convert_tf2.8_ckpt_to_saved_model.py

from object_detection.utils import config_util
from object_detection.builders import model_builder

#Where to save the SavedModel
output_directory = './try2_trained_model'
#Recover our saved model with the latest checkpoint:
pipeline_config = 'pipeline.config'
#Put the last ckpt from training in here, don't use long pathnames:
model_dir = './trained_checkpoints/ckpt-31'

#Build model from config file
configs = config_util.get_configs_from_pipeline_file(pipeline_config)
model_config = configs['model']
detection_model = model_builder.build(
      model_config=model_config, is_training=False)

# Restore latest checkpoint
ckpt = tf.train.Checkpoint(
      model=detection_model)
ckpt.restore(model_dir)

#Define function which will perform inferencing
#Function perform detection of the object on image in tensor format: 
import tensorflow as tf
def get_model_detection_function(model):
  """Get a tf.function for detection."""

  @tf.function
  def detect_fn(image):
    """Detect objects in image."""
    image, shapes = model.preprocess(image)
    prediction_dict = model.predict(image, shapes)
    #detections = model.postprocess(prediction_dict, shapes)
    return model.postprocess(prediction_dict, shapes)
    #detections, prediction_dict, tf.reshape(shapes, [-1])

  return detect_fn

detect_fn = get_model_detection_function(detection_model)

tf.saved_model.save(
    detection_model , output_directory ,
    signatures={
        'detect' : detect_fn.get_concrete_function(
            tf.TensorSpec([1, 300, 300, 3], dtype=tf.float32, name='detect'))
        }
)