Meta-learning experiments

[m09]

Here are the steps I want to go through to test out the meta-learning idea:

• Extract format analyzer parser (bblfsh + operator, whitespace and special cases handling)
• Transform parsed files (virtual nodes + uast) into a graph
• Create a model made of a GGNN encoder and a LSTM decoder with Deep Graph Library and PyTorch
• Overfit 1 file formatted by prettier to check that the model is expressive enough to learn the formatting of one file
• Overfit 1 project formatted by prettier, still to check expressiveness
• Gather a dataset of diverse and somewhat well maintained (ie formatted) projects to learn from (like @warenlg's top javascript repos dataset)
• Define an evaluation scheme made of both interpolation (modeling style on training repos) and extrapolation (modeling style on unseen repos)
• Test 4 approaches to train the model:
  • One model per repository (like style-analyzer)
  • One model for all repositories
  • One model for all repositories with multi-task learning (one task per repository)
  • One model for all repositories with meta-learning (one task per repository + learn to adapt)
• Plug the system into the visualizer to understand results
• If results seem promising, evaluate a bit more and report to give input to product