'Related Jobs' window feature

[TimZaman]

It's too bad #137 was abandoned. It was a huge amount of half-working features that i am trying to split out and work on one by one (the which i want to add). The related jobs view is very nice to me as i can see which datasets i can throw away when they are not used, and i can jump around easily between jobs using the name dataset.
Therefore the feature request to implement this, which i have already done. i will file a pull request soon.

[lukeyeager]

Why close this? I agree with you - I'd like to see this feature added soon.

[TimZaman]

Ah sorry Luke. I have this feature implemented in my own fork, but i worked very fast so I didnt put it in a nice branch to make a decent pull request. Actually, this feature is the one i miss most (from my own fork) when i use the vanilla Nvidia-master branch.

[lukeyeager]

Hey @TimZaman are you working on this already? I see some commits referencing this issue. If not, @Lucaszw can do it.

[TimZaman]

It's been in our fork for ages, and I have no clue how you guys can live without it. How about the Confusion Matrix? That's just so damn important to us (also in our fork, but only for torch models): sometimes your accuracy is 90%, and only the confusion matrix will show it's just 100% biased towards a specific label (due to dataset bias). The todo for the confmatrix would be implementing it for caffe too and maybe cleaning up the d3 for the graph a tiny tad.

To be on topic, i will commit the Related Jobs next week. Makes no sense doing double work; and it's fairly straightforward.

Example: 77% accuracy in graph.. but meanwhile the confusion matrix tells me:

[lukeyeager]

i will commit the Related Jobs next week

Cool, thanks!

How about the Confusion Matrix?

Well we do have this: #608.

But it looks like you're talking about something different - storing all the validation results and viewing those as a confusion matrix. Cool idea! I think that might be hard to do for Caffe...

[TimZaman]

Torch generally outputs the confusion matrix during validation, i set up a
simple regex that i catch and that's saved in the pickle file an updated
and sent with socketio with validation pass and voila. And i dont output
the confmat when there are more than 100 labels or it gets hefty. I didnt
dive deep in the caffe/torch implementation but i bet its similar-ish
except from the system calling and regexing.

On Friday, 20 May 2016, Luke Yeager notifications@github.com wrote:

i will commit the Related Jobs next week

Cool, thanks!

How about the Confusion Matrix?

Well we do have this: #608 #608.

But it looks like you're talking about something different - storing all
the validation results and viewing those as a confusion matrix. Cool
idea! I think that might be hard to do for Caffe...

—
You are receiving this because you were mentioned.
Reply to this email directly or view it on GitHub
#667 (comment)

[lukeyeager]

Whoops, forgot to close this when we merged #767. Thanks Tim. I like putting "Close #XXX" in the PR body so this gets done automatically:
https://github.com/blog/1506-closing-issues-via-pull-requests