Userguide updates #12
Conversation
| def get_detector_by_name(self, name: str) -> Optional[Detector]: | ||
| #TODO: Do this on server. | ||
| detector_list = self.list_detectors(page_size=100) | ||
| for d in detector_list.results: |
There was a problem hiding this comment.
does this work if there are more than 100 detectors ?
There was a problem hiding this comment.
It does not. But it will complain instead of failing silently.
| image_query = gl.submit_image_query(detector=detector, image="path/to/filename.jpeg") | ||
| # Show the results | ||
| print(f"The answer is {image_query.result}") |
There was a problem hiding this comment.
should we expose confidence here? its actually a little tricky because when a human reviews it the confidence becomes None (!).
also, this only gets the ML answer, which can be total junk. if you want to wait for a definitive answer, you have to poll it. i don't know what we do to encourage good behavior there or if it matters if people just busy wait and poll.
i think we can expose the example here of how to check threshold, wait for a good enough answer. its kindof awesome and a really unique feature.
There was a problem hiding this comment.
Yeah, this does require more thought, and my initial thoughts are probably wrong.
Seems like there are multiple levels:
A. We can specify the confidence_threshold at detector creation (or update) time. Then, when a posicheck comes:
- we return a boolean on whether we are above the confidence_threshold
- or we return
Nonefor result if we're under-confident. If we are above the confidence threshold, or it comes from a human, we return non-Noneresult.
B. We add an optional polling_timeout argument to submit_image_query, and we do optionally do the polling inside that function on behalf of the user. We can also add an optional confidence_threshold parameter to submit_image_query() that overrides the detector-level config.
Not sure if we need these now, or should just include a second, more detailed example code. Handling varying confidences definitely makes things more complicated.
There was a problem hiding this comment.
I think we should do at least one of these things , but I'd rather put that in another PR.
| image_query = gl.submit_image_query(detector=detector, image="path/to/filename.jpeg") | ||
| # Show the results | ||
| print(f"The answer is {image_query.result}") |
There was a problem hiding this comment.
Yeah, this does require more thought, and my initial thoughts are probably wrong.
Seems like there are multiple levels:
A. We can specify the confidence_threshold at detector creation (or update) time. Then, when a posicheck comes:
- we return a boolean on whether we are above the confidence_threshold
- or we return
Nonefor result if we're under-confident. If we are above the confidence threshold, or it comes from a human, we return non-Noneresult.
B. We add an optional polling_timeout argument to submit_image_query, and we do optionally do the polling inside that function on behalf of the user. We can also add an optional confidence_threshold parameter to submit_image_query() that overrides the detector-level config.
Not sure if we need these now, or should just include a second, more detailed example code. Handling varying confidences definitely makes things more complicated.
And a couple useful changes to the SDK.