Merge pull request #79 from Genentech/multinomial_axis

add multiomial axis option to poisson multinomial loss

src/grelu/lightning/losses.py

@@ -17,6 +17,8 @@ class PoissonMultinomialLoss(nn.Module):
         log_input: If True, the input is transformed with torch.exp to produce predicted
             counts. Otherwise, the input is assumed to already represent predicted
             counts.
+        multinomial_axis: Either "length" or "task", representing the axis along which the
+            multinomial distribution should be calculated.
         reduction: "mean" or "none".
     """
 
@@ -26,12 +28,17 @@ def __init__(
         eps: float = 1e-7,
         log_input: bool = True,
         reduction: str = "mean",
+        multinomial_axis: str = "length",
     ) -> None:
         super().__init__()
         self.eps = eps
         self.total_weight = total_weight
         self.log_input = log_input
         self.reduction = reduction
+        if multinomial_axis == "length":
+            self.axis = 2
+        elif multinomial_axis == "task":
+            self.axis = 1
 
     def forward(self, input: Tensor, target: Tensor) -> Tensor:
         """
@@ -54,8 +61,8 @@ def forward(self, input: Tensor, target: Tensor) -> Tensor:
             input += self.eps
 
         # Assuming count predictions
-        total_target = target.sum(axis=-1, keepdim=True)
-        total_input = input.sum(axis=-1, keepdim=True)
+        total_target = target.sum(axis=self.axis, keepdim=True)
+        total_input = input.sum(axis=self.axis, keepdim=True)
 
         # total count poisson loss, mean across targets
         poisson_term = F.poisson_nll_loss(
@@ -68,11 +75,11 @@ def forward(self, input: Tensor, target: Tensor) -> Tensor:
         log_p_input = torch.log(p_input)
 
         # multinomial loss
-        multinomial_dot = -torch.multiply(target, log_p_input)  # B x T x L
-        multinomial_term = multinomial_dot.mean(axis=-1, keepdim=True)  # B x T
+        multinomial_dot = -torch.multiply(target, log_p_input)
+        multinomial_term = multinomial_dot.mean(axis=self.axis, keepdim=True)
 
         # Combine
-        loss = multinomial_term + self.total_weight * poisson_term
+        loss = multinomial_term + (self.total_weight * poisson_term)
 
         if self.reduction == "mean":
             return loss.mean()