From 4650263390b9a2ddaf6b254e54e665061f5b44d4 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Fran=C3=A7ois=20Rozet?= <francois.rozet@outlook.com>
Date: Tue, 30 Aug 2022 13:37:10 +0200
Subject: [PATCH] =?UTF-8?q?=E2=9C=A8=20Implement=20balanced=20NRE=20(BNRE)?=
 =?UTF-8?q?=20loss=20(#3)?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 lampe/inference.py      | 57 +++++++++++++++++++++++++++++++++++++++++
 tests/test_inference.py | 12 +++++++++
 2 files changed, 69 insertions(+)

diff --git a/lampe/inference.py b/lampe/inference.py
index d567c98..96b3bae 100644
--- a/lampe/inference.py
+++ b/lampe/inference.py
@@ -139,6 +139,63 @@ def forward(self, theta: Tensor, x: Tensor) -> Tensor:
         return l1 + l0
 
 
+class BNRELoss(nn.Module):
+    r"""Creates a module that calculates the loss :math:`l` of a balanced NRE (BNRE)
+    classifier :math:`d_\phi`. Given a batch of :math:`N` pairs
+    :math:`\{ (\theta_i, x_i) \}`, the module returns
+
+    .. math::
+        \begin{align}
+            l & = \frac{1}{N} \sum_{i = 1}^N
+            \ell(d_\phi(\theta_i, x_i)) + \ell(1 - d_\phi(\theta_{i+1}, x_i)) \\
+              & + \lambda \left(1 - \frac{1}{N} \sum_{i = 1}^N
+                d_\phi(\theta_i, x_i) + d_\phi(\theta_{i+1}, x_i)
+            \right)^2
+        \end{align}
+
+    where :math:`\ell(p) = - \log p` is the negative log-likelihood.
+
+    References:
+        Towards Reliable Simulation-Based Inference with Balanced Neural Ratio Estimation
+        (Delaunoy et al., 2022)
+        https://arxiv.org/abs/2208.13624
+
+    Arguments:
+        estimator: A classifier network :math:`d_\phi(\theta, x)`.
+        lmbda: The weight :math:`\lambda` controlling the strength of the balancing
+            condition.
+    """
+
+    def __init__(self, estimator: nn.Module, lmbda: float = 100.0):
+        super().__init__()
+
+        self.estimator = estimator
+        self.lmbda = lmbda
+
+    def forward(self, theta: Tensor, x: Tensor) -> Tensor:
+        r"""
+        Arguments:
+            theta: The parameters :math:`\theta`, with shape :math:`(N, D)`.
+            x: The observation :math:`x`, with shape :math:`(N, L)`.
+
+        Returns:
+            The scalar loss :math:`l`.
+        """
+
+        theta_prime = torch.roll(theta, 1, dims=0)
+
+        log_r, log_r_prime = self.estimator(
+            torch.stack((theta, theta_prime)),
+            x,
+        )
+
+        l1 = -F.logsigmoid(log_r).mean()
+        l0 = -F.logsigmoid(-log_r_prime).mean()
+        lb = (torch.sigmoid(log_r) + torch.sigmoid(log_r_prime) - 1).mean().square()
+
+        return l1 + l0 + self.lmbda * lb
+
+
 class AMNRE(NRE):
     r"""Creates an arbitrary marginal neural ratio estimation (AMNRE) classifier
     network.
diff --git a/tests/test_inference.py b/tests/test_inference.py
index c6034f0..e9fd1e5 100644
--- a/tests/test_inference.py
+++ b/tests/test_inference.py
@@ -43,6 +43,18 @@ def test_NRELoss():
     assert l.requires_grad
 
 
+def test_BNRELoss():
+    estimator = NRE(3, 5)
+    loss = BNRELoss(estimator)
+
+    theta, x = randn(256, 3), randn(256, 5)
+
+    l = loss(theta, x)
+
+    assert l.shape == ()
+    assert l.requires_grad
+
+
 def test_AMNRE():
     estimator = AMNRE(3, 5)