def test_splitprior_logdet(input_size, distribution, n_times=10):
module = SplitPrior(input_size[1:], NegativeGaussianLoss).to('cuda')
half_c = input_size[1] // 2
nfeats = np.prod((half_c, *input_size[2:]))
x = torch.randn(input_size).to('cuda')
ldj_ours = module.logdet(x)
def func(*inputs):
inp = torch.stack(inputs, dim=0)
out, _ = module(inp)
out = out.sum(dim=0)
return out
J = torch.autograd.functional.jacobian(func,
tuple(x),
create_graph=False,
strict=False)
J = torch.stack(J, dim=0)
J = J[:, :, :, :, :half_c, :, :]
J = J.view(x.size(0), nfeats, nfeats)
logdet_pytorch = torch.slogdet(J)[1]
logdet_pytorch += module.base.log_prob(module.transform(x)[0][:, half_c:])
ldj_ours = ldj_ours.cpu().detach().numpy()
ldj_pytorch = logdet_pytorch.cpu().detach().numpy()
np.testing.assert_allclose(ldj_ours, ldj_pytorch, atol=1e-4)
def create_model(num_blocks=3,
block_size=48,
sym_recon_grad=False,
actnorm=False,
split_prior=False,
recon_loss_weight=1.0):
current_size = (3, 32, 32)
alpha = 1e-6
layers = [
Dequantization(UniformDistribution(size=current_size)),
Normalization(translation=0, scale=256),
Normalization(translation=-alpha, scale=1 / (1 - 2 * alpha)),
LogitTransform(),
]
for l in range(num_blocks):
layers.append(Squeeze())
current_size = (current_size[0] * 4, current_size[1] // 2,
current_size[2] // 2)
for k in range(block_size):
if actnorm:
layers.append(ActNorm(current_size[0]))
layers.append(Conv1x1(current_size[0]))
layers.append(Coupling(current_size))
if split_prior and l < num_blocks - 1:
layers.append(SplitPrior(current_size, NegativeGaussianLoss))
current_size = (current_size[0] // 2, current_size[1],
current_size[2])
return FlowSequential(NegativeGaussianLoss(size=current_size), *layers)
def create_model(num_blocks=2, block_size=16, sym_recon_grad=False,
actnorm=False, split_prior=False, recon_loss_weight=100.0):
alpha = 1e-6
layers = [
Dequantization(UniformDistribution(size=(1, 28, 28))),
Normalization(translation=0, scale=256),
Normalization(translation=-alpha, scale=1 / (1 - 2 * alpha)),
LogitTransform(),
]
current_size = (1, 28, 28)
for l in range(num_blocks):
layers.append(Squeeze())
current_size = (current_size[0]*4, current_size[1]//2, current_size[2]//2)
for k in range(block_size):
if actnorm:
layers.append(ActNorm(current_size[0]))
layers.append(SelfNormConv(current_size[0], current_size[0], (1, 1),
bias=True, stride=1, padding=0,
sym_recon_grad=sym_recon_grad,
recon_loss_weight=recon_loss_weight))
layers.append(Coupling(current_size))
if split_prior and l < num_blocks - 1:
layers.append(SplitPrior(current_size, NegativeGaussianLoss))
current_size = (current_size[0] // 2, current_size[1], current_size[2])
return FlowSequential(NegativeGaussianLoss(size=current_size),
*layers)
def test_splitprior_inverse(input_size, distribution, n_times=10):
module = SplitPrior(input_size[1:], NegativeGaussianLoss).to('cuda')
half_c = input_size[1] // 2
for _ in range(n_times):
input = torch.randn(input_size).to('cuda')
forward, logdet = module(input)
reverse = module.reverse(forward)
full_reverse = torch.cat([reverse[:, :half_c], input[:, half_c:]],
dim=1)
inp = input.cpu().detach().numpy()
outp = full_reverse.cpu().detach().view(input_size).numpy()
np.testing.assert_allclose(inp, outp, atol=1e-3)