Notes on 31 July 2020

Agenda:

State of DataLoader and Dataset in FastAI.jl
- What about Flux.DataLoader is missing?
PR to MLDatasets.jl to make datasets iteratable/map-like
- Already iterable but it would be better to extend Base (currently uses custom traintensor function)
State of transforms
Should we put together an example gist for the ML call?
Discussion on side-effects in Python code

Flux.DataLoader missing a lot features from this list
https://github.com/lorenzoh/DataLoaders.jl has a lot of the features we need
- Author is not responsive
- We could still use this package for inspiration
Can we compose with Dagger.jl and Transducers.jl
- Dagger, yes
- Transducers.jl will require implementing a larger set of interface
- Might need to implement SplittablesBase.jl
Start a repo for DataLoaders and get vanilla up and running
- In parallel branch work on Transducers.jl etc.
- Try to get vanilla example up and running with FastAI.jl
Support for fancy indexing
- Probably an easy refactor once the community settles on a package
- Will need Flux.jl to support fancy indexing first
Transformers design is complicated by weird FastAI type hierarchy/abstractions
- OK to be more "pythonic" at first
- Julia implementations will leverage existing packages
- Take shortcuts through the abstraction when possible
- Open issue on tracker to get feedback from community on design decisions
Meeting time
- Reschedule according to new poll
- Try to get everyone this time
- Settle on regular meeting time