See Medium blog: Accelerate Big Transfer (BiT) model inference with Intel® OpenVINO™

tf_freeze_graph.py

"""
 Copyright (c) 2022 Intel Corporation

 Licensed under the Apache License, Version 2.0 (the "License");
 you may not use this file except in compliance with the License.
 You may obtain a copy of the License at
      http://www.apache.org/licenses/LICENSE-2.0
 Unless required by applicable law or agreed to in writing, software
 distributed under the License is distributed on an "AS IS" BASIS,
 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License.
"""

import argparse
import logging as log
import sys, os
import re
from distutils.version import LooseVersion
from argparse import ArgumentParser

os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"

try:
    import tensorflow.compat.v1 as tf_v1
    # disable eager execution of TensorFlow 2 environment immediately
    tf_v1.disable_eager_execution()
    import tensorflow as tf
    from tensorflow.python.framework.convert_to_constants import (
        convert_variables_to_constants_v2,
    )
except ImportError:
    import tensorflow as tf_v1

def build_argparser():

    usage = """python tf_freeze_graph.py -i </path/to/tf/savedmodel/dir>"""

    parser = ArgumentParser(
        prog="python tf_freeze_graph.py",
        description="Convert TF SavedModel to frozen pb file",
        epilog=usage,
    )
    parser.add_argument("-i", "--input_dir", help="Path to TF SavedModel directory", required=True)

    return parser.parse_args()

def get_graph_def(saved_model_dir):
    # enable eager execution temporarily while TensorFlow 2 model is being loaded
    tf_v1.enable_eager_execution()

    try:
        # Code to extract Keras model.
        imported = tf.keras.models.load_model(saved_model_dir, compile=False)
    except:
        imported = tf.saved_model.load(saved_model_dir)

    # to get a signature by key throws KeyError for TF 1.x SavedModel format in case TF 2.x installed
    concrete_func = imported.signatures[
        tf.saved_model.DEFAULT_SERVING_SIGNATURE_DEF_KEY
    ]
    # the aggressive inlining parameter needs to freeze a table of embeddings for Keras Embedding operation
    # and a model with Embedding operation cannot properly converted to IR without this function parameter
    frozen_func = convert_variables_to_constants_v2(
        concrete_func, lower_control_flow=True, aggressive_inlining=True
    )
    graph_def = frozen_func.graph.as_graph_def(add_shapes=True)

    return graph_def

def main():
    args = build_argparser()

    model_name = os.path.basename(args.input_dir)
    dir_name = os.path.dirname(args.input_dir)
    if len(dir_name) == 0 :
        dir_name = "."
    graph_def = get_graph_def(args.input_dir)
    is_text = False
    new_ext = ".pbtxt" if is_text else ".pb"
    out_dir = "".join([dir_name, "/frozen_", model_name])
    out_name = "frozen_graph.pb"

    tf_v1.import_graph_def(graph_def, name="")
    tf_v1.train.write_graph(graph_def, out_dir, out_name, as_text=is_text)

    print(f"\nFrozen graph saved at: {out_dir}/{out_name}\n")

if __name__ == "__main__":
    sys.exit(main() or 0)