Merge pull request #6 from davidnabergoj/dev

Performance improvements and fixes

examples/Variational inference.md

@@ -0,0 +1,46 @@
+We show how to fit normalizing flows using stochastic variational inference (SVI). Whereas traditional maximum
+likelihood estimation requires a fixed dataset of samples, SVI lets us optimize NF parameters with the unnormalized
+target log density function.
+
+As an example, we define the unnormalized log density of a diagonal Gaussian. We assume this target has 10 dimensions
+with mean 5 and variance 9 in each dimension:
+
+```python
+import torch
+
+torch.manual_seed(0)
+
+event_shape = (10,)
+true_mean = torch.full(size=event_shape, fill_value=5.0)
+true_variance = torch.full(size=event_shape, fill_value=9.0)
+
+
+def target_log_prob(x: torch.Tensor):
+    return torch.sum(-((x - true_mean) ** 2 / (2 * true_variance)), dim=1)
+```
+
+We define the flow and run the variational fit:
+
+```python
+from normalizing_flows import Flow
+from normalizing_flows.bijections import RealNVP
+
+torch.manual_seed(0)
+flow = Flow(RealNVP(event_shape=event_shape))
+flow.variational_fit(target_log_prob, show_progress=True)
+```
+
+We plot samples from the trained flow. We also print estimated marginal means and variances. We see that the estimates are roughly accurate.
+```python
+import matplotlib.pyplot as plt
+
+torch.manual_seed(0)
+x_flow = flow.sample(10000).detach()
+
+plt.figure()
+plt.scatter(x_flow[:, 0], x_flow[:, 1])
+plt.show()
+
+print(f'{torch.mean(x_flow, dim=0) = }')
+print(f'{torch.var(x_flow, dim=0) = }')
+```

normalizing_flows/bijections/continuous/base.py

@@ -209,7 +209,7 @@ def forward(self, t, states):
                 s_.requires_grad_(True)
             dy = self.diffeq(t, y, *states[2:])
             divergence = self.divergence_step(dy, y)
-        return tuple([dy, -divergence] + [torch.zeros_like(s_).requires_grad_(True) for s_ in states[2:]])
+        return tuple([dy, divergence] + [torch.zeros_like(s_).requires_grad_(True) for s_ in states[2:]])
 
 
 class RegularizedApproximateODEFunction(ApproximateODEFunction):

normalizing_flows/bijections/continuous/otflow.py

@@ -1,3 +1,4 @@
+import math
 from typing import Union, Tuple
 
 import torch
@@ -29,19 +30,21 @@ def __init__(self, event_size: int, hidden_size: int, step_size: float = 0.01):
 
         divisor = max(event_size ** 2, 10)
 
-        K0_delta = torch.randn(size=(hidden_size, event_size)) / divisor
-        b0_delta = torch.randn(size=(hidden_size,)) / divisor
+        self.K0_delta = nn.Parameter(torch.randn(size=(hidden_size, event_size)) / divisor)
+        self.b0 = nn.Parameter(torch.randn(size=(hidden_size,)) / divisor)
 
-        K1_delta = torch.randn(size=(hidden_size, hidden_size)) / divisor
-        b1_delta = torch.randn(size=(hidden_size,)) / divisor
+        self.K1_delta = nn.Parameter(torch.randn(size=(hidden_size, hidden_size)) / divisor)
+        self.b1 = nn.Parameter(torch.randn(size=(hidden_size,)) / divisor)
 
-        self.K0 = nn.Parameter(torch.eye(hidden_size, event_size) + K0_delta)
-        self.b0 = nn.Parameter(0 + b0_delta)
+        self.step_size = step_size
 
-        self.K1 = nn.Parameter(torch.eye(hidden_size, hidden_size) + K1_delta)
-        self.b1 = nn.Parameter(0 + b1_delta)
+    @property
+    def K0(self):
+        return torch.eye(*self.K0_delta.shape) + self.K0_delta / 1000
 
-        self.step_size = step_size
+    @property
+    def K1(self):
+        return torch.eye(*self.K1_delta.shape) + self.K1_delta / 1000
 
