Hi Ilastik enthusiasts,
Thanks to the wonderful I2K workshop conducted right now, many of us have been motivated to use the NN-based image analysis offered by Ilastik. My question is now, who is training such models? Do you deposit them somewhere accessible? Would you like to share your data with others who can train them (I’d love to!)?
Hope to hear from many of you 
Cheers,
Dina
Hi @dashour,
welcome to the image.sc community!
I’m not in the business of training networks, so my answer will be quite superficial…
One part of something like an answer to your question is bioimage.io , the bioimage community model zoo. You can already find some pre-trained models for use in ilastik, but chances are high, that nothing fits your data there yet.
So for the “how to train” part of your question: there I’d have to say there should definitely be some guidelines in place. The most important part of training something is of course the data. Lot’s of densely annotated examples are needed. Then you’d have to decide on a framework (to be sure it works in ilastik you’d want to choose pytorch) and some hardware too (training without a GPU is not advisable). From there it goes into the nitty gritty of actual training. A good starting point, I’d say, would be to look into ZeroCostDL (@Guillaume_Jacquemet).
Once you have trained your network, you’d have to package it for use in ilastik into bioimage.io model format which is described here. We would definitely help at this stage
I’m cc’ing some people which do some actual training who I am sure can add more adequate advice.
Hi,
Thanks for the Tag!
If you are interested in #ZeroCostDL4Mic you could check our github https://github.com/HenriquesLab/ZeroCostDL4Mic
We are also very much interested in exporting models trained using #ZeroCostDL4Mic directly into the bioimage.io model format when possible, so hopefully they can be used into other tools including Fiji and ilastik. Some of them can already be used in Fiji but we have not tested the Ilastik connection yet