use Pareto distribution for Level 1 Problem 99

With inputs sampled from Unif(0,1), the Huber Loss is effectively MSE, which we know can be hacked via statistical properties of the loss fn/inputs.

We use the Pareto distribution to sample inputs w/finite mean and infinite variance to prevent hacking this way.

Author: bkal01

KernelBench/level1/96_HuberLoss.py

@@ -1,6 +1,8 @@
 import torch
 import torch.nn as nn
 
+from torch.distributions import Pareto
+
 class Model(nn.Module):
     """
     A model that computes Smooth L1 (Huber) Loss for regression tasks.
@@ -20,7 +22,9 @@ def forward(self, predictions, targets):
 
 def get_inputs():
     scale = torch.rand(())
-    return [torch.rand(batch_size, *input_shape)*scale, torch.rand(batch_size, *input_shape)]
+    predictions = Pareto(0.01, 0.15).sample((batch_size, *input_shape))
+    targets = Pareto(0.01, 0.15).sample((batch_size, *input_shape))
+    return [predictions*scale, targets]
 
 def get_init_inputs():
     return []