Remove unneeded get_reduction_factor function

test/test_sum_hessian.py

@@ -83,48 +83,13 @@
 
 def autograd_sum_hessian(layer, input, targets):
     """Compute the Hessian of a loss module w.r.t. its input."""
-
-    def get_reduction_factor(layer, input, targets):
-        """Determine reduction factor of individual losses.
-
-        Take the first sample and its label, clone it N times,
-        compare the individual loss with the loss of that batch.
-
-        For an `[N, D]` input:
-
-        CrossEntropyLoss(reduction="mean") -> N
-        CrossEntropyLoss(reduction="sum") -> 1
-        MSELoss(reduction="mean") -> N * D
-        MSELoss(reduction="sum") -> 1
-        """
-        N = input.shape[0]
-
-        sample = input[0].unsqueeze(0)
-        repeat = [N] + [1 for _ in input.shape[1:]]
-        sample_repeated = sample.repeat(*repeat)
-
-        target = targets[0].unsqueeze(0)
-        repeat = [N] + [1 for _ in targets.shape[1:]]
-        target_repeated = target.repeat(*repeat)
-
-        sample_loss = layer(sample, target)
-        batch_loss = layer(sample_repeated, target_repeated)
-
-        factor = torch.round(N * sample_loss / batch_loss).item()
-        print("Factor", factor)
-
-        return factor
-
-    factor = get_reduction_factor(layer, input, targets)
-    factor = 1
-
     input.requires_grad = True
     loss = layer(input, targets)
     hessian = autograd_hessian(loss, input)
 
     sum_hessian = sum_hessian_blocks(hessian, input)
 
-    return sum_hessian / factor
+    return sum_hessian
 
 
 def sum_hessian_blocks(hessian, input):