Models Versions Save

This release of the Official Models targets TensorFlow 2.12.0. Note that Research/tutorial/sample models have been removed.

This release fixes various typing annotation issues in official/vision that do not work for Python version <3.9.

This release includes a few updates to RetinaNet model under official/vision:

1. Skip generating image_info in SavedModel is input_type is tflite.
2. Normalize anchors during TFLite post-processing in object detection.
3. Add v3 NMS for Edge TPU.
4. Support class agnostic NMS for v2 NMS.
5. Support output decoded boxes (before NMS) even when NMS is applied.
6. A few other bug fix to make detection generator more robust.

Fix typing annotation in vision model export that breaks package import.

Match the changes with Pypi package and tensorflow.org documentation.

This release of the Official Models targets TensorFlow 2.11.0. Note that Research/tutorial/sample models have been removed.

New features:

1. MobileNetV2 backbone for Mask RCNN
2. Panoptic deeplab supported;
3. freeze_backbone supported, and many bug fixes
4. FNet and SparseMixer implementations
5. Move ViT models from projects to vision main backbone folder

Release branch is: https://github.com/tensorflow/models/tree/r2.11

Fix open cv python 3.10 build issue. A lot of downstream clients need to build with 3.10.

What's Changed

• Remove opencv-python-headless strict version requirement by @joezoug in https://github.com/tensorflow/models/pull/10831

New Contributors

• @joezoug made their first contribution in https://github.com/tensorflow/models/pull/10831

Full Changelog: https://github.com/tensorflow/models/compare/v2.10.0...v2.10.1

This release of the Official Models targets TensorFlow 2.10.0. Note that Research/tutorial/sample models have been removed.

TF versions of requirements have been updated to use the corresponding major version.

Updates:

1. MobileNetV2 backbone for Mask RCNN.
2. Panoptic deeplab supported.
3. freeze_backbone supported, and many bug fixes.

Pypi package for 2.10 cut is: https://pypi.org/project/tf-models-official/2.10.0/.

Adds a prefetch_buffer_size option to set the prefetch buffer size for input_reader instead of autotune. This resolves the issue of increasing buffer size until the CPU memory is exhausted.

Updated pip package: https://pypi.org/project/tf-models-official/2.7.2/

Fix a few import path and init problems.

We have the initial support to import the model garden libraries as: import tensorflow_models as tfm tfm.nlp.XXX

Pypi package for 2.9.2 cut is: https://pypi.org/project/tf-models-official/2.9.2/

Note: how the symbols are exported is defined by: https://github.com/tensorflow/models/tree/master/tensorflow_models/ folder