Dataset visualization - #331
Conversation
|
Great stuff, thanks @groar! I'm trying to test it out, but I keep getting this error: I tried it with PNG, JPEG and raw data in LMDBs - always got the same error. |
|
That's super weird @lukeyeager. I can't understand why you get those kinds of errors (it's the same problem for #137) while I don't. It works for PNG, JPEG, grayscale or color, or any other settings for me. I'm going to investigate the problem, but I can't manage to reproduce it. |
|
Can anyone else test this? Is it working on Travis? The latest build timed out, I restarted it ... |
|
I didn't write any test that uses it. Will do that ! |
|
Ok I pushed a test for the view, so we'll see. |
|
That did it for me. Maybe it was just the little UI suggestions:
Otherwise this looks pretty awesome! I'd like to get @jmancewicz to sign off on this before merging. |
|
yeah I tried the |
|
Thanks @groar, this is very useful! On my CIFAR10 dataset I have 37500 training examples. If I select 10 items per page the list of pages goes from 0 to 3750 (instead of 0--3749 or 1--3750) and on page #3750 there are no images to display. Also, when I am trying to explore my validation DB, if I click on any link of the exploration page I seem to be redirected to the train DB. |
|
Thanks for the feedback @gheinrich and @lukeyeager ! I handled all the UI bugs and improvements you mentioned and squashed the commits. |
13d1d1c to
61a256d
Compare
|
Looks good to me! Squash this into one commit on top of the |
|
Looks good to me too, thanks! |
|
This is very nice. Thanks. Everything looks good. |
20d0ea1 to
a347dee
Compare
|
@groar would you like to rebase on the tip of the master branch? |
regenerated doc
|
Done that ! |
|
Thanks @groar! |


Closes #21
Implemented some way to visualize a dataset db as a gallery, solving #21. I did that to debug #330, as some transformations were not doing what I expected.
You can browse a dataset page by page, reading directly from the db. For now it's only working with lmdb . You can filter by one of the labels from the classification task, as well as change the number of images displayed.
The explorer is accessible from a button in the dataset show view, once a task is done.

The UI is a bit rough and can probably be improved but at least it's functional.
I'll write some tests when I have enough time.