def test_surjection_is_well_behaved(self):
batch_size = 10
shape = [8, 4, 4]
x = torch.randn(batch_size, *shape)
surjections = [
(Augment(StandardNormal([1, 4, 4]), x_size=8,
split_dim=1), (9, 4, 4)),
(Augment(StandardNormal([4, 4, 4]), x_size=8,
split_dim=1), (12, 4, 4)),
(Augment(StandardNormal([7, 4, 4]), x_size=8,
split_dim=1), (15, 4, 4)),
(Augment(StandardNormal([8, 2, 4]), x_size=4,
split_dim=2), (8, 6, 4)),
(Augment(StandardNormal([8, 4, 2]), x_size=4,
split_dim=3), (8, 4, 6)),
(Augment(ConditionalMeanNormal(nn.Conv2d(8, 1, kernel_size=1)),
x_size=8,
split_dim=1), (9, 4, 4)),
(Augment(ConditionalMeanNormal(nn.Conv2d(8, 4, kernel_size=1)),
x_size=8,
split_dim=1), (12, 4, 4)),
(Augment(ConditionalMeanNormal(nn.Conv2d(8, 7, kernel_size=1)),
x_size=8,
split_dim=1), (15, 4, 4))
]
for surjection, new_shape in surjections:
with self.subTest(surjection=surjection):
self.assert_surjection_is_well_behaved(surjection,
x,
z_shape=(batch_size,
*new_shape),
z_dtype=x.dtype)
def __init__(self, num_flows, actnorm, affine, scale_fn_str, hidden_units,
activation, range_flow, augment_size, base_dist, cond_size):
D = 2 # Number of data dimensions
A = D + augment_size # Number of augmented data dimensions
P = 2 if affine else 1 # Number of elementwise parameters
# initialize context. Only upsample context in ContextInit if latent shape doesn't change during the flow.
context_init = MLP(input_size=cond_size,
output_size=D,
hidden_units=hidden_units,
activation=activation)
# initialize flow with either augmentation or Abs surjection
if augment_size > 0:
assert augment_size % 2 == 0
transforms = [Augment(StandardNormal((augment_size, )), x_size=D)]
else:
transforms = []
transforms = [SimpleAbsSurjection()]
if range_flow == 'logit':
transforms += [
ScaleBijection(scale=torch.tensor([[1 / 4, 1 / 4]])),
Logit()
]
elif range_flow == 'softplus':
transforms += [SoftplusInverse()]
# apply coupling layer flows
for _ in range(num_flows):
net = nn.Sequential(
MLP(A // 2 + D,
P * A // 2,
hidden_units=hidden_units,
activation=activation), ElementwiseParams(P))
if affine:
transforms.append(
ConditionalAffineCouplingBijection(
net, scale_fn=scale_fn(scale_fn_str)))
else:
transforms.append(ConditionalAdditiveCouplingBijection(net))
if actnorm: transforms.append(ActNormBijection(D))
transforms.append(Reverse(A))
transforms.pop()
if base_dist == "uniform":
base = StandardUniform((A, ))
else:
base = StandardNormal((A, ))
super(SRFlow, self).__init__(base_dist=base,
transforms=transforms,
context_init=context_init)
return nn.Sequential(
DenseNet(in_channels=channels // 2,
out_channels=channels,
num_blocks=1,
mid_channels=64,
depth=8,
growth=16,
dropout=0.0,
gated_conv=True,
zero_init=True), ElementwiseParams2d(2))
model = Flow(base_dist=StandardNormal((24, 8, 8)),
transforms=[
UniformDequantization(num_bits=8),
Augment(StandardUniform((3, 32, 32)), x_size=3),
AffineCouplingBijection(net(6)),
ActNormBijection2d(6),
Conv1x1(6),
AffineCouplingBijection(net(6)),
ActNormBijection2d(6),
Conv1x1(6),
AffineCouplingBijection(net(6)),
ActNormBijection2d(6),
Conv1x1(6),
AffineCouplingBijection(net(6)),
ActNormBijection2d(6),
Conv1x1(6),
Squeeze2d(),
Slice(StandardNormal((12, 16, 16)), num_keep=12),
AffineCouplingBijection(net(12)),
def __init__(self,
data_shape,
cond_shape,
num_bits,
num_scales,
num_steps,
actnorm,
conditional_channels,
