Add sample batching to avoid memory issues.

harmonic/evidence.py

@@ -212,7 +212,7 @@ def get_masks(self, chain_start_ixs: jnp.ndarray) -> jnp.ndarray:
 
         return masks_arr
 
-    def add_chains(self, chains):
+    def add_chains(self, chains, num_slices=None):
         """Add new chains and calculate an estimate of the inverse evidence, its
         variance, and the variance of the variance.
 
@@ -228,6 +228,10 @@ def add_chains(self, chains):
             chains (Chains): An instance of the chains class containing the chains to
                 be used in the calculation.
 
+            num_slices (int): Number of slices into which the samples are divided row-wise
+                when using flow models to avoid memory issues. If None, the samples are
+                considered all-together. Defaults to None.
+
         Raises:
 
             ValueError: Raised if the input number of chains to not match the
@@ -247,10 +251,36 @@ def add_chains(self, chains):
         Y = chains.ln_posterior
         nchains = self.nchains
 
+        if not num_slices is None:
+            if num_slices > X.shape[0]:
+                raise ValueError(
+                    "Can't split chains into more blocks than there are samples."
+                )
+
         if self.batch_calculation:
-            lnpred = self.model.predict(x=X)
-            lnargs = lnpred - Y
-            lnargs = lnargs.at[jnp.isinf(lnargs)].set(jnp.nan)
+            if num_slices:
+                # Number of rows in each slice
+                slice_size = X.shape[0] // num_slices
+                lnpred_list = []
+
+                # Calculate lnpred in row-wise slices
+                for i in range(num_slices):
+                    start_row = i * slice_size
+                    end_row = (i + 1) * slice_size if i < num_slices - 1 else X.shape[0]
+                    X_slice = X[start_row:end_row]
+
+                    # Predict for each row slice and append result
+                    lnpred_slice = self.model.predict(x=X_slice)
+                    lnpred_list.append(lnpred_slice)
+
+                # Concatenate all row slice predictions
+                lnpred = jnp.concatenate(lnpred_list, axis=0)
+                lnargs = lnpred - Y
+                lnargs = lnargs.at[jnp.isinf(lnargs)].set(jnp.nan)
+            else:
+                lnpred = self.model.predict(x=X)
+                lnargs = lnpred - Y
+                lnargs = lnargs.at[jnp.isinf(lnargs)].set(jnp.nan)
 
         else:
             lnpred = np.zeros_like(Y)