How do I export (download) a source's data?
To export metadata, points, or annotations, first go to the source's Images page:
Once you're on that page, locate the "Image Actions" box below the grid of images:
In the first box, select 'Export'. In the second box, select the export type you want (depending on the type, you may also see additional options). In the third box, choose whether you want to export data for all image search results, or just the results on that page. Click 'Go' to proceed with the export (this may take a long time to finish). To read more details about each export type, click the yellow button with the '?'.
How do I export (download) the image files themselves?
Unfortunately, for now, you can only download image files one at a time by visiting each image individually.
I can't find my preferred label short-code among the existing labels. Should I create a new label?
In general, no. The short code you see on the site-wide label list is only a default. Short codes can be customized on a per-source basis.
Once you add some labels to your source's labelset, you should see an "Edit (customize) label codes" link near the bottom of your source's Labelset page:
Click that link to start editing the label codes for your source's labelset:
So, when should you create a new label? When you can't find an existing label that semantically represents the same concept you want to label, based on the label's full name and description.
When does CoralNet decide to train a new classifier? Do I have to do something to trigger it?
The first classifier is trained once the source has at least 20 images with 'Confirmed' status. An image becomes confirmed when all of its points have confirmed annotations (should say "All Done" in the annotation tool).
The next classifier training will occur when the number of confirmed images has increased by a factor of 1.1. For example, if the previous classifier training occurred at 100 confirmed images, then the next classifier training will occur when there are at least 100 x 1.1 = 110 confirmed images.
This doesn't mean that a new classifier will be saved after every 1.1x increase in confirmed images. The newly-trained classifier will only be kept if its accuracy is at least 1.01x better than the previous classifier. For example, if the old classifier gets 70% accuracy (tested on the evaluation set, which is 1/8 of the available confirmed images), then the new classifier must get at least 70.7% accuracy, otherwise the new classifier will be discarded.
How long does CoralNet take to train a new classifier?
Train time depends on the total number of confirmed images in the source (directly proportional), and the number of points in each image (roughly proportional to the square root). You can use this table to get a rough time estimate:
|Points per image||Train time per image (seconds)|
If someone else is ahead of you in the backend processing queue, you may have to wait for other sources to get processed first, which may mean you have to wait additional hours or days.
Sometimes, something does break in the processing pipeline, and has to be fixed manually by us. If you have to wait over 4 days, you can ask us on the Google Group forum to see if something is wrong.
We are hoping to add some better status information on classifier training in the near future, such as a display which would tell you if your classifier is currently being trained, or tell you if you're waiting in the queue behind other sources.
There used to be an annotation statistics page with charts. Where did it go?
We needed to make a lot of changes to this page to keep it working in CoralNet Beta (2016), and we just didn't get around to it, so we removed the page for now. Hopefully it can make a return in the near future.