Transfer learning doesn't seem to take parameters from existing checkpoint

In my config file I do have the following directives
fine_tune_checkpoint: "ssd_inception_v2_coco_2018_01_28/model.ckpt"
from_detection_checkpoint: true
load_all_detection_checkpoint_vars: true

I the same high initial loss even after the model was trained for a while and loss was much lower and I start new training from the new checkpoint.
I also see this info printed:
I0906 17:17:05.336543 140295991609152 config_util.py:488] Maybe overwriting load_pretrained: True
I0906 17:17:05.337264 140295991609152 config_util.py:498] Ignoring config override key: load_pretrained

I checked config_util.py and in _maybe_update_config_with_key_value it does not seem to have such a key in the list of key and thus does not recognize it, the function so it return False which causes merge_external_params_with_configs to print this info statement.

Thanks