TensorFlow Version Compatibility

This document is for users who need backwards compatibility across different versions of TensorFlow (either for code or data), and for developers who want to modify TensorFlow while preserving compatibility.

Semantic Versioning 2.0

TensorFlow follows Semantic Versioning 2.0 (semver) for its public API. Each release version of TensorFlow has the form MAJOR.MINOR.PATCH. For example, TensorFlow version 1.2.3 has MAJOR version 1, MINOR version 2, and PATCH version 3. Changes to each number have the following meaning:

• MAJOR: Potentially backwards incompatible changes. Code and data that worked with a previous major release will not necessarily work with the new release. However, in some cases existing TensorFlow graphs and checkpoints may be migratable to the newer release; see Compatibility of graphs and checkpoints for details on data compatibility.

• MINOR: Backwards compatible features, speed improvements, etc. Code and data that worked with a previous minor release and which depends only on the public API will continue to work unchanged. For details on what is and is not the public API, see What is covered.

• PATCH: Backwards compatible bug fixes.

For example, release 1.0.0 introduced backwards incompatible changes from release 0.12.1. However, release 1.1.1 was backwards compatible with release 1.0.0.

What is covered

Only the public APIs of TensorFlow are backwards compatible across minor and patch versions. The public APIs consist of

• All the documented Python functions and classes in the tensorflow module and its submodules, except for

  • functions and classes in tf.contrib
  • functions and classes whose names start with _ (as these are private)
  • functions, arguments, properties and classes whose name starts with experimental, or whose fully qualified name includes a module called experimental Note that the code in the examples/ and tools/ directories is not reachable through the tensorflow Python module and is thus not covered by the compatibility guarantee.

  If a symbol is available through the tensorflow Python module or its submodules, but is not documented, then it is not considered part of the public API.

• The C API.

• The following protocol buffer files:

  • attr_value
  • config
  • event
  • graph
  • op_def
  • reader_base
  • summary
  • tensor
  • tensor_shape
  • types

What is not covered

Some API functions are explicitly marked as "experimental" and can change in backward incompatible ways between minor releases. These include:

• Experimental APIs: The tf.contrib module and its submodules in Python and any functions in the C API or fields in protocol buffers that are explicitly commented as being experimental. In particular, any field in a protocol buffer which is called "experimental" and all its fields and submessages can change at any time.

• Other languages: TensorFlow APIs in languages other than Python and C, such as:

• C++ (exposed through header files in tensorflow/cc).
• Java,
• Go
• JavaScript

• Details of composite ops: Many public functions in Python expand to several primitive ops in the graph, and these details will be part of any graphs saved to disk as GraphDefs. These details may change for minor releases. In particular, regressions tests that check for exact matching between graphs are likely to break across minor releases, even though the behavior of the graph should be unchanged and existing checkpoints will still work.

• Floating point numerical details: The specific floating point values computed by ops may change at any time. Users should rely only on approximate accuracy and numerical stability, not on the specific bits computed. Changes to numerical formulas in minor and patch releases should result in comparable or improved accuracy, with the caveat that in machine learning improved accuracy of specific formulas may result in decreased accuracy for the overall system.

• Random numbers: The specific random numbers computed by the random ops may change at any time. Users should rely only on approximately correct distributions and statistical strength, not the specific bits computed. However, we will make changes to random bits rarely (or perhaps never) for patch releases. We will, of course, document all such changes.

• Version skew in distributed Tensorflow: Running two different versions of TensorFlow in a single cluster is unsupported. There are no guarantees about backwards compatibility of the wire protocol.

• Bugs: We reserve the right to make backwards incompatible behavior (though not API) changes if the current implementation is clearly broken, that is, if it contradicts the documentation or if a well-known and well-defined intended behavior is not properly implemented due to a bug. For example, if an optimizer claims to implement a well-known optimization algorithm but does not match that algorithm due to a bug, then we will fix the optimizer. Our fix may break code relying on the wrong behavior for convergence. We will note such changes in the release notes.

