def __init__(self, pretrained_model, latent_size):
self.flow_shape = pretrained_model.base_dist.shape
self.latent_size = latent_size
# initialize transforms with first scale of pretrained models
transforms = pretrained_model.transforms[0:28]
# Replace slice layer with a compression flow
mencoder = ConditionalNormal(ConvEncoderNet(
in_channels=48,
out_channels=768,
mid_channels=[64, 128, 256],
max_pool=True,
batch_norm=True),
split_dim=1)
mdecoder = ConditionalNormal(ConvDecoderNet(
in_channels=768,
out_shape=(48 * 2, 8, 8),
mid_channels=[256, 128, 64],
batch_norm=True,
in_lambda=lambda x: x.view(x.shape[0], x.shape[1], 1, 1)),
split_dim=1)
mid_vae = VAE(encoder=mencoder, decoder=mdecoder)
reshape = Reshape(input_shape=(768, ), output_shape=(12, 8, 8))
transforms.extend([mid_vae, reshape])
# Non-dimension preserving flows
current_shape = pretrained_model.base_dist.shape
flat_dim = current_shape[0] * current_shape[1] * current_shape[2]
fencoder = ConditionalNormal(
MLP(flat_dim,
2 * latent_size,
hidden_units=[512, 256],
activation='relu',
in_lambda=lambda x: x.view(x.shape[0], flat_dim)))
fdecoder = ConditionalNormal(MLP(
latent_size,
2 * flat_dim,
hidden_units=[256, 512],
activation='relu',
out_lambda=lambda x: x.view(x.shape[0], current_shape[0] * 2,
current_shape[1], current_shape[2])),
split_dim=1)
# append last scale of pretrained model and extend with the compressive VAE
transforms.extend(pretrained_model.transforms[29:])
final_vae = VAE(encoder=fencoder, decoder=fdecoder)
transforms.append(final_vae)
# Base distribution for non-dimension preserving portion of flow
base1 = StandardNormal((latent_size, ))
super(CompressPretrained, self).__init__(base_dist=[None, base1],
transforms=transforms)
def test_stochastic_transform_is_well_behaved(self):
batch_size = 8
data_size = 10
latent_size = 2
x = torch.randn(batch_size, data_size)
encoder = ConditionalNormal(nn.Linear(data_size,2*latent_size))
decoder = ConditionalNormal(nn.Linear(latent_size,2*data_size))
transform = VAE(decoder=decoder, encoder=encoder)
self.assert_stochastic_transform_is_well_behaved(transform, x, z_shape=(batch_size, latent_size), z_dtype=torch.float)
#model = NDPFlow(base_dist=[StandardNormal((16,7,7)), StandardNormal((latent_size,))],
model = NDPFlow(base_dist=[None, StandardNormal((latent_size,))],
transforms=[
UniformDequantization(num_bits=8),
ActNormBijection2d(1),
ScalarAffineBijection(scale=2.0),
ScalarAffineBijection(shift=-0.5),
Squeeze2d(),
ActNormBijection2d(4), Conv1x1(4), AffineCouplingBijection(net(4)),
ActNormBijection2d(4), Conv1x1(4), AffineCouplingBijection(net(4)),
Squeeze2d(),
ActNormBijection2d(16), Conv1x1(16), AffineCouplingBijection(net(16)),
ActNormBijection2d(16), Conv1x1(16), AffineCouplingBijection(net(16)),
VAE(encoder=encoder, decoder=decoder)
]).to(device)
print(model)
###########
## Optim ##
###########
optimizer = Adam(model.parameters(), lr=1e-3)
###########
## Train ##
###########
print('Training...')
