Fix VarAutoEncoder reparameterize returning mu+std at inference (#8413)

monai/networks/nets/fullyconnectednet.py

@@ -103,6 +103,8 @@ class VarFullyConnectedNet(nn.Module):
         act: activation type and arguments. Defaults to PReLU.
         bias: whether to have a bias term in linear units. Defaults to True.
         adn_ordering: order of operations in :py:class:`monai.networks.blocks.ADN`.
+        use_mean_at_inference: whether to return the posterior mean (rather than ``mu + std``)
+            as the latent code during inference. Defaults to False for backward compatibility.
 
     Examples::
 
@@ -122,11 +124,13 @@ def __init__(
         act: tuple | str | None = Act.PRELU,
         bias: bool = True,
         adn_ordering: str | None = None,
+        use_mean_at_inference: bool = False,
     ) -> None:
         super().__init__()
         self.in_channels = in_channels
         self.out_channels = out_channels
         self.latent_size = latent_size
+        self.use_mean_at_inference = use_mean_at_inference
 
         self.encode = nn.Sequential()
         self.decode = nn.Sequential()
@@ -172,12 +176,27 @@ def decode_forward(self, z: torch.Tensor, use_sigmoid: bool = True) -> torch.Ten
         return x
 
     def reparameterize(self, mu: torch.Tensor, logvar: torch.Tensor) -> torch.Tensor:
-        std = torch.exp(0.5 * logvar)
+        """Sample a latent code using the reparameterization trick.
+
+        At inference (eval mode), if ``use_mean_at_inference`` is enabled, the posterior
+        mean is returned directly. Otherwise, ``mu + std`` is returned, matching the
+        original behaviour. During training, returns ``mu + eps * std`` with
+        ``eps ~ N(0, I)``.
 
-        if self.training:  # multiply random noise with std only during training
-            std = torch.randn_like(std).mul(std)
+        Args:
+            mu: Posterior mean, shape ``(batch, latent_size)``.
+            logvar: Log-variance of the posterior, same shape as ``mu``.
 
-        return std.add_(mu)
+        Returns:
+            Sampled latent code, same shape as ``mu``.
+        """
+        if not self.training and self.use_mean_at_inference:
+            # At inference the latent code is the posterior mean; the random
+            # term is only added during training (the reparameterization trick).
+            return mu
+        std = torch.exp(0.5 * logvar)
+        eps = torch.randn_like(std) if self.training else torch.ones_like(std)
+        return mu + eps * std
 
     def forward(self, x: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
         mu, logvar = self.encode_forward(x)

monai/networks/nets/varautoencoder.py

@@ -51,6 +51,8 @@ class VarAutoEncoder(AutoEncoder):
             According to `Performance Tuning Guide <https://pytorch.org/tutorials/recipes/recipes/tuning_guide.html>`_,
             if a conv layer is directly followed by a batch norm layer, bias should be False.
         use_sigmoid: whether to use the sigmoid function on final output. Defaults to True.
+        use_mean_at_inference: whether to return the posterior mean (rather than ``mu + std``)
+            as the latent code during inference. Defaults to False for backward compatibility.
 
     Examples::
 
@@ -90,9 +92,11 @@ def __init__(
         dropout: tuple | str | float | None = None,
         bias: bool = True,
         use_sigmoid: bool = True,
+        use_mean_at_inference: bool = False,
     ) -> None:
         self.in_channels, *self.in_shape = in_shape
         self.use_sigmoid = use_sigmoid
+        self.use_mean_at_inference = use_mean_at_inference
 
         self.latent_size = latent_size
         self.final_size = np.asarray(self.in_shape, dtype=int)
@@ -142,12 +146,27 @@ def decode_forward(self, z: torch.Tensor, use_sigmoid: bool = True) -> torch.Ten
         return x
 
     def reparameterize(self, mu: torch.Tensor, logvar: torch.Tensor) -> torch.Tensor:
+        """Sample a latent code using the reparameterization trick.
+
+        At inference (eval mode), if ``use_mean_at_inference`` is enabled, the posterior
+        mean is returned directly. Otherwise, ``mu + std`` is returned, matching the
+        original behaviour. During training, returns ``mu + eps * std`` with
+        ``eps ~ N(0, I)``.
+
+        Args:
+            mu: Posterior mean, shape ``(batch, latent_size)``.
+            logvar: Log-variance of the posterior, same shape as ``mu``.
+
+        Returns:
+            Sampled latent code, same shape as ``mu``.
+        """
+        if not self.training and self.use_mean_at_inference:
+            # At inference the latent code is the posterior mean; the random
+            # term is only added during training (the reparameterization trick).
+            return mu
         std = torch.exp(0.5 * logvar)
-
-        if self.training:  # multiply random noise with std only during training
-            std = torch.randn_like(std).mul(std)
-
-        return std.add_(mu)
+        eps = torch.randn_like(std) if self.training else torch.ones_like(std)
+        return mu + eps * std
 
     def forward(self, x: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
         mu, logvar = self.encode_forward(x)

tests/networks/nets/test_fullyconnectednet.py

@@ -64,6 +64,56 @@ def test_vfc_shape(self, input_param, input_shape, expected_shape):
             result = net.forward(torch.randn(input_shape).to(device))[0]
             self.assertEqual(result.shape, expected_shape)
 
