TF2JAX is an experimental library for converting TensorFlow functions/graphs to JAX functions.

Specifically, it aims to transform a tf.function, e.g.

@tf.function
def tf_fn(x):
  return tf.sin(tf.cos(x))

to a python function equivalent to the following JAX code.

def jax_fn(x):
  return jnp.sin(jnp.cos(x))

Users are able to apply additional JAX transforms (e.g. jit, grad, vmap, make_jaxpr, etc.) to the converted function as they would any other code written in JAX.

[TOC]

You can install the latest released version of TF2JAX from PyPI via:

pip install tf2jax

or you can install the latest development version from GitHub:

pip install git+https://github.com/google-deepmind/tf2jax.git

TF2JAX enables existing TensorFlow functions and models (including SavedModel and TensorFlow Hub) to be reused and/or fine-tuned within JAX codebases. The conversion process is transparent to the users, which is useful for debugging and introspection.

This also provide a pathway for JAX users to integrate JAX functions serialized via jax2tf.convert, back into their existing JAX codebases.

See section at the end for comparison with an alternative approach provided by jax2tf.call_tf.

This is experimental code with potentially unstable API, and there are no guarantees for using it at this point in time. We highly recommend you thoroughly test the resulting JAX functions to ensure they meet your requirements.

The rest of this document assumes the following imports:

import jax
import jax.numpy as jnp
import numpy as np
import tensorflow as tf  # Assumes this is v2.
import tf2jax

An example using the convert API and the Sonnet v2 MLP.

import sonnet.v2 as snt

model = snt.nets.MLP((64, 10,))

@tf.function
def forward(x):
  return model(x)

x = np.random.normal(size=(128, 16)).astype(np.float32)

# TF -> JAX, jax_params are the network parameters of the MLP
jax_func, jax_params = tf2jax.convert(forward, np.zeros_like(x))

# Call JAX, also return updated jax_params (e.g. variable, batchnorm stats)
jax_outputs, jax_params = jax_func(jax_params, x)

tf2jax.convert has the signature convert(fn: tf.Function, *args, **kwargs), where fn(*args, **kwargs) is used to trace the function fn and generates the corresponding tf.GraphDef. The zeros_like is not necessary, only used here to demonstrate the JAX function is not memorizing the outputs.

If your function is pure, i.e. it does not capture any variables, then you can drop the parameters from the inputs and outputs of the converted function with tf2jax.convert_functional.

@tf.function
def forward(x):
  return tf.sin(tf.cos(x))

jax_func = tf2jax.convert_functional(forward, np.zeros_like(x))
jax_outputs = jax_func(x)

A TensorFlow function that make use of random ops will be converted to a JAX function that takes a PRNG key as a keyword-only argument. TF2JAX will complain loudly if a PRNG key is required but not provided.

jax_outputs, jax_params = jax_func(jax_params, x, rng=jax.random.PRNGKey(42))

Custom gradient support is highly experimental, please report any errors.

@tf.function
@tf.custom_gradient
def forward(x):
  e = tf.exp(x)
  def grad(dy):
    return dy * tf.sin(x) + e  # # This is deliberately the wrong gradient.
  return tf.reduce_sum(e), grad

with tf2jax.override_config("convert_custom_gradient", True):
  jax_func = tf2jax.convert_functional(forward, np.zeros_like(x))

jax_grads = jax.grad(jax_func)(x)

SavedModel is the preferred format for serializing TF2 functions.

model = tf.Module()
model.f = forward
model.f(x)  # Dummy call.
tf.saved_model.save(model, "/tmp/blah")

restored = tf.saved_model.load("/tmp/blah")
jax_func, jax_params = tf2jax.convert(restored.f, np.zeros_like(x))

If the restored function has an unambiguous signature, i.e. it was only traced once prior to export. Then TF2JAX can convert the function directly from its GraphDef without tracing it again.

jax_func, jax_params = tf2jax.convert_from_restored(restored.f)

The (legacy, TF1) TF-Hub format is supported with minor boilerplate.

import tensorflow_hub as hub

hub_model = hub.load("/tmp/blah")
jax_func, jax_params = tf2jax.convert(tf.function(hub_model), tf.zeros_like(x))
jax_outputs, updated_jax_params = jax_func(jax_params, x)

tf2jax.convert_functional can convert the outputs of jax2tf.convert back into JAX code.

# Some JAX function.
def forward(*inputs):
  ...

# JAX -> TF
tf_func = jax2tf.convert(forward)

# JAX -> TF -> JAX
jax_func = tf2jax.convert_functional(tf.function(tf_func), *tree.map_structure(np.zeros_like, inputs))

# JAX -> TF -> SavedModel -> TF
model = tf.Module()
model.f = tf.function(tf_func)
model.f(*tree.map_structure(tf.zeros_like, inputs))  # Dummy call.
tf.saved_model.save(model, "/tmp/blah")
restored = tf.saved_model.load("/tmp/blah")

# JAX -> TF -> SavedModel -> TF -> JAX
jax_too_func = tf2jax.convert_functional(restored.f, *tree.map_structure(np.zeros_like, inputs))

The behaviour of TF2JAX can be configured globally via tf2jax.update_config, or configured locally via the context manager tf2jax.override_config.