• Error messages: We reserve the right to change the text of error messages. In addition, the type of an error may change unless the type is specified in the documentation. For example, a function documented to raise an InvalidArgument exception will continue to raise InvalidArgument, but the human-readable message contents can change.

Compatibility of graphs and checkpoints

You'll sometimes need to preserve graphs and checkpoints. Graphs describe the data flow of ops to be run during training and inference, and checkpoints contain the saved tensor values of variables in a graph.

Many TensorFlow users save graphs and trained models to disk for later evaluation or additional training, but end up running their saved graphs or models on a later release. In compliance with semver, any graph or checkpoint written out with one version of TensorFlow can be loaded and evaluated with a later version of TensorFlow with the same major release. However, we will endeavor to preserve backwards compatibility even across major releases when possible, so that the serialized files are usable over long periods of time.

Graphs are serialized via the GraphDef protocol buffer. To facilitate (rare) backwards incompatible changes to graphs, each GraphDef has a version number separate from the TensorFlow version. For example, GraphDef version 17 deprecated the inv op in favor of reciprocal. The semantics are:

• Each version of TensorFlow supports an interval of GraphDef versions. This interval will be constant across patch releases, and will only grow across minor releases. Dropping support for a GraphDef version will only occur for a major release of TensorFlow.

• Newly created graphs are assigned the latest GraphDef version number.

• If a given version of TensorFlow supports the GraphDef version of a graph, it will load and evaluate with the same behavior as the TensorFlow version used to generate it (except for floating point numerical details and random numbers), regardless of the major version of TensorFlow. In particular, all checkpoint files will be compatible.

• If the GraphDef upper bound is increased to X in a (minor) release, there will be at least six months before the lower bound is increased to X. For example (we're using hypothetical version numbers here):

  • TensorFlow 1.2 might support GraphDef versions 4 to 7.
  • TensorFlow 1.3 could add GraphDef version 8 and support versions 4 to 8.
  • At least six months later, TensorFlow 2.0.0 could drop support for versions 4 to 7, leaving version 8 only.

Finally, when support for a GraphDef version is dropped, we will attempt to provide tools for automatically converting graphs to a newer supported GraphDef version.

Graph and checkpoint compatibility when extending TensorFlow

This section is relevant only when making incompatible changes to the GraphDef format, such as when adding ops, removing ops, or changing the functionality of existing ops. The previous section should suffice for most users.

Backward and partial forward compatibility

Our versioning scheme has three requirements:

• Backward compatibility to support loading graphs and checkpoints created with older versions of TensorFlow.
• Forward compatibility to support scenarios where the producer of a graph or checkpoint is upgraded to a newer version of TensorFlow before the consumer.
• Enable evolving TensorFlow in incompatible ways. For example, removing ops, adding attributes, and removing attributes.

Note that while the GraphDef version mechanism is separate from the TensorFlow version, backwards incompatible changes to the GraphDef format are still restricted by Semantic Versioning. This means functionality can only be removed or changed between MAJOR versions of TensorFlow (such as 1.7 to 2.0). Additionally, forward compatibility is enforced within Patch releases (1.x.1 to 1.x.2 for example).

To achieve backward and forward compatibility and to know when to enforce changes in formats, graphs and checkpoints have metadata that describes when they were produced. The sections below detail the TensorFlow implementation and guidelines for evolving GraphDef versions.

Independent data version schemes

There are different data versions for graphs and checkpoints. The two data formats evolve at different rates from each other and also at different rates from TensorFlow. Both versioning systems are defined in core/public/version.h. Whenever a new version is added, a note is added to the header detailing what changed and the date.

Data, producers, and consumers

We distinguish between the following kinds of data version information:

• producers: binaries that produce data. Producers have a version (producer) and a minimum consumer version that they are compatible with (min_consumer).
• consumers: binaries that consume data. Consumers have a version (consumer) and a minimum producer version that they are compatible with (min_producer).