+    def test_vfc_reparameterize_eval_returns_mu(self):
+        """A VFC latent code is deterministic at eval (equals mu) and stochastic at train.
+
+        Regression test for the #8413 reparameterize bug, which returned ``mu + std``
+        at inference instead of ``mu``.
+        """
+        net = VarFullyConnectedNet(
+            in_channels=10,
+            out_channels=10,
+            latent_size=30,
+            encode_channels=(15, 20, 25),
+            decode_channels=(15, 20, 25),
+            use_mean_at_inference=True,
+        ).to(device)
+        data = torch.randn(3, 10).to(device)
+
+        with eval_mode(net):
+            _, mu1, _, z1 = net(data)
+            _, _, _, z2 = net(data)
+        self.assertTrue(torch.allclose(z1, mu1))
+        self.assertTrue(torch.allclose(z1, z2))
+
+        net.train()
+        with torch.no_grad():
+            _, mu_t, _, zt1 = net(data)
+            _, _, _, zt2 = net(data)
+        self.assertFalse(torch.allclose(zt1, mu_t))
+        self.assertFalse(torch.allclose(zt1, zt2))
+
+    def test_vfc_reparameterize_default_keeps_original_behaviour(self):
+        """By default, eval returns ``mu + std`` (deterministic) for backward compatibility.
+
+        The default must preserve the pre-#8413 behaviour: at inference the standard
+        deviation is added to the mean without random noise.
+        """
+        net = VarFullyConnectedNet(
+            in_channels=10,
+            out_channels=10,
+            latent_size=30,
+            encode_channels=(15, 20, 25),
+            decode_channels=(15, 20, 25),
+        ).to(device)
+        data = torch.randn(3, 10).to(device)
+
+        with eval_mode(net):
+            _, mu1, _, z1 = net(data)
+            _, _, _, z2 = net(data)
+        self.assertFalse(torch.allclose(z1, mu1))
+        self.assertTrue(torch.allclose(z1, z2))
+
 
 if __name__ == "__main__":
     unittest.main()

tests/networks/nets/test_varautoencoder.py

@@ -122,6 +122,57 @@ def test_script(self):
         test_data = torch.randn(2, 1, 32, 32)
         test_script_save(net, test_data, rtol=1e-3, atol=1e-3)
 
+    def test_reparameterize_eval_returns_mu(self):
+        """A VarAutoEncoder latent code is deterministic at eval (equals mu) and stochastic at train.
+
+        Regression test for #8413, where eval returned ``mu + std`` instead of ``mu``.
+        """
+        net = VarAutoEncoder(
+            spatial_dims=2,
+            in_shape=(1, 32, 32),
+            out_channels=1,
+            latent_size=4,
+            channels=(4, 8),
+            strides=(2, 2),
+            use_mean_at_inference=True,
+        ).to(device)
+        data = torch.randn(2, 1, 32, 32).to(device)
+
+        with eval_mode(net):
+            _, mu1, _, z1 = net(data)
+            _, _, _, z2 = net(data)
+        self.assertTrue(torch.allclose(z1, mu1))
+        self.assertTrue(torch.allclose(z1, z2))
+
+        net.train()
+        with torch.no_grad():
+            _, mu_t, _, zt1 = net(data)
+            _, _, _, zt2 = net(data)
+        self.assertFalse(torch.allclose(zt1, mu_t))
+        self.assertFalse(torch.allclose(zt1, zt2))
+
+    def test_reparameterize_default_keeps_original_behaviour(self):
+        """By default, eval returns ``mu + std`` (deterministic) for backward compatibility.
+
+        The default must preserve the pre-#8413 behaviour: at inference the standard
+        deviation is added to the mean without random noise.
+        """
+        net = VarAutoEncoder(
+            spatial_dims=2,
+            in_shape=(1, 32, 32),
+            out_channels=1,
+            latent_size=4,
+            channels=(4, 8),
+            strides=(2, 2),
+        ).to(device)
+        data = torch.randn(2, 1, 32, 32).to(device)
+
+        with eval_mode(net):
+            _, mu1, _, z1 = net(data)
+            _, _, _, z2 = net(data)
+        self.assertFalse(torch.allclose(z1, mu1))
+        self.assertTrue(torch.allclose(z1, z2))
+
 
 if __name__ == "__main__":
     unittest.main()