lowres_encoder_channels,
lowres_encoder_blocks,
lowres_encoder_depth,
lowres_upsampler_channels,
pooling,
compression_ratio,
coupling_network,
coupling_blocks,
coupling_channels,
coupling_dropout=0.0,
coupling_gated_conv=None,
coupling_depth=None,
coupling_mixtures=None,
dequant="flow",
dequant_steps=4,
dequant_context=32,
dequant_blocks=2,
augment_steps=4,
augment_context=32,
augment_blocks=2,
augment_size=None,
checkerboard_scales=[],
tuple_flip=True):
if len(compression_ratio) == 1 and num_scales > 1:
compression_ratio = [compression_ratio[0]] * (num_scales - 1)
assert all([
compression_ratio[s] >= 0.0 and compression_ratio[s] < 1.0
for s in range(num_scales - 1)
])
# initialize context. Only upsample context in ContextInit if latent shape doesn't change during the flow.
context_init = ContextInit(num_bits=num_bits,
in_channels=cond_shape[0],
out_channels=lowres_encoder_channels,
mid_channels=lowres_encoder_channels,
num_blocks=lowres_encoder_blocks,
depth=lowres_encoder_depth,
dropout=coupling_dropout)
transforms = []
current_shape = data_shape
if dequant == 'uniform':
transforms.append(UniformDequantization(num_bits=num_bits))
elif dequant == 'flow':
dequantize_flow = DequantizationFlow(
data_shape=data_shape,
num_bits=num_bits,
num_steps=dequant_steps,
coupling_network=coupling_network,
num_context=dequant_context,
num_blocks=dequant_blocks,
mid_channels=coupling_channels,
depth=coupling_depth,
dropout=0.0,
gated_conv=False,
num_mixtures=coupling_mixtures,
checkerboard=True,
tuple_flip=tuple_flip)
transforms.append(
VariationalDequantization(encoder=dequantize_flow,
num_bits=num_bits))
# Change range from [0,1]^D to [-0.5, 0.5]^D
transforms.append(ScalarAffineBijection(shift=-0.5))
# Initial squeezing
if current_shape[1] >= 128 and current_shape[2] >= 128:
# H x W -> 64 x 64
transforms.append(Squeeze2d())
current_shape = (current_shape[0] * 4, current_shape[1] // 2,
current_shape[2] // 2)
if current_shape[1] >= 64 and current_shape[2] >= 64:
# H x W -> 32 x 32
transforms.append(Squeeze2d())
current_shape = (current_shape[0] * 4, current_shape[1] // 2,
current_shape[2] // 2)
if 0 not in checkerboard_scales or (current_shape[1] > 32
and current_shape[2] > 32):
# Only go to 16 x 16 if not doing checkerboard splits first
transforms.append(Squeeze2d())
current_shape = (current_shape[0] * 4, current_shape[1] // 2,
current_shape[2] // 2)
# add in augmentation channels if desired
if augment_size is not None and augment_size > 0:
#transforms.append(Augment(StandardUniform((augment_size, current_shape[1], current_shape[2])), x_size=current_shape[0]))
#transforms.append(Augment(StandardNormal((augment_size, current_shape[1], current_shape[2])), x_size=current_shape[0]))
augment_flow = AugmentFlow(data_shape=current_shape,
augment_size=augment_size,
num_steps=augment_steps,
coupling_network=coupling_network,
mid_channels=coupling_channels,
num_context=augment_context,
num_mixtures=coupling_mixtures,
num_blocks=augment_blocks,
dropout=0.0,
checkerboard=True,
tuple_flip=tuple_flip)
transforms.append(
Augment(encoder=augment_flow, x_size=current_shape[0]))
current_shape = (current_shape[0] + augment_size, current_shape[1],
current_shape[2])
for scale in range(num_scales):
# First and Third scales use checkerboard split pattern
checkerboard = scale in checkerboard_scales