     @staticmethod
     def sigma(x):
@@ -114,7 +117,7 @@ def hessian_trace(self,
 
         t0 = torch.sum(
             torch.multiply(
-                (self.sigma_prime_prime(torch.nn.functional.linear(s, self.K0, self.b0)) * z1),
+                self.sigma_prime_prime(torch.nn.functional.linear(s, self.K0, self.b0)) * z1,
                 torch.nn.functional.linear(ones, self.K0[:, :-1] ** 2)
             ),
             dim=1
@@ -138,9 +141,13 @@ def hessian_trace(self,
 
 
 class OTPotential(TimeDerivative):
-    def __init__(self, event_size: int, hidden_size: int, **kwargs):
+    def __init__(self, event_size: int, hidden_size: int = None, **kwargs):
         super().__init__()
+
         # hidden_size = m
+        if hidden_size is None:
+            hidden_size = max(math.log(event_size), 4)
+
         r = min(10, event_size)
 
         # Initialize w to 1
@@ -159,7 +166,7 @@ def __init__(self, event_size: int, hidden_size: int, **kwargs):
         self.resnet = OTResNet(event_size + 1, hidden_size, **kwargs)  # (x, t) has d+1 elements
 
     def forward(self, t, x):
-        return -self.gradient(concatenate_x_t(x, t))
+        return self.gradient(concatenate_x_t(x, t))
 
     def gradient(self, s):
         # Equation 12
@@ -187,8 +194,8 @@ def hessian_trace(self, s: torch.Tensor, u0: torch.Tensor = None, z1: torch.Tens
 
 
 class OTFlowODEFunction(ExactODEFunction):
-    def __init__(self, n_dim):
-        super().__init__(OTPotential(n_dim, hidden_size=30))
+    def __init__(self, n_dim, **kwargs):
+        super().__init__(OTPotential(n_dim, **kwargs))
 
     def compute_log_det(self, t, x):
         return self.diffeq.hessian_trace(concatenate_x_t(x, t)).view(-1, 1)  # Need an empty dim at the end

normalizing_flows/bijections/finite/autoregressive/architectures.py

@@ -180,7 +180,7 @@ def __init__(self, event_shape, n_layers: int = 2, **kwargs):
 
 
 class UMNNMAF(BijectiveComposition):
-    def __init__(self, event_shape, n_layers: int = 2, **kwargs):
+    def __init__(self, event_shape, n_layers: int = 1, **kwargs):
         if isinstance(event_shape, int):
             event_shape = (event_shape,)
         bijections = [ElementwiseAffine(event_shape=event_shape)]

normalizing_flows/bijections/finite/autoregressive/conditioner_transforms.py → normalizing_flows/bijections/finite/autoregressive/conditioning/transforms.py

@@ -1,12 +1,12 @@
 import math
-from typing import Tuple, Union, Type, List
+from typing import Tuple, Union
 
 import torch
 import torch.nn as nn
 
-from normalizing_flows.bijections.finite.autoregressive.conditioners.context import Concatenation, ContextCombiner, \
+from normalizing_flows.bijections.finite.autoregressive.conditioning.context import Concatenation, ContextCombiner, \
     Bypass
-from normalizing_flows.utils import get_batch_shape, pad_leading_dims
+from normalizing_flows.utils import get_batch_shape
 
 
 class ConditionerTransform(nn.Module):
@@ -83,12 +83,12 @@ def forward(self, x: torch.Tensor, context: torch.Tensor = None):
         else:
             if self.n_global_parameters == self.n_transformer_parameters:
                 # All transformer parameters are learned globally
-                output = torch.zeros(*batch_shape, *self.parameter_shape)
+                output = torch.zeros(*batch_shape, *self.parameter_shape, device=x.device)
                 output[..., self.global_parameter_mask] = self.global_theta_flat
                 return output
             else:
                 # Some transformer parameters are learned globally, some are predicted
-                output = torch.zeros(*batch_shape, *self.parameter_shape)
+                output = torch.zeros(*batch_shape, *self.parameter_shape, device=x.device)
                 output[..., self.global_parameter_mask] = self.global_theta_flat
                 output[..., ~self.global_parameter_mask] = self.predict_theta_flat(x, context)
                 return output

