I used to have my torchlife library (https://sachinruk.github.io/torchlife//index.html) in Fastai v1 but since then I converted it to pytorch lightning because I didn’t keep up with the v2 library in 2020. Perhaps this was a mistake.
What I realised in the whole process was that designing a good library is incredibly hard.
I’ve put Survival Analysis theory here: https://sachinruk.github.io/torchlife//SAT
One area that I feel this library is performing poorly is with AFT models and shown below. I’m guessing that I’ve chosen a poor learning rate to get what I’m seeing here: https://sachinruk.github.io/torchlife//model#Modelling-Distribution-with-AFT-models
Considering that I’ve been working on this on and off for about 1 year I’m pretty close to calling quits. But thought I’d ask if there is anyone here who’d like to contribute to this library/ take over just incase.
I’m happy to convert this to fastai v2. But fyi I can’t use the tabular data class because there are two y’s during training time, a binary column saying whether a death event occurred or not (censored), and a continuous variable of time.
Apologies for the long spiel but I’m hoping that Survival Analysis + DL is of use to someone.