Ideas¶
Advanced Features¶
Dynamic Shape Tracing¶
Integrate with torch.fx or jax.make_jaxpr to extract graph information for cases where static AST analysis fails (
e.g. data-dependent control flow that resolves to static graph).
Remote Snapshot Registry¶
Currently, snapshots/ are stored in the repo. Move to a remote registry (like conda-forge or a dedicated repo)
allowing users to download mappings for specific framework versions (e.g. ml_switcheroo sync --remote torch==1.13.0).
Type Inference Engine¶
Enhance the PurityScanner to perform basic type inference, allowing for more robust “Method vs Property” resolution
without relying strictly on API naming conventions.
Benchmarking¶
Compare same architecture, precision, hyperparameters across different datasets and accelerators. Show performance from each ML framework, in order to confidently—and with evidence—make haughty claims like:
“ml-switcheroo speeds up ML by up to XXX%”
Novel source/target frameworks¶
Math for theorem provers¶
To/fro Lean or Coq. Maybe u
WASM¶
Also for WASM doc frontend compile ML model to WASM then execute WASM in browser with some way of giving input and showing output… very much TBD
Mobile friendly doc¶
Show always-on-top selection of target, source framework Show always-on-top selection of example Show two tabs, source target Activate target tab when “Translate” clicked