt_phi = ClaytonPhi(torch.tensor(5.))
cop = Copula(gt_phi)
train_ll = -torch.log(cop(train_data, mode='pdf'))
test_ll = -torch.log(cop(test_data, mode='pdf'))
print('train_ll', torch.mean(train_ll))
print('test_ll', torch.mean(test_ll))
print('Train ideal ll:', torch.mean(train_ll))
print('Test ideal ll:', torch.mean(test_ll))
phi = Phi(depth, widths, lc_w_range, shift_w_range)
net = Copula(phi)
expt(train_data, test_data, net, optim_name,
optim_args, identifier_id, num_epochs, batch_size, chkpt_freq)
def run(x_index, y_index,
x_flip, y_flip,
Phi,
depth, widths, lc_w_range, shift_w_range,
optim_name, optim_args,
num_epochs, batch_size, chkpt_freq,
frac_rand):
id = get_info()
identifier_id = '%s%s' % (identifier, id)
train_data = X_train[:, [x_index, y_index]]
train_data = add_train_random_noise(train_data,
int(train_data.shape[0]*frac_rand))
test_data = X_test[:, [x_index, y_index]]
if x_flip:
train_data[:, 0] = 1-train_data[:, 0]
test_data[:, 0] = 1-test_data[:, 0]
if y_flip:
train_data[:, 1] = 1-train_data[:, 1]
test_data[:, 1] = 1-test_data[:, 1]
phi = Phi(depth, widths, lc_w_range, shift_w_range)
net = Copula(phi)
expt(train_data, test_data, net, optim_name,
optim_args, identifier_id, num_epochs, batch_size, chkpt_freq)
def gen_data(phi,
ndims,
N,
seed):
torch.set_default_tensor_type(torch.DoubleTensor)
net = Copula(phi)
s = sample(net, ndims, N, seed=seed)
log_ll = -torch.log(net(s, 'pdf'))
print('mean log_ll:', torch.mean(log_ll))
plot_samples(s)
return s, log_ll
def run(Phi, initial_theta, optim_name, optim_args, num_epochs, batch_size,
chkpt_freq, frac_rand):
id = get_info()
identifier_id = '%s%s' % (identifier, id)
train_data = X_train
train_data = add_train_random_noise(train_data,
int(X_train.shape[0] * frac_rand))
test_data = X_test
phi = Phi(torch.tensor(initial_theta))
net = Copula(phi)
expt(train_data, test_data, net, optim_name, optim_args, identifier_id,
num_epochs, batch_size, chkpt_freq)
def run(Phi, depth, widths, lc_w_range, shift_w_range, optim_name, optim_args,
num_epochs, batch_size, chkpt_freq, frac_rand):
id = get_info()
identifier_id = '%s%s' % (identifier, id)
train_data = X_train
train_data = add_train_random_noise(train_data,
int(X_train.shape[0] * frac_rand))
test_data = X_test
phi = Phi(depth, widths, lc_w_range, shift_w_range)
net = Copula(phi)
expt(train_data, test_data, net, optim_name, optim_args, identifier_id,
num_epochs, batch_size, chkpt_freq)
def run(data_name, num_train, num_test, Phi, depth, widths, lc_w_range,
shift_w_range, optim_name, optim_args, num_epochs, batch_size,
chkpt_freq):
id = get_info()
identifier_id = '%s%s' % (identifier, id)
train_data, test_data = load_data(data_name, num_train, num_test)
train_ll, test_ll = load_log_ll(data_name, num_train, num_test)
print('Train ideal ll:', torch.mean(train_ll))
print('Test ideal ll:', torch.mean(test_ll))
phi = Phi(depth, widths, lc_w_range, shift_w_range)
net = Copula(phi)
expt(train_data, test_data, net, optim_name, optim_args, identifier_id,
num_epochs, batch_size, chkpt_freq)