normalizing_flows/bijections/finite/autoregressive/layers.py

@@ -1,7 +1,7 @@
 import torch
 
-from normalizing_flows.bijections.finite.autoregressive.conditioner_transforms import FeedForward
-from normalizing_flows.bijections.finite.autoregressive.conditioners.coupling_masks import HalfSplit
+from normalizing_flows.bijections.finite.autoregressive.conditioning.transforms import FeedForward
+from normalizing_flows.bijections.finite.autoregressive.conditioning.coupling_masks import HalfSplit
 from normalizing_flows.bijections.finite.autoregressive.layers_base import MaskedAutoregressiveBijection, \
     InverseMaskedAutoregressiveBijection, ElementwiseBijection, CouplingBijection
 from normalizing_flows.bijections.finite.autoregressive.transformers.linear.affine import Scale, Affine, Shift
@@ -223,8 +223,8 @@ class UMNNMaskedAutoregressive(MaskedAutoregressiveBijection):
     def __init__(self,
                  event_shape: torch.Size,
                  context_shape: torch.Size = None,
-                 n_hidden_layers: int = 1,
-                 hidden_dim: int = 5,
+                 n_hidden_layers: int = None,
+                 hidden_dim: int = None,
                  **kwargs):
         transformer: ScalarTransformer = UnconstrainedMonotonicNeuralNetwork(
             event_shape=event_shape,

normalizing_flows/bijections/finite/autoregressive/layers_base.py

@@ -1,11 +1,11 @@
-from typing import Tuple, Optional, Union
+from typing import Tuple, Union
 
 import torch
+import torch.nn as nn
 
-from normalizing_flows.bijections.finite.autoregressive.conditioners.base import Conditioner, NullConditioner
-from normalizing_flows.bijections.finite.autoregressive.conditioner_transforms import ConditionerTransform, Constant, \
+from normalizing_flows.bijections.finite.autoregressive.conditioning.transforms import ConditionerTransform, \
     MADE
-from normalizing_flows.bijections.finite.autoregressive.conditioners.coupling_masks import CouplingMask
+from normalizing_flows.bijections.finite.autoregressive.conditioning.coupling_masks import CouplingMask
 from normalizing_flows.bijections.finite.autoregressive.transformers.base import TensorTransformer, ScalarTransformer
 from normalizing_flows.bijections.base import Bijection
 from normalizing_flows.utils import flatten_event, unflatten_event, get_batch_shape
@@ -14,12 +14,10 @@
 class AutoregressiveBijection(Bijection):
     def __init__(self,
                  event_shape,
-                 conditioner: Optional[Conditioner],
                  transformer: Union[TensorTransformer, ScalarTransformer],
                  conditioner_transform: ConditionerTransform,
                  **kwargs):
         super().__init__(event_shape=event_shape)
-        self.conditioner = conditioner
         self.conditioner_transform = conditioner_transform
         self.transformer = transformer
 
@@ -57,7 +55,7 @@ def __init__(self,
                  coupling_mask: CouplingMask,
                  conditioner_transform: ConditionerTransform,
                  **kwargs):
-        super().__init__(coupling_mask.event_shape, None, transformer, conditioner_transform, **kwargs)
+        super().__init__(coupling_mask.event_shape, transformer, conditioner_transform, **kwargs)
         self.coupling_mask = coupling_mask
 
         assert conditioner_transform.input_event_shape == (coupling_mask.constant_event_size,)
@@ -112,7 +110,7 @@ def __init__(self,
             context_shape=context_shape,
             **kwargs
         )
-        super().__init__(transformer.event_shape, None, transformer, conditioner_transform)
+        super().__init__(transformer.event_shape, transformer, conditioner_transform)
 
     def apply_conditioner_transformer(self, inputs, context, forward: bool = True):
         h = self.conditioner_transform(inputs, context)
@@ -160,7 +158,23 @@ class ElementwiseBijection(AutoregressiveBijection):
     def __init__(self, transformer: ScalarTransformer, fill_value: float = None):
         super().__init__(
             transformer.event_shape,
-            NullConditioner(),
             transformer,
-            Constant(transformer.event_shape, transformer.parameter_shape, fill_value=fill_value)
+            None
         )
+
+        if fill_value is None:
+            self.value = nn.Parameter(torch.randn(*transformer.parameter_shape))
+        else:
+            self.value = nn.Parameter(torch.full(size=transformer.parameter_shape, fill_value=fill_value))
+
+    def prepare_h(self, batch_shape):
+        tmp = self.value[[None] * len(batch_shape)]
+        return tmp.repeat(*batch_shape, *([1] * len(self.transformer.parameter_shape)))
+
+    def forward(self, x: torch.Tensor, context: torch.Tensor = None) -> Tuple[torch.Tensor, torch.Tensor]:
+        h = self.prepare_h(get_batch_shape(x, self.event_shape))
+        return self.transformer.forward(x, h)
+
+    def inverse(self, z: torch.Tensor, context: torch.Tensor = None) -> Tuple[torch.Tensor, torch.Tensor]:
+        h = self.prepare_h(get_batch_shape(z, self.event_shape))
+        return self.transformer.inverse(z, h)

normalizing_flows/bijections/finite/autoregressive/transformers/integration/unconstrained_monotonic_neural_network.py