context_out_channels = min(current_shape[0], coupling_channels)
context_out_shape = (context_out_channels, current_shape[1],
current_shape[2] //
2) if checkerboard else (context_out_channels,
current_shape[1],
current_shape[2])
# reshape the context to the current size for all flow steps at this scale
context_upsampler_net = UpsamplerNet(
in_channels=lowres_encoder_channels,
out_shape=context_out_shape,
mid_channels=lowres_upsampler_channels)
transforms.append(
ContextUpsampler(context_net=context_upsampler_net,
direction='forward'))
for step in range(num_steps):
flip = (step % 2 == 0) if tuple_flip else False
if len(conditional_channels) == 0:
if actnorm:
transforms.append(ActNormBijection2d(current_shape[0]))
transforms.append(Conv1x1(current_shape[0]))
else:
if actnorm:
transforms.append(
ConditionalActNormBijection2d(
cond_shape=current_shape,
out_channels=current_shape[0],
mid_channels=conditional_channels))
transforms.append(
ConditionalConv1x1(cond_shape=current_shape,
out_channels=current_shape[0],
mid_channels=conditional_channels,
slogdet_cpu=True))
if coupling_network in ["conv", "densenet"]:
transforms.append(
SRCoupling(x_size=context_out_shape,
y_size=current_shape,
mid_channels=coupling_channels,
depth=coupling_depth,
num_blocks=coupling_blocks,
dropout=coupling_dropout,
gated_conv=coupling_gated_conv,
coupling_network=coupling_network,
checkerboard=checkerboard,
flip=flip))
elif coupling_network == "transformer":
transforms.append(
SRMixtureCoupling(x_size=context_out_shape,
y_size=current_shape,
mid_channels=coupling_channels,
dropout=coupling_dropout,
num_blocks=coupling_blocks,
num_mixtures=coupling_mixtures,
checkerboard=checkerboard,
flip=flip))
# Upsample context (for the previous flows, only if moving in the inverse direction)
transforms.append(
ContextUpsampler(context_net=context_upsampler_net,
direction='inverse'))
if scale < num_scales - 1:
if pooling == 'none' or compression_ratio[scale] == 0.0:
# fully bijective flow with multi-scale architecture
transforms.append(Squeeze2d())
current_shape = (current_shape[0] * 4,
current_shape[1] // 2,
current_shape[2] // 2)
elif pooling == 'slice':
# slice some of the dimensions (channel-wise) out from further flow steps
unsliced_channels = int(
max(
1, 4 * current_shape[0] *
(1.0 - compression_ratio[scale])))
sliced_channels = int(4 * current_shape[0] -
unsliced_channels)
noise_shape = (sliced_channels, current_shape[1] // 2,
current_shape[2] // 2)
transforms.append(Squeeze2d())
transforms.append(
Slice(StandardNormal(noise_shape),
num_keep=unsliced_channels,
dim=1))
current_shape = (unsliced_channels, current_shape[1] // 2,
current_shape[2] // 2)
elif pooling == 'max':
# max pooling to compress dimensions spatially, h//2 and w//2
noise_shape = (current_shape[0] * 3, current_shape[1] // 2,
current_shape[2] // 2)
decoder = StandardHalfNormal(noise_shape)
transforms.append(
SimpleMaxPoolSurjection2d(decoder=decoder))
current_shape = (current_shape[0], current_shape[1] // 2,
current_shape[2] // 2)
elif pooling == "mvae":
# Compressive flow: reduce the dimensionality of data by 2 (channel-wise)
compressed_channels = max(
1,
int(current_shape[0] *
(1.0 - compression_ratio[scale])))
latent_size = compressed_channels * current_shape[
1] * current_shape[2]
vae_channels = [
current_shape[0] * 2, current_shape[0] * 4,
current_shape[0] * 8
]
