freeze_variables has no effect when training with model_main_tf2.py

I noticed the freeze_variables train config is never accounted for when training with model_main_tf2.py

Versions

• Tensorflow 2.5
• Commit: latest (Mon Jul 19 12:32:17 2021 -0700)
• tf_env_collect.sh output here

1. The entire URL of the file you are using

https://github.com/tensorflow/models/blob/master/research/object_detection/model_main_tf2.py
https://github.com/tensorflow/models/blob/master/research/object_detection/model_lib_v2.py

2. Describe the bug

When I provide something in the freeze_variables config (here in train.proto), e.g. freeze_variables: ['.*conv.'], and then run training with model_main_tf2.py as in the tutorial, no variables are frozen and all layers are trained. I went through the model and training protocol configuration steps in model_lib_v2.py, in the train_loop method which is called in model_main_tf2.py, and it never seems to touch the freeze_variables option.

3. Steps to reproduce

Follow the steps of the tutorial to set up a training run for the SSD resnet50 v1 fpn model, and use the config file included with the model. Add freeze_variables: ['.*conv.'] to train_config. If you add print(detection_model.trainable_variables) to line 673 of model_lib_v2.py (right before the training begins), it outputs all variables, i.e. all variables are still trainable.

4. Expected behavior

Variables matching the provided freeze_variable regex should no longer be trainable.