Support for Tensorflow 2.0 in Object Detection API

Question

I am using TF Object Detection API for training a model that I will eventually deploy using tf-serving. I plan to take the output of this network (at intermediate CNN layers) and build additional networks for other functionality besides object detection. I plan to develop new networks in TF 2.0. However, for this to work smoothly, (I suspect) that Object Detection API will need to be updated to support TF-2.0.

Is there a plan to update the TF OD API to TF 2.0?

I tried running the model_builder_test and it failed as shown below, because tensorflow 2.0 doesn't support contrib anymore.

Traceback (most recent call last):
  File "object_detection/builders/model_builder_test.py", line 23, in <module>
    from object_detection.builders import model_builder
  File "/models/research/object_detection/builders/model_builder.py", line 19, in <module>
    from object_detection.builders import box_predictor_builder
  File "/models/research/object_detection/builders/box_predictor_builder.py", line 18, in <module>
    from object_detection.core import box_predictor
  File "/research/object_detection/core/box_predictor.py", line 35, in <module>
    slim = tf.contrib.slim
AttributeError: module 'tensorflow' has no attribute 'contrib'

Rumormonger

Answer 1

Tensorflow 2.0 is still in Alpha. Fortunately, support for Tensorflow 2.0 will eventually be added. See this issue

Bent

Answer 2

Yes there is a plan to migrate the Object Detection API to TF2.0, but it will take months. In the meantime please use TF1.x for the API.

Flathead

Following the official release of TF 2.0 this month, I was wondering if anyone is aware of any updates on the migration of Object Detection API to TF2.0? – Subdued