def test_forward_inverse_are_consistent(self):
features = 100
batch_size = 50
inputs = torch.randn(batch_size, features)
transforms = [
norm.BatchNorm(features=features, affine=affine)
for affine in [True, False]
]
self.eps = 1e-6
for transform in transforms:
with self.subTest(transform=transform):
transform.eval()
self.assert_forward_inverse_are_consistent(transform, inputs)
def test_forward(self):
features = 100
batch_size = 50
bn_eps = 1e-5
self.eps = 1e-4
for affine in [True, True]:
with self.subTest(affine=affine):
inputs = torch.randn(batch_size, features)
transform = norm.BatchNorm(features=features,
affine=affine,
eps=bn_eps)
outputs, logabsdet = transform(inputs)
self.assert_tensor_is_good(outputs, [batch_size, features])
self.assert_tensor_is_good(logabsdet, [batch_size])
mean, var = inputs.mean(0), inputs.var(0)
outputs_ref = (inputs - mean) / torch.sqrt(var + bn_eps)
logabsdet_ref = torch.sum(
torch.log(1.0 / torch.sqrt(var + bn_eps)))
logabsdet_ref = torch.full([batch_size], logabsdet_ref.item())
if affine:
outputs_ref *= transform.weight
outputs_ref += transform.bias
logabsdet_ref += torch.sum(torch.log(transform.weight))
self.assert_tensor_is_good(outputs_ref, [batch_size, features])
self.assert_tensor_is_good(logabsdet_ref, [batch_size])
print(outputs, outputs_ref)
self.assertEqual(outputs, outputs_ref)
self.assertEqual(logabsdet, logabsdet_ref)
transform.eval()
outputs, logabsdet = transform(inputs)
self.assert_tensor_is_good(outputs, [batch_size, features])
self.assert_tensor_is_good(logabsdet, [batch_size])
mean = transform.running_mean
var = transform.running_var
outputs_ref = (inputs - mean) / torch.sqrt(var + bn_eps)
logabsdet_ref = torch.sum(
torch.log(1.0 / torch.sqrt(var + bn_eps)))
logabsdet_ref = torch.full([batch_size], logabsdet_ref.item())
if affine:
outputs_ref *= transform.weight
outputs_ref += transform.bias
logabsdet_ref += torch.sum(torch.log(transform.weight))
self.assert_tensor_is_good(outputs_ref, [batch_size, features])
self.assert_tensor_is_good(logabsdet_ref, [batch_size])
self.assertEqual(outputs, outputs_ref)
self.assertEqual(logabsdet, logabsdet_ref)
def test_inverse(self):
features = 100
batch_size = 50
inputs = torch.randn(batch_size, features)
for affine in [True, False]:
with self.subTest(affine=affine):
transform = norm.BatchNorm(features=features, affine=affine)
with self.assertRaises(base.InverseNotAvailable):
transform.inverse(inputs)
transform.eval()
outputs, logabsdet = transform.inverse(inputs)
self.assert_tensor_is_good(outputs, [batch_size, features])
self.assert_tensor_is_good(logabsdet, [batch_size])