encoder = ConditionalNormal(ConvEncoderNet(
in_channels=current_shape[0],
out_channels=latent_size,
mid_channels=vae_channels,
max_pool=True,
batch_norm=True),
split_dim=1)
decoder = ConditionalNormal(ConvDecoderNet(
in_channels=latent_size,
out_shape=(current_shape[0] * 2, current_shape[1],
current_shape[2]),
mid_channels=list(reversed(vae_channels)),
batch_norm=True,
in_lambda=lambda x: x.view(x.shape[0], x.shape[1], 1, 1
)),
split_dim=1)
transforms.append(VAE(encoder=encoder, decoder=decoder))
transforms.append(
Reshape(input_shape=(latent_size, ),
output_shape=(compressed_channels,
current_shape[1],
current_shape[2])))
# after reducing channels with mvae, squeeze to reshape latent space before another sequence of flows
transforms.append(Squeeze2d())
current_shape = (
compressed_channels * 4, # current_shape[0] * 4
current_shape[1] // 2,
current_shape[2] // 2)
else:
raise ValueError(
"pooling argument must be either mvae, slice, max, or none"
)
else:
if actnorm:
transforms.append(ActNormBijection2d(current_shape[0]))
# for reference save the shape output by the bijective flow
self.latent_size = current_shape[0] * current_shape[1] * current_shape[
2]
self.flow_shape = current_shape
super(SRPoolFlow, self).__init__(base_dist=ConvNormal2d(current_shape),
transforms=transforms,
context_init=context_init)
## Specify data ##
##################
train_loader, test_loader = get_data(args)
###################
## Specify model ##
###################
assert args.augdim % 2 == 0
D = 2 # Number of data dimensions
A = 2 + args.augdim # Number of augmented data dimensions
P = 2 if args.affine else 1 # Number of elementwise parameters
transforms = [Augment(StandardNormal((args.augdim, )), x_size=D)]
for _ in range(args.num_flows):
net = nn.Sequential(
MLP(A // 2,
P * A // 2,
hidden_units=args.hidden_units,
activation=args.activation), ElementwiseParams(P))
if args.affine:
transforms.append(
AffineCouplingBijection(net, scale_fn=scale_fn(args.scale_fn)))
else:
transforms.append(AdditiveCouplingBijection(net))
if args.actnorm: transforms.append(ActNormBijection(D))
transforms.append(Reverse(A))
transforms.pop()
def __init__(self, num_flows, actnorm, affine, scale_fn_str, hidden_units,
activation, range_flow, augment_size, base_dist):
D = 2 # Number of data dimensions
if base_dist == "uniform":
classifier = MLP(D,
D // 2,
hidden_units=hidden_units,
activation=activation,
out_lambda=lambda x: x.view(-1))
transforms = [
ElementAbsSurjection(classifier=classifier),
ShiftBijection(shift=torch.tensor([[0.0, 4.0]])),
ScaleBijection(scale=torch.tensor([[1 / 4, 1 / 8]]))
]
base = StandardUniform((D, ))
else:
A = D + augment_size # Number of augmented data dimensions
P = 2 if affine else 1 # Number of elementwise parameters
# initialize flow with either augmentation or Abs surjection
if augment_size > 0:
assert augment_size % 2 == 0
transforms = [
Augment(StandardNormal((augment_size, )), x_size=D)
]
else:
transforms = [SimpleAbsSurjection()]
if range_flow == 'logit':
transforms += [
ScaleBijection(scale=torch.tensor([[1 / 4, 1 / 4]])),
Logit()
]
elif range_flow == 'softplus':
transforms += [SoftplusInverse()]
# apply coupling layer flows
for _ in range(num_flows):
net = nn.Sequential(
MLP(A // 2,
P * A // 2,
hidden_units=hidden_units,
activation=activation), ElementwiseParams(P))
if affine:
transforms.append(
