7239447415c21586291d5c993e30fba5df54b9db,foolbox/models/pytorch.py,PyTorchModel,predictions_and_gradient,#PyTorchModel#Any#Any#,105

Before Change

        images = image[np.newaxis]
        images = torch.from_numpy(images).to(self.device)

        if self._old_pytorch():  // pragma: no cover
            target = Variable(target)
            images = Variable(images, requires_grad=True)
        else:
            images.requires_grad_()

        predictions = self._model(images)
        ce = nn.CrossEntropyLoss()
        loss = ce(predictions, target)
        loss.backward()
        grad = images.grad

        if self._old_pytorch():  // pragma: no cover
            predictions = predictions.data
        predictions = predictions.to("cpu")

        if not self._old_pytorch():
            predictions = predictions.detach()
        predictions = predictions.numpy()
        predictions = np.squeeze(predictions, axis=0)
        assert predictions.ndim == 1
        assert predictions.shape == (self.num_classes(),)

        if self._old_pytorch():  // pragma: no cover
            grad = grad.data
        grad = grad.to("cpu")
        if not self._old_pytorch():
            grad = grad.detach()
        grad = grad.numpy()
        grad = np.squeeze(grad, axis=0)
        grad = self._process_gradient(dpdx, grad)
        assert grad.shape == input_shape

After Change

        assert predictions.ndim == 1
        assert predictions.shape == (self.num_classes(),)

        grad = grad.detach().cpu().numpy()
        grad = np.squeeze(grad, axis=0)
        grad = self._process_gradient(dpdx, grad)
        assert grad.shape == input_shape

Instances