@@ -32,14 +32,19 @@ class UnconstrainedMonotonicNeuralNetwork(UnconstrainedMonotonicTransformer):
     """
     def __init__(self,
                  event_shape: Union[torch.Size, Tuple[int, ...]],
-                 n_hidden_layers: int = 2,
+                 n_hidden_layers: int = None,
                  hidden_dim: int = None):
         super().__init__(event_shape, g=self.neural_network_forward, c=torch.tensor(-100.0))
+
+        if n_hidden_layers is None:
+            n_hidden_layers = 1
         self.n_hidden_layers = n_hidden_layers
+
         if hidden_dim is None:
-            hidden_dim = max(5 * int(math.log(self.n_dim)), 4)
+            hidden_dim = max(int(math.log(self.n_dim)), 4)
         self.hidden_dim = hidden_dim
-        self.const = 1000  # for stability
+
+        self.const = 1  # for stability
 
         # weight, bias have self.hidden_dim elements
         self.n_input_params = 2 * self.hidden_dim
@@ -118,10 +123,8 @@ def neural_network_forward(inputs, parameters: List[torch.Tensor]):
 
     def base_forward_1d(self, x: torch.Tensor, params: List[torch.Tensor]) -> Tuple[torch.Tensor, torch.Tensor]:
         x_r = x.view(-1, 1, 1)
-        integral_flat = self.integral(x_r, params)
-        log_det_flat = self.g(x_r, params).log()  # We can apply log since g is always positive
-        output = integral_flat.view_as(x)
-        log_det = log_det_flat.view_as(x)
+        output = self.integral(x_r, params).view_as(x)
+        log_det = self.g(x_r, params).log().view_as(x)  # We can apply log since g is always positive
         return output, log_det
 
     def inverse_1d(self, z: torch.Tensor, h: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:

normalizing_flows/bijections/finite/autoregressive/transformers/spline/linear_rational.py

@@ -83,9 +83,9 @@ def compute_knots(self, u_x, u_y, u_l, u_d):
         # u_y acts as a delta
         # u_d acts as a delta
         knots_x = self.compute_bins(u_x, self.min_input, self.max_input, self.min_bin_width)
-        knots_y = self.compute_bins(u_x + u_y / 1000, self.min_output, self.max_output, self.min_bin_height)
+        knots_y = self.compute_bins(u_x + u_y / 100, self.min_output, self.max_output, self.min_bin_height)
         knots_lambda = torch.sigmoid(u_l)
-        knots_d = self.compute_derivatives(self.const + u_d / 1000)
+        knots_d = self.compute_derivatives(self.const + u_d / 100)
         return knots_x, knots_y, knots_d, knots_lambda
 
     def forward_1d(self, x, h):

normalizing_flows/bijections/finite/autoregressive/util.py

@@ -1,5 +1,5 @@
 from functools import lru_cache
-from typing import Tuple, List
+from typing import List, Optional, Tuple
 
 import numpy as np
 import torch.autograd
@@ -10,13 +10,13 @@
 
 class GaussLegendre(torch.autograd.Function):
     @staticmethod
-    def forward(ctx, f, a: torch.Tensor, b: torch.Tensor, n: int, h: List[torch.Tensor]) -> torch.Tensor:
+    def forward(ctx, f, a: torch.Tensor, b: torch.Tensor, n: int, *h: List[torch.Tensor]) -> torch.Tensor:
         ctx.f, ctx.n = f, n
         ctx.save_for_backward(a, b, *h)
         return GaussLegendre.quadrature(f, a, b, n, h)
 
     @staticmethod
-    def backward(ctx, grad_area: torch.Tensor) -> Tuple[torch.Tensor, ...]:
+    def backward(ctx, grad_area: torch.Tensor) -> Tuple[Optional[torch.Tensor], ...]:
         f, n = ctx.f, ctx.n
         a, b, *h = ctx.saved_tensors
 
@@ -62,4 +62,4 @@ def nodes(n: int, **kwargs) -> Tuple[torch.Tensor, torch.Tensor]:
 
 
 def gauss_legendre(f, a, b, n, h):
-    return GaussLegendre.apply(f, a, b, n, h)
+    return GaussLegendre.apply(f, a, b, n, *h)