Coming out of a SwiftAI community meeting, we thought it’d be good to have a running to-do list for people to get involved, and tighter communication between the core fast.ai team and wider community.
The idea is to have a list or table of the planned parts of SwiftAI and their current progress. That way someone who wants to contribute can see at a glance what needs help, and hopefully it’ll be easy for everyone in and around fast.ai to be on the same page about what current priorities and challenges are.
@jeremy & co, if you like the idea could you turn this into a wiki?
Below is a starting framework for SwiftAI’s current state of development. Feel free to edit.
SwiftAI is built from a series of Jupyter notebooks. A piece of software is developed, documented, and tested in one notebook, and is imported by the next. The notebooks in the swiftai/nbs directory output Swift modules in swfitai/Sources/SwiftAI/. You can learn about this development technique in course-v3 Part 2: Deep Learning from the Foundations. Lessons 13 & 14 are specific to Swift.
Modules:
- << wish-list & planned functionality, along with current status >>
Implemented:
- Dense Layer
- SGD
- Adam
- Once Cycle
- to-do
Unimplemented:
- to-do
Lessons:
- all unimplemented
Build Notebooks:
- 00_load_data.ipynb
- 01_matmul.ipynb
- 01a_fastai_layers.ipynb
- 02_fully_connected.ipynb
- 02a_why_sqrt5.ipynb
- 03_minibatch_training.ipynb
- 04_callbacks.ipynb
- 05_anneal.ipynb
- 05b_early_stopping.ipynb
- 06_cuda.ipynb
- 07_batchnorm.ipynb
- 08_data_block.ipynb
- 08a_heterogeneous_dictionary.ipynb
- 09_optimizer.ipynb
- 10_mixup_ls.ipynb
- 11_imagenette.ipynb