RuntimeError

[YOU-TO-BE]

RuntimeError Traceback (most recent call last)
in
9 # Ready to roll!
10
---> 11 backprop.visualize(owl, target_class, guided=True)

~\Downloads\flashtorch-master\flashtorch\saliency\backprop.py in visualize(self, input_, target_class, guided, use_gpu, figsize, cmap, alpha, return_output)
180 # (title, [(image1, cmap, alpha), (image2, cmap, alpha)])
181 ('Input image',
--> 182 [(format_for_plotting(denormalize(input_)), None, None)]),
183 ('Gradients across RGB channels',
184 [(format_for_plotting(standardize_and_clip(gradients)),

~\Downloads\flashtorch-master\flashtorch\utils_init_.py in denormalize(tensor)
117
118 for channel, mean, std in zip(denormalized[0], means, stds):
--> 119 channel.mul_(std).add_(mean)
120
121 return denormalized

RuntimeError: Output 0 of UnbindBackward is a view and is being modified inplace. This view is the output of a function that returns multiple views. Such functions do not allow the output views to be modified inplace. You should replace the inplace operation by an out-of-place one.

[drscotthawley]

Came here to post this because I just got the error on Colab:

----> 3 backprop.visualize(img, target_class, guided=True, use_gpu=True)

1 frames
/usr/local/lib/python3.7/dist-packages/flashtorch/saliency/backprop.py in visualize(self, input_, target_class, guided, use_gpu, figsize, cmap, alpha, return_output)
    180             # (title, [(image1, cmap, alpha), (image2, cmap, alpha)])
    181             ('Input image',
--> 182              [(format_for_plotting(denormalize(input_)), None, None)]),
    183             ('Gradients across RGB channels',
    184              [(format_for_plotting(standardize_and_clip(gradients)),

/usr/local/lib/python3.7/dist-packages/flashtorch/utils/__init__.py in denormalize(tensor)
    117 
    118     for channel, mean, std in zip(denormalized[0], means, stds):
--> 119         channel.mul_(std).add_(mean)
    120 
    121     return denormalized

RuntimeError: Output 0 of UnbindBackward is a view and is being modified inplace. This view is the output of a function that returns multiple views. Such functions do not allow the output views to be modified inplace. You should replace the inplace operation by an out-of-place one.

@MisaOgura: Any idea what's going wrong here?

I see a related closed issue in torchvision that suggests wrapping in torch.no_grad() : pytorch/vision#3025 (comment)
...but adding no_grad() would break the backpropagation, wouldn't it? So... Not sure how to fix this.

EDIT: seems the breaking change happened in pytorch 1.7, and affected many other packages. But haven't found another package that had this issue and was trying to visualize gradients.

[ziyi-bear]

@MisaOgura Is There any update version for this problem? or any suggest workround?

I also facing this issue, when using pytorch1.8.1 with gpu cuda111 and also tried without the gpu, but still no luck

thanks