def __init__(self, data_shape, num_bits,
base_distribution, num_scales, num_steps, actnorm,
vae_hidden_units,
coupling_network,
dequant, dequant_steps, dequant_context,
coupling_blocks, coupling_channels, coupling_dropout,
coupling_gated_conv=None, coupling_depth=None, coupling_mixtures=None):
assert len(base_distribution) == 1, "Only a single base distribution is supported"
transforms = []
current_shape = data_shape
if num_steps == 0: num_scales = 0
if dequant == 'uniform' or num_steps == 0 or num_scales == 0:
# no bijective flows defaults to only using uniform dequantization
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=coupling_blocks,
mid_channels=coupling_channels,
depth=coupling_depth,
dropout=coupling_dropout,
gated_conv=coupling_gated_conv,
num_mixtures=coupling_mixtures)
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))
for scale in range(num_scales):
# squeeze to exchange height and width for more channels
transforms.append(Squeeze2d())
current_shape = (current_shape[0] * 4,
current_shape[1] // 2,
current_shape[2] // 2)
# Dimension preserving components
for step in range(num_steps):
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))
else:
transforms.append(
MixtureCoupling(in_channels=current_shape[0],
mid_channels=coupling_channels,
num_mixtures=coupling_mixtures,
num_blocks=coupling_blocks,
dropout=coupling_dropout))
# Non-dimension preserving flows: reduce the dimensionality of data by 2 (channel-wise)
if actnorm: transforms.append(ActNormBijection2d(current_shape[0]))
assert current_shape[0] % 2 == 0, f"Current shape {current_shape[1]}x{current_shape[2]} must be divisible by two"
latent_size = (current_shape[0] * current_shape[1] * current_shape[2]) // 2
encoder = ConditionalNormal(
ConvEncoderNet(in_channels=current_shape[0],
out_channels=latent_size,
mid_channels=vae_hidden_units,
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_hidden_units)),
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))
current_shape = (current_shape[0] // 2,
current_shape[1],
current_shape[2])
if scale < num_scales - 1:
# reshape latent sample to have height and width
transforms.append(Reshape(input_shape=(latent_size,), output_shape=current_shape))
# Base distribution for dimension preserving portion of flow
if base_distribution == "n":
base_dist = StandardNormal((latent_size,))
elif base_distribution == "c":
base_dist = ConvNormal2d((latent_size,))
elif base_distribution == "u":
base_dist = StandardUniform((latent_size,))
else:
raise ValueError("Base distribution must be one of n=Normal, u=Uniform, or c=ConvNormal")
# for reference save the shape output by the bijective flow
self.latent_size = latent_size
self.flow_shape = current_shape
super(MultilevelCompressiveFlow, self).__init__(base_dist=[None, base_dist], transforms=transforms)
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)
def __init__(self,
data_shape,
num_bits,
base_distributions,
num_scales,
num_steps,
actnorm,
vae_hidden_units,
latent_size,
vae_activation,
coupling_network,
dequant,
dequant_steps,
dequant_context,
coupling_blocks,
coupling_channels,
coupling_dropout,
coupling_growth=None,
coupling_gated_conv=None,
coupling_depth=None,
coupling_mixtures=None):
transforms = []
current_shape = data_shape
if num_steps == 0: num_scales = 0
if dequant == 'uniform' or num_steps == 0 or num_scales == 0:
# no bijective flows defaults to only using uniform dequantization
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=coupling_blocks,
mid_channels=coupling_channels,
depth=coupling_depth,
growth=coupling_growth,
dropout=coupling_dropout,
gated_conv=coupling_gated_conv,
num_mixtures=coupling_mixtures)
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 squeeze
transforms.append(Squeeze2d())
current_shape = (current_shape[0] * 4, current_shape[1] // 2,
current_shape[2] // 2)
# Dimension preserving flows
for scale in range(num_scales):
