chore: remove use of vmap in stats-pooling layer (#1706)

pyannote/audio/models/blocks/pooling.py

@@ -26,53 +26,53 @@
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
-from einops import rearrange
 
 
-class StatsPool(nn.Module):
-    """Statistics pooling
+def _pool(sequences: torch.Tensor, weights: torch.Tensor) -> torch.Tensor:
+    """Helper function to compute statistics pooling
 
-    Compute temporal mean and (unbiased) standard deviation
-    and returns their concatenation.
+    Assumes that weights are already interpolated to match the number of frames
+    in sequences and that they encode the activation of only one speaker.
 
-    Reference
-    ---------
-    https://en.wikipedia.org/wiki/Weighted_arithmetic_mean
+    Parameters
+    ----------
+    sequences : (batch, features, frames) torch.Tensor
+        Sequences of features.
+    weights : (batch, frames) torch.Tensor
+        (Already interpolated) weights.
 
+    Returns
+    -------
+    output : (batch, 2 * features) torch.Tensor
+        Concatenation of mean and (unbiased) standard deviation.
     """
 
-    def _pool(self, sequences: torch.Tensor, weights: torch.Tensor) -> torch.Tensor:
-        """Helper function to compute statistics pooling
+    weights = weights.unsqueeze(dim=1)
+    # (batch, 1, frames)
 
-        Assumes that weights are already interpolated to match the number of frames
-        in sequences and that they encode the activation of only one speaker.
+    v1 = weights.sum(dim=2) + 1e-8
+    mean = torch.sum(sequences * weights, dim=2) / v1
 
-        Parameters
-        ----------
-        sequences : (batch, features, frames) torch.Tensor
-            Sequences of features.
-        weights : (batch, frames) torch.Tensor
-            (Already interpolated) weights.
+    dx2 = torch.square(sequences - mean.unsqueeze(2))
+    v2 = torch.square(weights).sum(dim=2)
 
-        Returns
-        -------
-        output : (batch, 2 * features) torch.Tensor
-            Concatenation of mean and (unbiased) standard deviation.
-        """
+    var = torch.sum(dx2 * weights, dim=2) / (v1 - v2 / v1 + 1e-8)
+    std = torch.sqrt(var)
 
-        weights = weights.unsqueeze(dim=1)
-        # (batch, 1, frames)
+    return torch.cat([mean, std], dim=1)
 
-        v1 = weights.sum(dim=2) + 1e-8
-        mean = torch.sum(sequences * weights, dim=2) / v1
 
-        dx2 = torch.square(sequences - mean.unsqueeze(2))
-        v2 = torch.square(weights).sum(dim=2)
+class StatsPool(nn.Module):
+    """Statistics pooling
 
-        var = torch.sum(dx2 * weights, dim=2) / (v1 - v2 / v1 + 1e-8)
-        std = torch.sqrt(var)
+    Compute temporal mean and (unbiased) standard deviation
+    and returns their concatenation.
 
-        return torch.cat([mean, std], dim=1)
+    Reference
+    ---------
+    https://en.wikipedia.org/wiki/Weighted_arithmetic_mean
+
+    """
 
     def forward(
         self, sequences: torch.Tensor, weights: Optional[torch.Tensor] = None
@@ -112,17 +112,20 @@ def forward(
             has_speaker_dimension = True
 
         # interpolate weights if needed
-        _, _, num_frames = sequences.shape
-        _, _, num_weights = weights.shape
+        _, _, num_frames = sequences.size()
+        _, num_speakers, num_weights = weights.size()
         if num_frames != num_weights:
             warnings.warn(
                 f"Mismatch between frames ({num_frames}) and weights ({num_weights}) numbers."
             )
             weights = F.interpolate(weights, size=num_frames, mode="nearest")
 
-        output = rearrange(
-            torch.vmap(self._pool, in_dims=(None, 1))(sequences, weights),
-            "speakers batch features -> batch speakers features",
+        output = torch.stack(
+            [
+                _pool(sequences, weights[:, speaker, :])
+                for speaker in range(num_speakers)
+            ],
+            dim=1,
         )
 
         if not has_speaker_dimension: