search

modAL-python / modAL

A modular active learning framework for Python
https://modAL-python.github.io/
MIT License
2.17k stars 316 forks source link

bayesian DL #51

Open damienlancry opened 5 years ago

damienlancry commented 5 years ago

HI I opened this issue to discuss the implementation of the acquisition functions that you said you would like to make a feature in #48. I am interested in contributing. where should it be implemented? in uncertainty.py?

cosmic-cortex commented 5 years ago

I have created the feature/bayesianDL branch for this purpose. I think that the acquisition functions should be implemented in the modAL.bayesianDL module.

I am not sure that we can implement these functions in a backend-agnostic way. If not, I propose to implement two versions for each acquisition function. Related to this, I have been thinking about the future of modAL a lot and I think it would be a good direction to rebuild modAL like the early Keras, where you could specify the backend (Theano, TensorFlow or CNTK), but the frontend interface was the same. With this architecture, the required sklearn API for the estimator can be dropped, which would be pretty awesome.

One more thing. Unfortunately, my Ubuntu development setup has just given up, so I need to set up my computer again. Since I have just moved to a new country and started a new job, I might not have time to do it in the following few days. In any case, I'll try to do it as soon as I can!