Changelog

Release 0.21.0

Additions

Add PluggableDeviceLibrary

Changes

Upgrade to TensorFlow 2.13.0
Add support for saving to and restoring from checkpoints

Release 0.20.0

Additions

Implemented Send and Sync for SessionRunArgs

Changes

Migrated from random to rand
Upgraded to TensorFlow 2.11.0

Fixes

Removed binaries from the repository

Release 0.19.0

Additions

Improved debug formatting of tensors
Add accessors for operation named inputs and outputs
Support loading TensorFlow library dynamically at runtime

Changes

Upgrade to TensorFlow 2.9.1

Release 0.18.0

Additions

Added eager API support (Big thanks to dskkato!)
Add pre-trained Keras model use example

Changes

Upgrade to TensorFlow 2.8
Update addition and regression_savedmodel examples' python code for tf2.5
Migrate CI from Travis to GitHub Actions

Fixes

Codegen cleanup
Fixed rustfmt on Windows

Release 0.17.0

Additions

Support Tensor::from and Shape::from for arrays (requires Rust 1.51)
Add dtype and shape accessors to TensorInfo
Implement Scope::with_xla_cluster
Add Scope::with_kernel_label

Changes

Upgrade to TensorFlow 2.5
Deprecate Session::from_saved_model in favor of SavedModelBundle::load

Release 0.16.1

Fixes

Fix Windows build

Release 0.16.0

Additions

Stabilize new graph generation code (removed experimental_training feature)
Add Scope::{with_device, with_control_dependencies}
Add optional support for Tensor conversions to/from ndarray::Array
Add Library::op_list
Allow tensorflow-sys to download prebuilt windows releases

Changes

Improve ergonomics for graph building
- Allow conversions for arguments to generated ops
- Implement From for arrays for Shape
- Allow VariableBuilder::shape to take Into

Fixes

Fix memory safety bug in Operation::control_inputs
Allow 0 colons in output names, default to index 0
Fix docs.rs (hopefully)

Release 0.15.0

Additions

Add generated code for all standard ops
- Currently guarded by experimental_training feature
Add RecordReader for TFRecords
Add support for creating saved models
Document that BFloat16 is not an IEEE-754 16-bit float
Implement Send and Sync for Status
Add Tensor::get and Tensor::set

Changes

Use std::alloc instead of aligned_alloc

Release 0.14.0

Additions

Support for high-level graph building in pure Rust
- Adds Scope, ops module, etc.
- Currently guarded by experimental_training feature
- Includes a basic xor example
Support requesting run metadata from Session::run
Implement TensorType for half::f16
Add From<&[i64]> and From<&[u64]> for Shape
Add Tensor::shape
Add Shape::new

Changes

Change return type of Graph::add_gradients to return optional gradients

Fixes

Fix memory initialization bug in Operation::output_consumers