AffineCouplingBijection(
net, scale_fn=scale_fn(scale_fn_str)))
else:
transforms.append(AdditiveCouplingBijection(net))
if actnorm: transforms.append(ActNormBijection(D))
transforms.append(Reverse(A))
transforms.pop()
base = StandardNormal((A, ))
super(UnconditionalFlow, self).__init__(base_dist=base,
transforms=transforms)
def reduction_layer(channels, items):
return [
*perm_norm_bi(channels),
*perm_norm_bi(channels),
*perm_norm_bi(channels),
Squeeze2d(4),
Slice(StandardNormal((channels * 2, items)), num_keep=channels * 2),
]
model = Flow(
base_dist=StandardNormal((base_channels * (2**5), n_items // (4**4))),
transforms=[
UniformDequantization(num_bits=8),
Augment(StandardUniform((base_channels * 1, n_items)),
x_size=base_channels),
*reduction_layer(base_channels * (2**1), n_items // (4**1)),
*reduction_layer(base_channels * (2**2), n_items // (4**2)),
*reduction_layer(base_channels * (2**3), n_items // (4**3)),
*reduction_layer(base_channels * (2**4), n_items // (4**4)),
# *reduction_layer(base_channels*(2**5), n_items//(4**4)),
*perm_norm_bi(base_channels * (2**5))
# AffineCouplingBijection(net(base_channels*2)), ActNormBijection2d(base_channels*2), Conv1x1(base_channels*2),
# AffineCouplingBijection(net(base_channels*2)), ActNormBijection2d(base_channels*2), Conv1x1(base_channels*2),
# AffineCouplingBijection(net(base_channels*2)), ActNormBijection2d(base_channels*2), Conv1x1(base_channels*2),
# Squeeze2d(), Slice(StandardNormal((base_channels*2, n_items//4)), num_keep=base_channels*2),
# AffineCouplingBijection(net(base_channels*2)), ActNormBijection2d(base_channels*2), Conv1x1(base_channels*2),
# AffineCouplingBijection(net(base_channels*2)), ActNormBijection2d(base_channels*2), Conv1x1(base_channels*2),
# AffineCouplingBijection(net(base_channels*2)), ActNormBijection2d(base_channels*2), Conv1x1(base_channels*2),
]).to(device)
def __init__(self,
data_shape,
num_bits,
num_scales,
num_steps,
actnorm,
pooling,
compression_ratio,
coupling_network,
coupling_blocks,
coupling_channels,
coupling_dropout=0.0,
coupling_gated_conv=None,
coupling_depth=None,
coupling_mixtures=None,
dequant="flow",
dequant_steps=4,
dequant_context=32,
dequant_blocks=2,
augment_steps=4,
augment_context=32,
augment_blocks=2,
augment_size=None,
checkerboard_scales=[],
tuple_flip=True):
if len(compression_ratio) == 1 and num_scales > 1:
compression_ratio = [compression_ratio[0]] * (num_scales - 1)
assert all([
compression_ratio[s] >= 0 and compression_ratio[s] < 1
for s in range(num_scales - 1)
])
transforms = []
current_shape = data_shape
if dequant == 'uniform':
transforms.append(UniformDequantization(num_bits=num_bits))
elif dequant == 'flow':
dequantize_flow = DequantizationFlow(
data_shape=data_shape,
num_bits=num_bits,
num_steps=dequant_steps,
coupling_network=coupling_network,
num_context=dequant_context,
num_blocks=dequant_blocks,
mid_channels=coupling_channels,
depth=coupling_depth,
dropout=0.0,
gated_conv=False,
num_mixtures=coupling_mixtures,
checkerboard=True,
tuple_flip=tuple_flip)
transforms.append(
VariationalDequantization(encoder=dequantize_flow,
num_bits=num_bits))
# Change range from [0,1]^D to [-0.5, 0.5]^D
transforms.append(ScalarAffineBijection(shift=-0.5))
# Initial squeezing
if current_shape[1] >= 128 and current_shape[2] >= 128:
# H x W -> 64 x 64
transforms.append(Squeeze2d())
