Improve handling of zero division in mod_c and intdiv_c

* In python, when computing x % 0, we get ZeroDivisionError("division by zero"). mod_c should thus probably behave the same when t2 is the zero vector.
* Change ValueError to ZeroDivisionError in intdiv_c
* Make mod_c reuse intdiv_c. This makes it raise ZeroDivisionError when called with t2=0.

Author: ValerianRey

src/torchjd/sparse/_linalg.py

@@ -37,17 +37,12 @@ def mod_c(t1: Tensor, t2: Tensor) -> Tensor:
         [8, 12]^T %c [2, 3]^T = [0, 0]^T
         [8, 12]^T %c [2, 4]^T = [2, 0]^T
         [8, 12]^T %c [3, 3]^T = [2, 6]^T
-        [8, 12]^T %c [0, 0]^T = [8, 12]^T
         [8, 12]^T %c [2, 0]^T = [0, 12]^T
         [8, 12]^T %c [0, 2]^T = [8, 0]^T
+        [8, 12]^T %c [0, 0]^T => ZeroDivisionError
     """
 
-    non_zero_indices = torch.nonzero(t2)
-    if len(non_zero_indices) == 0:
-        return t1
-    else:
-        min_divider = (t1[non_zero_indices] // t2[non_zero_indices]).min()
-        return t1 - min_divider * t2
+    return t1 - intdiv_c(t1, t2) * t2
 
 
 def intdiv_c(t1: Tensor, t2: Tensor) -> Tensor:
@@ -63,14 +58,14 @@ def intdiv_c(t1: Tensor, t2: Tensor) -> Tensor:
         [8, 12]^T //c [2, 3]^T = 4
         [8, 12]^T //c [2, 4]^T = 3
         [8, 12]^T //c [3, 3]^T = 2
-        [8, 12]^T //c [0, 0]^T => ValueError
         [8, 12]^T //c [2, 0]^T = 4
         [8, 12]^T //c [0, 2]^T = 6
+        [8, 12]^T //c [0, 0]^T => ZeroDivisionError
     """
 
     non_zero_indices = torch.nonzero(t2)
     if len(non_zero_indices) == 0:
-        raise ValueError("Cannot divide by the zero vector.")
+        raise ZeroDivisionError("division by zero")
     else:
         min_divider = (t1[non_zero_indices] // t2[non_zero_indices]).min()
         return min_divider

tests/unit/sparse/test_structured_sparse_tensor.py

@@ -429,7 +429,6 @@ def test_fix_zero_stride_columns(
         (tensor([8, 12]), tensor([2, 3]), tensor([0, 0])),
         (tensor([8, 12]), tensor([2, 4]), tensor([2, 0])),
         (tensor([8, 12]), tensor([3, 3]), tensor([2, 6])),
-        (tensor([8, 12]), tensor([0, 0]), tensor([8, 12])),
         (tensor([8, 12]), tensor([2, 0]), tensor([0, 12])),
         (tensor([8, 12]), tensor([0, 2]), tensor([8, 0])),
     ],
@@ -442,6 +441,11 @@ def test_mod_c(
     assert torch.equal(mod_c(t1, t2), expected)
 
 
+def test_mod_c_by_0_raises():
+    with raises(ZeroDivisionError):
+        mod_c(tensor([3, 4]), tensor([0, 0]))
+
+
 @mark.parametrize(
     ["t1", "t2", "expected"],
     [
@@ -461,5 +465,5 @@ def test_intdiv_c(
 
 
 def test_intdiv_c_by_0_raises():
-    with raises(ValueError):
+    with raises(ZeroDivisionError):
         intdiv_c(tensor([3, 4]), tensor([0, 0]))