qupath / qupath-docs

QuPath documentation
http://qupath.readthedocs.io/
17 stars 22 forks source link

Fix StarDist script #7

Closed dkurt closed 3 years ago

petebankhead commented 3 years ago

No idea how you came across this, but thanks @dkurt :)

(Having read many of your posts, I see StarDist can now be converted to an OpenCV-compatible .pb - so this may well become the main approach used by QuPath in the future.)

dkurt commented 3 years ago

@petebankhead, Hi! Many thanks for such kind of attention to my job :)

Yeah, we started playing with QuPath this week. There are plans to enable Intel's OpenVINO as well (in the similar way as TensorFlow extension is implemented). Just need some time to study QuPath internal structure.

petebankhead commented 3 years ago

I'm intrigued and would be happy to chat about any of this :) Deep learning inference is still a somewhat secret feature of QuPath, partly because we haven't established the most practical/maintainable design for our main use cases, and partly because we just don't have the capacity to easily support the number of researchers who might want to use it. But it will be a key feature & coming as quickly as we can figure out these details - as well as our tools to annotate efficiently (which are perhaps relevant to many other domains).

Internal structure is changing to make this possible, and is open to being influenced :)

The developments in OpenCV are helping a lot. The more we can use OpenCV / dependencies that can be distributed easily, the better for our users - most do not want to code or build anything from source.

I still have frequent UnsupportedLayer problems with some of the few relevant models researchers have shared in this space, but I find tf2onnx with --output_frozen_graph has been the simplest way to get some of them (including StarDist) working. And for any models I train myself I make sure to use a compatible architecture.

Anyway, this is an unexpected outcome of that errant 's'. Our forum is here or my email should be easily findable through University of Edinburgh if there's anything you want to discuss more deeply :)