diff --git a/tests/conftest.py b/tests/conftest.py
index ae4035d..57836b4 100644
--- a/tests/conftest.py
+++ b/tests/conftest.py
@@ -27,7 +27,7 @@ def rng(request):
 @pytest.fixture(scope='session', params=[
     *product(
         [Linear],
-        prodict(in_features=[16], out_features=[16], bias=[True, False]),
+        prodict(in_features=[16], out_features=[15], bias=[True, False]),
     ),
     *product(
         [Conv1d, Conv2d, Conv3d, ConvTranspose1d, ConvTranspose2d, ConvTranspose3d],
@@ -36,7 +36,7 @@ def rng(request):
 ])
 def module_linear(rng, request):
     module_type, kwargs = request.param
-    return module_type(**kwargs).eval()
+    return module_type(**kwargs).to(torch.float64).eval()
 
 
 @pytest.fixture(scope='session')
@@ -47,8 +47,8 @@ def module_batchnorm(module_linear):
         ((Conv3d, ConvTranspose3d), BatchNorm3d),
     ]
     feature_index_map = [
-        ((Linear, ConvTranspose1d, ConvTranspose2d, ConvTranspose3d), 1),
-        ((Conv1d, Conv2d, Conv3d), 0),
+        ((ConvTranspose1d, ConvTranspose2d, ConvTranspose3d), 1),
+        ((Linear, Conv1d, Conv2d, Conv3d), 0),
     ]
 
     batchnorm_type = None
@@ -67,12 +67,12 @@ def module_batchnorm(module_linear):
     if feature_index is None:
         raise RuntimeError('No feature index for linear layer found.')
 
-    return batchnorm_type(num_features=module_linear.weight.shape[feature_index]).eval()
+    return batchnorm_type(num_features=module_linear.weight.shape[feature_index]).to(torch.float64).eval()
 
 
 @pytest.fixture(scope='session')
 def data_input(rng, module_linear):
-    shape = (16,)
+    shape = (4,)
     setups = [
         (Conv1d, 1, 1),
         (ConvTranspose1d, 0, 1),
@@ -86,6 +86,6 @@ def data_input(rng, module_linear):
     else:
         for module_type, dim, ndims in setups:
             if isinstance(module_linear, module_type):
-                shape += (module_linear.weight.shape[dim],) + (16,) * ndims
+                shape += (module_linear.weight.shape[dim],) + (4,) * ndims
 
-    return torch.empty(*shape).normal_(generator=rng)
+    return torch.empty(*shape, dtype=torch.float64).normal_(generator=rng)