It seems a bug

[mjjackey]

It seems a bug

[avirambh]

Hi @mjjackey,
This code was developed with Pytorch 0.1.12, and it seems that after version 0.5, 0-dim tensor should be handled with an item() call. I think you can use Pytorch 0.4.0/1.
Pull requests for supporting later versions are welcome of course.

[mjjackey]

Thank you very much! @avirambh.
Sorry for the late reply, I have figured it out already.
And I have a question: how to vary the FLOPs or parameters to plot the figure for accuracy? Do I need to open a new issue for this question?

[avirambh]

We used the script referred by the authors of the paper in the original torch repository. We created the graphs manually.

[mjjackey]

We used the script referred by the authors of the paper in the original torch repository. We created the graphs manually.

@avirambh Hi, avirambn, I am confused that each classifier has been added in the code, but the FLOPs just be output once, how can I know FLOPs of each classifier for anytime prediction?

[avirambh]

@mjjackey we did it manually. If i'm not mistaken, we added prints and checked different model architectures.

[mjjackey]

@avirambh Thank you very much for your timely response! Recently, I am studying this paper, so have a lot of questions. :)
In terms of this question, I have found it the prints adding in the ['Conv2d'], this could be not accurate because 'Conv2d' is not limited in classifiers.