Update projected_gradient_descent_pytorch.py

salomonhotegni · salomonhotegni · commit c0da48c51477 · 2024-07-01T20:17:13.000+02:00
Solve non-writable NumPy array and device mismatch issues.

Signed-off-by: salomonhotegni &lt;salomon.hotegni@aims.ac.rw&gt;
diff --git a/art/attacks/evasion/projected_gradient_descent/projected_gradient_descent_pytorch.py b/art/attacks/evasion/projected_gradient_descent/projected_gradient_descent_pytorch.py
@@ -501,7 +501,7 @@ def _projection(
             )
         else:  # Optimal
             if norm == np.inf:  # Easy exact case
-                values_tmp = values_tmp.sign() * torch.minimum(values_tmp.abs(), torch.Tensor(eps))
+                values_tmp = values_tmp.sign() * torch.minimum(values_tmp.abs(), torch.tensor(eps).to(values_tmp.device))
             elif norm >= 1:  # Convex optim
                 raise NotImplementedError(
                     "Finite values of `norm_p >= 1` are currently not supported with `suboptimal=False`."

Original file line number	Diff line number	Diff line change
`@@ -501,7 +501,7 @@ def _projection(`
`501`	`501`	`)`
`502`	`502`	`else: # Optimal`
`503`	`503`	`if norm == np.inf: # Easy exact case`
`504`		`- values_tmp = values_tmp.sign() * torch.minimum(values_tmp.abs(), torch.Tensor(eps))`
	`504`	`+ values_tmp = values_tmp.sign() * torch.minimum(values_tmp.abs(), torch.tensor(eps).to(values_tmp.device))`
`505`	`505`	`elif norm >= 1: # Convex optim`
`506`	`506`	`raise NotImplementedError(`
`507`	`507`	"Finite values of `norm_p >= 1` are currently not supported with `suboptimal=False`."