Some Python objects were not bound to checkpointed values, likely due to changes in the Python program

Question

I am trying to train the CenterNet HourGlass104 512x512 model for object detection with:

python model_main_tf2.py --model_dir=models/centernet_hg104_512x512_coco17_tpu-8 --pipeline_config_path=models/centernet_hg104_512x512_coco17_tpu-8/pipeline.config

When starting training the error:

AssertionError: Some Python objects were not bound to checkpointed values, likely due to changes in the Python program: [SyncOnReadVariable:{
  0: <tf.Variable 'center_net_hourglass_feature_extractor/hourglass_network/encoder_decoder_block_5/encoder_decoder_block_6/encoder_decoder_block_7/encoder_decoder_block_8/encoder_decoder_block_9/residual_block_60/convolutional_block_62/batchnorm/moving_mean:0' shape=(384,) dtype=float32, numpy=
array([

is thrown.

I have found this Thread, but

fine_tune_checkpoint_type: "detection"

was already set.

Does anyone know what I am missing?

Edit (18/02/21):

I changed fine_tune_checkpoint_type to "fine_tune". With this I can start the training, but I get no output regarding the quality of the results. Checkpoint files are generated.