By default, TF2JAX will assert that the input shapes to the converted function are compatible with the input shapes of the original function. This is because some functions have shape dependent behaviours that will silently return the incorrect outputs after conversion, e.g. some batchnorm implementation.

jax_func = tf2jax.convert_functional(forward, np.zeros((10, 5), np.float32))

# This will raise an error.
jax_func(np.zeros((20, 5), np.float32))

# This will not.
with tf2jax.override_config("strict_shape_check", False):
  jax_func(np.zeros((20, 5), np.float32))

The input dtypes are not currently checked but this may change in the future.

jax_func = tf2jax.convert_functional(forward, np.zeros((10, 5), np.float32))

# This will not raise an error.
jax_func(np.zeros((20, 5), np.int32))

# This will.
with tf2jax.override_config("strict_dtype_check", True):
  jax_func(np.zeros((20, 5), np.int32))

TF2JAX allows users to trace the converted function with parameters and inputs of different precision than the original function, e.g. bfloat16 instead of float32. To aid this, the configuration force_const_float32_to_bfloat16 and force_const_float64_to_bfloat16 can be used to force float constants in the original function into bfloat16 precision, to avoid accidental type promotion.

@tf.function
def forward(x):
  return x + tf.constant(3.14, dtype=tf.float32)

with tf2jax.override_config("force_const_float32_to_bfloat16", True):
  jax_func = tf2jax.convert_functional(forward, np.zeros_like(x))
jax_bf16_outputs = jax_func(jnp.asarray(x, jnp.bfloat16))

If jax2tf.convert(..., with_gradient=False) is used to produce the initial TF function (and possibly exported as SavedModel), then TF2JAX will respect the inserted tf.raw_ops.PreventGradient ops and raise LookupError when computing gradients.

This can be disabled by setting the configuration raise_on_prevent_gradient to false (default is true), so that TF2JAX will only log a warning but otherwise allow the gradient to be computed as though the PreventGradient ops were not present.

@tf.function
def prevent(x):
  return tf.raw_ops.PreventGradient(input=x * x, message="prevented")

jax_func = tf2jax.convert_functional(prevent, 0.0)
jax.grad(jax_func)(3.14)  # Raise LookupError.

with tf2jax.config.override_config("raise_on_prevent_gradient", False):
  jax_func = tf2jax.convert_functional(prevent, 0.0)
g = jax.grad(jax_func)(3.14)  # Returns 6.28.

If the infer_cumulative_reduction_from_jax2tf flag is true (default) then TF2JAX will attempt to infer cummax, cummin, cumprod and cumsum operations from reduce_window operations generated by JAX2TF. This provides better performance because reduce_window implementation of these ops have O(N^2) complexity on CPU and GPU backends, and can suffer from long compilation times due to aggressive constant folding.

See jax2tf_cumulative_reduction for more context.

From JAX v0.4.7 and onward, jax2tf.convert preferred mode of operation (soon to be default) is native serialization in which the target function is lowered to StableHLO and wrapped in a single TensorFlow op, XlaCallModule.

The new native serialization format will more faithfully reproduce the semantics of the target function, at the cost of some reduced flexibility for downstream processing as the computation graph is no longer exposed as a tf.Graph.

XlaCallModule is supported by TF2JAX from v0.3.4 and onward.

However as this makes use of a custom JAX primitive that aims to encapsulate StableHLO payload found in XlaCallModule, it does not possess JAX rules for transformations such as (but not limited to) batching and differentiation.

• Differentiation: first order derivative of serialized function is still supported through custom gradients requested at serialization time with jax2tf.convert(..., with_gradient=True). This is the default behaviour.
• Batching: jax.vmap will fail, though users may be able to naively replicate the desired behavior with jax.lax.map, albeit with poorer performance.

Natively serialized JAX programs are platform specific (link). Executing a natively serialized program on platforms other than the one for which it was lowered, would raise a ValueError, e.g.:

ValueError: Unsupported backend: `cpu` not in `('tpu',)`.

This matches the behaviour of XlaCallModule.

Users can disable this check via a config flag, but the resulting program may be slower or fail to execute completely.

with tf2jax.override_config("xlacallmodule_strict_checks", False):
  jax_func(np.zeros((20, 5), np.float32))

Currently, only a subset of TensorFlow ops are supported, and not necessarily all functionalities are supported for some ops. The code will fail fast. Support for additional TensorFlow ops are added on a as-needed basis. Please submit your requests via Github issues or send in your pull requests.

There will likely to be some cases where the resulting JAX code is not equivalent to the TensorFlow code, both in terms of performance and numerical outputs. The goal is to minimise differences in the latter for model endpoints, ahead of improving performance.

TF2 control flows are supported with some limitations, e.g. for while loops, the cond and body functions cannot have side effects such as assigning to variables.

TF1 control flows are not supported.

jax2tf now also offers the experimental call_tf function which allows JAX to call TensorFlow functions. For compiled code, this works by staging out TensorFlow functions to XLA.

From the jax2tf documentation, as of 2022-07-22:

The function call_tf allows JAX functions to call TensorFlow functions. These functions can be called anywhere in a JAX computation, including in staging contexts jax.jit, jax.pmap, jax.xmap, or inside JAX's control-flow primitives. In non-staging contexts, the TensorFlow function is called in eager mode. For now, only reverse-mode autodiff is supported for these functions (no forward-mode autodiff, nor vmap).

The advantage of call_tf is that it implicitly covers all TensorFlow ops and supports custom_gradient by deferring to TensorFlow during eager execution and to XLA for compiled code.

The disadvantage is that it only supports a limited set of JAX transforms (jit, grad, pmap, remat) and otherwise appears as a "black box" to JAX (e.g. vmap is not supported, nor custom transforms). A TensorFlow function must be compileable to XLA if it is to be jitted after call_tf.

This repository is part of the DeepMind JAX Ecosystem, to cite TF2JAX please use the DeepMind JAX Ecosystem citation.

We are happy to receive pull requests that improve our coverage of TensorFlow ops.