Region Feature Extraction with TF2

I’m trying to get region feature extraction to work with TF2. I have a script where they use faster_rcnn_inception_resnet_v2_atrous_oidv4 from the TF1 model zoo. However, due to hardware restrictions, I have to use TF2. Following enabling proto with region feature output and region feature extraction, I picked faster_rcnn_inception_resnet_v2_640x640 from the TF2 model zoo and re-exported it according to the stack overflow answer mentioned above.
However, instead of the detection_features I’m getting those keys back:
['rpn_box_predictor_features', 'rpn_features_to_crop', 'image_shape', 'rpn_box_encodings', 'rpn_objectness_predictions_with_background', 'anchors', 'feature_maps', 'refined_box_encodings', 'class_predictions_with_background', 'proposal_boxes', 'box_classifier_features', 'proposal_boxes_normalized', 'final_anchors', 'num_proposals']
How can I enable the detection_features key with TF2?