module 'tensorflow._api.v1.compat.v2.compat' has no attribute 'v2'

Prerequisites

1. The entire URL of the file you are using

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

2. Describe the bug

Traceback (most recent call last):
  File "model_main_tf2.py", line 122, in <module>
    tf.compat.v1.app.run()
  File "/usr/local/lib/python3.6/dist-packages/tensorflow_core/python/platform/app.py", line 40, in run
    _run(main=main, argv=argv, flags_parser=_parse_flags_tolerate_undef)
  File "/usr/local/lib/python3.6/dist-packages/absl/app.py", line 299, in run
    _run_main(main, args)
  File "/usr/local/lib/python3.6/dist-packages/absl/app.py", line 250, in _run_main
    sys.exit(main(argv))
  File "model_main_tf2.py", line 106, in main
    strategy = tf.compat.v2.distribute.MirroredStrategy()
  File "/usr/local/lib/python3.6/dist-packages/tensorflow_core/python/util/module_wrapper.py", line 193, in __getattr__
    attr = getattr(self._tfmw_wrapped_module, name)
AttributeError: module 'tensorflow._api.v1.compat.v2.compat' has no attribute 'v2'

I am using model_main_tf2.py with the SSD MobileNet v2 320x320 model from https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/tf2_detection_zoo.md

python model_main_tf2.py -model_dir /root/models/research/object_detection --num_train_steps 400 --pipeline_config_path pipeline.config --sample_1_of_n_eval_examples 1

3. Steps to reproduce

I have taken the pipeline.config (for 13 classes and 400 steps) and the checkpoints provided for SSD MobileNet v2 320x320 pretrained model (see https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/tf2_detection_zoo.md).
The tfrecords are very small (2mb).

Also, I am using sagemaker.estimator (Framework: https://sagemaker.readthedocs.io/en/stable/api/training/estimators.html) to launch the training process on AWS Sagemaker, but this shouldn't make any difference, since I get the same error even on my local machine.

4. Additional context

I think line
strategy = tf.compat.v2.distribute.MirroredStrategy()
from model_main_tf2.py is not correctly interpreting tf (tf as import tensorflow.compat.v2 as tf). Substituting tf.compat.v2 by tf seems to fix the problem. However, it later crashes with the call to model_lib_v2 one line later (for the same reason).

I have seen other related issues (https://github.com/tensorflow/models/issues/9651, https://github.com/tensorflow/models/issues/8088), but none of them seems to have a clear answer, just to downgrade the whole models repository.

5. System information

I am using the Docker image: tensorflow/tensorflow:2.5.0-gpu

• OS Platform and Distribution: Ubuntu 18.04.
• TensorFlow installed from (source or binary): Docker image.
• TensorFlow version (use command below): 2.5.0
• Python version: 3.6.9
• CUDA/cuDNN version: cuda-11.2
• GPU model and memory: The GPU from AWS Sagemaker ml.p3.2xlarge