for step in range(num_steps):
if actnorm:
transforms.append(ActNormBijection2d(current_shape[0]))
transforms.append(Conv1x1(current_shape[0]))
if coupling_network in ["conv", "densenet"]:
transforms.append(
Coupling(in_channels=current_shape[0],
num_blocks=coupling_blocks,
mid_channels=coupling_channels,
depth=coupling_depth,
growth=coupling_growth,
dropout=coupling_dropout,
gated_conv=coupling_gated_conv,
coupling_network=coupling_network))
else:
transforms.append(
MixtureCoupling(in_channels=current_shape[0],
mid_channels=coupling_channels,
num_mixtures=coupling_mixtures,
num_blocks=coupling_blocks,
dropout=coupling_dropout))
if scale < num_scales - 1:
transforms.append(Squeeze2d())
current_shape = (current_shape[0] * 4, current_shape[1] // 2,
current_shape[2] // 2)
else:
if actnorm:
transforms.append(ActNormBijection2d(current_shape[0]))
# Base distribution for dimension preserving portion of flow
if len(base_distributions) > 1:
if base_distributions[0] == "n":
base0 = StandardNormal(current_shape)
elif base_distributions[0] == "c":
base0 = ConvNormal2d(current_shape)
elif base_distributions[0] == "u":
base0 = StandardUniform(current_shape)
else:
raise ValueError(
"Base distribution must be one of n=Noraml, u=Uniform, or c=ConvNormal"
)
else:
base0 = None
# for reference save the shape output by the bijective flow
self.flow_shape = current_shape
# Non-dimension preserving flows
flat_dim = current_shape[0] * current_shape[1] * current_shape[2]
encoder = ConditionalNormal(
MLP(flat_dim,
2 * latent_size,
hidden_units=vae_hidden_units,
activation=vae_activation,
in_lambda=lambda x: x.view(x.shape[0], flat_dim)))
decoder = ConditionalNormal(MLP(
latent_size,
2 * flat_dim,
hidden_units=list(reversed(vae_hidden_units)),
activation=vae_activation,
out_lambda=lambda x: x.view(x.shape[0], current_shape[0] * 2,
current_shape[1], current_shape[2])),
split_dim=1)
transforms.append(VAE(encoder=encoder, decoder=decoder))
# Base distribution for non-dimension preserving portion of flow
#self.latent_size = latent_size
if base_distributions[-1] == "n":
base1 = StandardNormal((latent_size, ))
elif base_distributions[-1] == "c":
base1 = ConvNormal2d((latent_size, ))
elif base_distributions[-1] == "u":
base1 = StandardUniform((latent_size, ))
else:
raise ValueError(
"Base distribution must be one of n=Noraml, u=Uniform, or c=ConvNormal"
)
super(VAECompressiveFlow, self).__init__(base_dist=[base0, base1],
transforms=transforms)
in_lambda=lambda x: 2 * x.view(x.shape[0], 784).float() - 1))
decoder = ConditionalBernoulli(MLP(latent_sizes[0], 784,
hidden_units=[512,256],
activation='relu',
out_lambda=lambda x: x.view(x.shape[0], 1, 28, 28)))
encoder2 = ConditionalNormal(MLP(latent_sizes[0], 2*latent_sizes[1],
hidden_units=[256,128],
activation='relu'))
decoder2 = ConditionalNormal(MLP(latent_sizes[1], 2*latent_sizes[0],
hidden_units=[256,128],
activation='relu'))
model = Flow(base_dist=StandardNormal((latent_sizes[-1],)),
transforms=[
VAE(encoder=encoder, decoder=decoder),
VAE(encoder=encoder2, decoder=decoder2),
]).to(device)
###########
## Optim ##
###########
optimizer = Adam(model.parameters(), lr=1e-3)
###########
## Train ##
###########
print('Training...')
for epoch in range(20):
encoder = ConditionalNormal(
MLP(784,
2 * latent_size,
hidden_units=[512, 256],
activation='relu',
in_lambda=lambda x: 2 * x.view(x.shape[0], 784).float() - 1))
decoder = ConditionalBernoulli(
MLP(latent_size,
784,
hidden_units=[512, 256],
activation='relu',
out_lambda=lambda x: x.view(x.shape[0], 1, 28, 28)))
model = Flow(base_dist=StandardNormal((latent_size, )),
transforms=[VAE(encoder=encoder, decoder=decoder)]).to(device)
###########
## Optim ##
###########
optimizer = Adam(model.parameters(), lr=1e-3)
###########
## Train ##
###########
print('Training...')
for epoch in range(15):
l = 0.0
for i, x in enumerate(train_loader):