Open bbernhard opened 6 years ago
agree that streamlining training something off this data might help incentivise more people to contribute.
does also seem like the world of build system automation is applicable. Elsewhere someone had talked to me about the idea of 'a build system specifically for training NN's', I wonder if something like that already exists
does also seem like the world of build system automation is applicable. Elsewhere someone had talked to me about the idea of 'a build system specifically for training NN's', I wonder if something like that already exists
interesting thought! I quickly googled a bit, but couldn't find anything useful :/ I guess Jenkins is still the best /most flexible ready-made solution out there..although I think it's not 100% suited for this use case.
But I think it shouldn't be that hard to develop a small web-based tensorflow frontend. I am thinking about a simple html site, where one can select all the labels the model should be trained on. In the background, the appropriate data is downloaded via the API and fed into the tensorflow pipeline. If everything is packaged inside the docker container, it would be pretty easy to spin an instance up (just docker pull ... & docker run ...
) and access it from the host system.
I am thinking about a simple html site, where one can select all the labels the model should be trained on.
That makes a lot of sense; a label-graph might help since you could select between 'detecting animals' or 'detecting cats vs dogs' etc.
fyi: there exists now a dockerized solution (see https://imagemonkey.io/libraries for details)
Setting up tensorflow in order to train a model on own data is quite complicated. It would be cool, if we could provide a docker image that integrates nicely with the ImageMonkey API (and is based on pyimagemonkey).
The training could then be started with a simple commandline script. Maybe also add a webui (jenkins)?