Convert Tensorflow Object Detection model to Tensorflow lite model in Easy Steps

Below three steps are my hands on solution for converting TF SSD models to TF lite models for object detection. Following are the steps and coverage of this topic

1. Download the model .tar file from github.
2. Convert the graph to tensorflow lite compatible graph.
3. Generate the tflite file from the graph
4. Troubleshooting - I have added some common issues that you might face and the solution.

Assumptions.

• Tensorflow is already installed on your machine.
• You know which object detection model you will use.
• Below is done for ssd inception model v2.
• You can find information about arguments in the command below on google docs.

Step 1. Download the model .tar from github
https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/detection_model_zoo.md

Step2. Convert the TF graph to TF Lite compatible graph.
When you install latest version of tensorflow, you have a utility in the tensorflow -> research -> object detections folder . Name of the utility is export_tflite_ssd_graph.py

• You can use the following command to generate the TF lite graph in your folder (in the below case, I made a folder named "my_tflite_mobile_inceptionv2"

Create the TFLite compatible graph using command like

• python export_tflite_ssd_graph.py --pipeline_config_path ./ssd_inception_v2_coco_2018_01_28/pipeline.config --trained_checkpoint_prefix ./ssd_inception_v2_coco_2018_01_28/model.ckpt --output_directory ./ssd_inception_v2_coco_2018_01_28/my_tflite_mobile_inceptionv2

Step3. Generate the tflite file from the graph.
This can be also done using TOCO convertor, that comes with tensorflow installation.
You can use the following command to do so

Create the tflite using toco converter (comes with tensorflow installtion)

• toco --graph_def_file="C:\Program Files\Python\Lib\site-packages\tensorflow\models\research\object_detection\ssd_inception_v2_coco_2018_01_28\my_tflite_mobile_inceptionv2\tflite_graph.pb" --output_file="C:\Program Files\Python\Lib\site-packages\tensorflow\models\research\object_detection\ssd_inception_v2_coco_2018_01_28\my_tflite_mobile_inceptionv2\tflite_mobile.tflite" --input_shapes=1,300,300,3 --input_arrays=normalized_input_image_tensor --output_arrays=TFLite_Detection_PostProcess,TFLite_Detection_PostProcess:1,TFLite_Detection_PostProcess:2,TFLite_Detection_PostProcess:3 --inference_type=FLOAT --allow_custom_ops

Troubleshooting some issues

Error :

Raise ValueError('SSD Inception V2 feature extractor always uses'

ValueError: SSD Inception V2 feature extractor always usesscope returned 

Solution :

Go to your pipeline.config and under feature_extractor right after conv_hyperparameters closing bracket paste the line

override_base_feature_extractor_hyperparams: true.

So the block looks like this

        activation: RELU_6
        batch_norm {
          decay: 0.999700009823
          center: true
          scale: true
          epsilon: 0.0010000000475
          train: true
        }
      }
      override_base_feature_extractor_hyperparams: true
    }
    box_coder {
      faster_rcnn_box_coder {
        y_scale: 10.0
        x_scale: 10.0
        height_scale: 5.0
        width_scale: 5.0
      }

From <https://github.com/tensorflow/models/issues/7411>