Ivy promises framework freedom, but currently charges one-way fares

What it does

Ivy is a transpiler for machine learning code. It takes functions, classes, or even entire third-party libraries written for one framework and rewrites them into another—think of it as a dialect translator for tensor operations. The API centers on ivy.transpile for framework-to-framework conversion and ivy.trace_graph for stripping away wrappers to produce a clean computational graph.

The interesting bit

The clever part is the lazy transpilation of whole modules: you can point Ivy at a library like Kornia (PyTorch-only) and get a TensorFlow-compatible proxy that converts each function only when first called, caching the result for subsequent runs. That avoids paying the upfront cost of translating code you might never touch.

Key highlights

• Supports transpiling from PyTorch to TensorFlow, JAX, or NumPy.
• Can transpile individual functions eagerly or entire modules lazily.
• Includes trace_graph to remove redundant wrapping and trace an efficient graph.
• Active CI with transpiler and integration test workflows.

Caveats

• The conversion matrix is lopsided: PyTorch is currently the only reliable source, while TensorFlow, JAX, and NumPy are mainly targets. You cannot yet transpile to PyTorch or from the others in any practical sense.
• The README contains a small but visible typo in the trace_graph example (referencing an undefined jax_fn), which suggests the docs are moving fast and not always in lockstep.

Verdict

Worth a look if you have a PyTorch model or library that needs to run in a TensorFlow or JAX pipeline and you lack the patience for a manual rewrite. If you need to move code into PyTorch, this is not your tool—yet.