current_shape = (current_shape[0] * 4, current_shape[1] // 2,
current_shape[2] // 2)
if current_shape[1] >= 64 and current_shape[2] >= 64:
# H x W -> 32 x 32
transforms.append(Squeeze2d())
current_shape = (current_shape[0] * 4, current_shape[1] // 2,
current_shape[2] // 2)
if 0 not in checkerboard_scales or (current_shape[1] > 32
and current_shape[2] > 32):
# Only go to 16 x 16 if not doing checkerboard splits first
transforms.append(Squeeze2d())
current_shape = (current_shape[0] * 4, current_shape[1] // 2,
current_shape[2] // 2)
# add in augmentation channels if desired
if augment_size is not None and augment_size > 0:
augment_flow = AugmentFlow(data_shape=current_shape,
augment_size=augment_size,
num_steps=augment_steps,
coupling_network=coupling_network,
mid_channels=coupling_channels,
num_context=augment_context,
num_mixtures=coupling_mixtures,
num_blocks=augment_blocks,
dropout=0.0,
checkerboard=True,
tuple_flip=tuple_flip)
transforms.append(
Augment(encoder=augment_flow, x_size=current_shape[0]))
current_shape = (current_shape[0] + augment_size, current_shape[1],
current_shape[2])
for scale in range(num_scales):
checkerboard = scale in checkerboard_scales
for step in range(num_steps):
flip = (step % 2 == 0) if tuple_flip else False
if actnorm:
transforms.append(ActNormBijection2d(current_shape[0]))
transforms.append(Conv1x1(current_shape[0]))
if coupling_network == "conv":
transforms.append(
Coupling(in_channels=current_shape[0],
num_blocks=coupling_blocks,
mid_channels=coupling_channels,
depth=coupling_depth,
dropout=coupling_dropout,
gated_conv=coupling_gated_conv,
coupling_network=coupling_network,
checkerboard=checkerboard,
flip=flip))
else:
transforms.append(
MixtureCoupling(in_channels=current_shape[0],
mid_channels=coupling_channels,
num_mixtures=coupling_mixtures,
num_blocks=coupling_blocks,
dropout=coupling_dropout,
checkerboard=checkerboard,
flip=flip))
if scale < num_scales - 1:
if pooling in ['bijective', 'none'
] or compression_ratio[scale] == 0.0:
transforms.append(Squeeze2d())
current_shape = (current_shape[0] * 4,
current_shape[1] // 2,
current_shape[2] // 2)
elif pooling == 'slice':
# slice some of the dimensions (channel-wise) out from further flow steps
unsliced_channels = int(
max(1, 4 * current_shape[0] *
(1.0 - sliced_ratio[scale])))
sliced_channels = int(4 * current_shape[0] -
unsliced_channels)
noise_shape = (sliced_channels, current_shape[1] // 2,
current_shape[2] // 2)
transforms.append(Squeeze2d())
transforms.append(
Slice(StandardNormal(noise_shape),
num_keep=unsliced_channels,
dim=1))
current_shape = (unsliced_channels, current_shape[1] // 2,
current_shape[2] // 2)
elif pooling == 'max':
noise_shape = (current_shape[0] * 3, current_shape[1] // 2,
current_shape[2] // 2)
decoder = StandardHalfNormal(noise_shape)
transforms.append(
SimpleMaxPoolSurjection2d(decoder=decoder))
current_shape = (current_shape[0], current_shape[1] // 2,
current_shape[2] // 2)
else:
raise ValueError(
f"Pooling argument must be either slice, max or none, not: {pooling}"
)
else:
if actnorm:
transforms.append(ActNormBijection2d(current_shape[0]))
# for reference save the shape output by the bijective flow
self.flow_shape = current_shape
self.latent_size = current_shape[0] * current_shape[1] * current_shape[
2]
super(PoolFlow, self).__init__(base_dist=ConvNormal2d(current_shape),
transforms=transforms)