-
Notifications
You must be signed in to change notification settings - Fork 8
Conversation
Merged in the new Model Training section as well to make it easier to review these changes in the full context of the spec. |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I like "behave as expected". This is basically based on the knowledge and judgment of the developer.
In a future release we may want to expand this and allow for transfer learning applications. For example, if they have used the model in a region/domain that has limited validation data what has been the performance.
@calebrob6 @batic No pressure, but if you want to leave comments on this, please do. I'll plan on merging by 4 PM EDT today if there are no pressing concerns. |
The section very aptly addresses where "the model will work well". In my experience, we often know also where/when the model will not perform (well, or at all). As such, perhaps the section should be extended (or re-worded) so that the information therein can also identify areas/... where the model is not to be used. |
a new geographic region. This section is meant to capture that information in a systematic way, | ||
while also providing publishers the opportunity to describe their recommendations in a more | ||
free-form style. | ||
|
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
A short sentence/paragraph could be added, noting that the section can also provide possible conditions, where the model will not perform.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Updated in a829170 to have separate recommendations
and cautions
sections.
That's a really good point. It seems like restructuring this to include "recommended" and "discouraged" sections with the same structure might be the best approach. That way a users could apply the same filter logic (just inverted) to find both recommended and discouraged conditions. |
The section has been updated in a829170 to include a field for "recommendations" and a field for "cautions." Each of these is a list of objects describing the conditions that apply to that recommendation/caution. |
@batic I'm going to merge this so we can cut an initial release. If you have any concerns about the structure, feel free to open an issue and we can touch things up in a future PR. Thanks for your input on this! |
Adds a Usage Recommendation section for describing conditions under which a model should behave as expected.
This section is based on the outline from the original Google Doc, except that it leaves out the "Input Data" element from that overview. Finding a way of consistently representing restrictions on input data (e.g. GSD, bands, etc.) in a way that is both machine-readable and understandable to humans is important, but also a difficult problem. In the interest of pushing towards an initial release of this spec that we can begin testing, I think it might be best to tackle that problem as a separate PR so that we can have a more focused discussion.
The term "behave as expected" is admittedly a bit vague. It didn't seem appropriate to use the term "perform well" since there is nothing in the current spec that guarantees that a model "performs well" even under conditions identical to the training environment. If anyone has suggestions on how to better define this I'm open to input.