Output accuracies per class.

[rmanor]

No description provided.

[bhack]

Is this something like #759?

[rmanor]

No, AFAIU, #759 is when a sample image has several labels and you just compute the accuracy differently (it still outputs one accuracy number, i.e. "correct classification").
Here I'm talking about the usual case when the problem is not binary classification but you have M classes, e.g. MNIST (10 classes), ImageNet (1000 classes), and you want to see the accuracy per class.

[seanbell]

While this is a nice feature, I think it should be an option that is enabled in AccuracyParam, where the default is not to use it. Otherwise, it breaks existing code if you simply overwrite existing functionality with your new feature.

[rmanor]

I agree, I will update the code.

[sguada]

I think it should be a different top blob, that way if one ask for two tops
then the first blob will contain the overall accuracy, and the second could
contain the per class accuracy.

Also you have to make sure it works for top-k accuracy, and not only for top-1 accuracy.

2014-11-19 12:55 GMT-08:00 rmanor notifications@github.com:

I agree, I will update the code.

—
Reply to this email directly or view it on GitHub
#1442 (comment).

[rmanor]

Hi, I refactored the change.
Now it will output per class accuracies in the second top blob, if exists.

[rmanor]

ping? Thanks.

[bhack]

@rmanor Can you wait forever?

[rmanor]

@bhack What?
I didn't know how to update this PR so I'm going to start a new one synced to master.

[bhack]

OK I think that you was tired to waiting for a review

[rmanor]

Yeah, but I thought I give it another try before giving up :)

On Sat, Aug 15, 2015 at 7:40 PM bhack notifications@github.com wrote:

OK I think that you was tired to waiting for a review

—
Reply to this email directly or view it on GitHub
#1442 (comment).

[rmanor]

Replaced this with #2935