ValueError: Gradient term in PyTorch model is None.

[dalqattan]

I tried to generate adversarial images using PGD.

I got the following error

The .grad attribute of a Tensor that is not a leaf Tensor is being accessed. Its .grad attribute won't be populated during autograd.backward(). If you indeed want the .grad field to be populated for a non-leaf Tensor, use .retain_grad() on the non-leaf Tensor. If you access the non-leaf Tensor by mistake, make sure you access the leaf Tensor instead. See github.com/pytorch/pytorch/pull/30531 for more informations. (Triggered internally at aten/src/ATen/core/TensorBody.h:486.)

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
[<ipython-input-27-1c78037930d6>](https://localhost:8080/#) in <cell line: 1>()
     21 
     22     target = estimator.predict(x)
---> 23     adversarial_image_pgd = pgd.generate(x=x, y=target)
     24 
     25     adversarial_image_pgd = adversarial_image_pgd.squeeze(0).transpose(1, 2, 0)

4 frames
[/usr/local/lib/python3.10/dist-packages/art/attacks/evasion/projected_gradient_descent/projected_gradient_descent.py](https://localhost:8080/#) in generate(self, x, y, **kwargs)
    198         """
    199         logger.info("Creating adversarial samples.")
--> 200         return self._attack.generate(x=x, y=y, **kwargs)
    201 
    202     @property

[/usr/local/lib/python3.10/dist-packages/art/attacks/evasion/projected_gradient_descent/projected_gradient_descent_numpy.py](https://localhost:8080/#) in generate(self, x, y, **kwargs)
    420                 self._i_max_iter = i_max_iter
    421 
--> 422                 adv_x = self._compute(
    423                     adv_x,
    424                     x,

[/usr/local/lib/python3.10/dist-packages/art/attacks/evasion/fast_gradient.py](https://localhost:8080/#) in _compute(self, x, x_init, y, mask, eps, eps_step, project, random_init, batch_id_ext, decay, momentum)
    538 
    539             # Get perturbation
--> 540             perturbation = self._compute_perturbation(batch, batch_labels, mask_batch, decay, momentum)
    541 
    542             # Compute batch_eps and batch_eps_step

[/usr/local/lib/python3.10/dist-packages/art/attacks/evasion/fast_gradient.py](https://localhost:8080/#) in _compute_perturbation(self, x, y, mask, decay, momentum)
    393 
    394         # Get gradient wrt loss; invert it if attack is targeted
--> 395         grad = self.estimator.loss_gradient(x, y) * (1 - 2 * int(self.targeted))
    396 
    397         # Write summary

[/usr/local/lib/python3.10/dist-packages/art/estimators/object_detection/pytorch_yolo.py](https://localhost:8080/#) in loss_gradient(self, x, y, **kwargs)
    402                 grads = x_grad.grad.clone()
    403         else:
--> 404             raise ValueError("Gradient term in PyTorch model is `None`.")
    405 
    406         if self.clip_values is not None:

ValueError: Gradient term in PyTorch model is `None`.

Nay help?
Thanks

Duaa

[beat-buesser]

Hi @dalqattan Which version of ART are you using? And can you show a small example script to reproduce the error?