How to convert mobilenet v2 to pb?

System information

• What is the top-level directory of the model you are using: research/object-detection
• Have I written custom code (as opposed to using a stock example script provided in TensorFlow): No
• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Linux Ubuntu 18.04 (NVidia Docker: nvcr.io/nvidia/tensorrt:19.11)
• TensorFlow installed from (source or binary): binary (pip3 install tensorflow==1.12)
• TensorFlow version: v1.12.0-0-ga6d8ffae09 1.12.0
• Bazel version: N/A
• CUDA/cuDNN version:
  nvcc: NVIDIA (R) Cuda compiler driver
  Copyright (c) 2005-2019 NVIDIA Corporation
  Built on Sun_Jul_28_19:07:16_PDT_2019
  Cuda compilation tools, release 10.1, V10.1.243
• GPU model and memory: Quadro P5000
• Exact command to reproduce: see below.

Describe the problem

I am trying to convert sd_mobile net_v2_coco and convert then to uff - > CudaEngine. The original frozen_inference_graph.pb converts perfectly, but frozen_inference_graph after work object_detection/export_inference_graph.py no. If you compare the two models by hash sums-they are different. How to properly convert checkpoint (and ckpt) to get the original frozen_inference_graph.pb?

I run the conversion like this:

export PYTHONPATH=$(realpath ../tensorflow-models/research):$(realpath ../tensorflow-models/research/slim)
INPUT_TYPE=image_tensor
PIPELINE_CONFIG_PATH=$(realpath ssd_mobilenet_v2_13/pipeline.config)
TRAINED_CKPT_PREFIX=$(realpath ssd_mobilenet_v2_13/model.ckpt)
EXPORT_DIR=$(realpath ssd_mobilenet_v2_13/export/)
python3 -m object_detection.export_inference_graph \
    --input_type=image_tensor \
    --pipeline_config_path=${PIPELINE_CONFIG_PATH} \
    --trained_checkpoint_prefix=${TRAINED_CKPT_PREFIX} \
    --output_directory=${EXPORT_DIR}