Each piece of versioned data has a VersionDef versions field which records the producer that made the data, the min_consumer that it is compatible with, and a list of bad_consumers versions that are disallowed.

By default, when a producer makes some data, the data inherits the producer's producer and min_consumer versions. bad_consumers can be set if specific consumer versions are known to contain bugs and must be avoided. A consumer can accept a piece of data if the following are all true:

• consumer >= data's min_consumer
• data's producer >= consumer's min_producer
• consumer not in data's bad_consumers

Since both producers and consumers come from the same TensorFlow code base, core/public/version.h contains a main data version which is treated as either producer or consumer depending on context and both min_consumer and min_producer (needed by producers and consumers, respectively). Specifically,

• For GraphDef versions, we have TF_GRAPH_DEF_VERSION, TF_GRAPH_DEF_VERSION_MIN_CONSUMER, and TF_GRAPH_DEF_VERSION_MIN_PRODUCER.
• For checkpoint versions, we have TF_CHECKPOINT_VERSION, TF_CHECKPOINT_VERSION_MIN_CONSUMER, and TF_CHECKPOINT_VERSION_MIN_PRODUCER.

Add a new attribute with default to an existing op

Following the guidance below gives you forward compatibility only if the set of ops has not changed:

1. If forward compatibility is desired, set strip_default_attrs to True while exporting the model using either the tf.saved_model.builder.SavedModelBuilder.add_meta_graph_and_variables and tf.saved_model.builder.SavedModelBuilder.add_meta_graph methods of the SavedModelBuilder class, or tf.estimator.Estimator.export_savedmodel
2. This strips off the default valued attributes at the time of producing/exporting the models. This makes sure that the exported tf.MetaGraphDef does not contain the new op-attribute when the default value is used.
3. Having this control could allow out-of-date consumers (for example, serving binaries that lag behind training binaries) to continue loading the models and prevent interruptions in model serving.

Evolving GraphDef versions

This section explains how to use this versioning mechanism to make different types of changes to the GraphDef format.

Add an op

Add the new op to both consumers and producers at the same time, and do not change any GraphDef versions. This type of change is automatically backward compatible, and does not impact forward compatibility plan since existing producer scripts will not suddenly use the new functionality.

Add an op and switch existing Python wrappers to use it

1. Implement new consumer functionality and increment the GraphDef version.
2. If it is possible to make the wrappers use the new functionality only in cases that did not work before, the wrappers can be updated now.
3. Change Python wrappers to use the new functionality. Do not increment min_consumer, since models that do not use this op should not break.

Remove or restrict an op's functionality

1. Fix all producer scripts (not TensorFlow itself) to not use the banned op or functionality.
2. Increment the GraphDef version and implement new consumer functionality that bans the removed op or functionality for GraphDefs at the new version and above. If possible, make TensorFlow stop producing GraphDefs with the banned functionality. To do so, add the REGISTER_OP(...).Deprecated(deprecated_at_version, message).
3. Wait for a major release for backward compatibility purposes.
4. Increase min_producer to the GraphDef version from (2) and remove the functionality entirely.

Change an op's functionality

1. Add a new similar op named SomethingV2 or similar and go through the process of adding it and switching existing Python wrappers to use it. To ensure forward compatibility use the checks suggested in compat.py when changing the Python wrappers.
2. Remove the old op (Can only take place with a major version change due to backward compatibility).
3. Increase min_consumer to rule out consumers with the old op, add back the old op as an alias for SomethingV2, and go through the process to switch existing Python wrappers to use it.
4. Go through the process to remove SomethingV2.

Ban a single unsafe consumer version

1. Bump the GraphDef version and add the bad version to bad_consumers for all new GraphDefs. If possible, add to bad_consumers only for GraphDefs which contain a certain op or similar.
2. If existing consumers have the bad version, push them out as soon as possible.