export_inference_graph.py

"""Export inference graph."""

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import argparse
import os
import sys

import tensorflow as tf

import deeplab_model
from utils import preprocessing


parser = argparse.ArgumentParser()

parser.add_argument('--model_dir', type=str, default='./model',
                    help="Base directory for the model. "
                         "Make sure 'model_checkpoint_path' given in 'checkpoint' file matches "
                         "with checkpoint name.")

parser.add_argument('--export_dir', type=str, default='dataset/export_output',
                    help='The directory where the exported SavedModel will be stored.')

parser.add_argument('--base_architecture', type=str, default='resnet_v2_101',
                    choices=['resnet_v2_50', 'resnet_v2_101'],
                    help='The architecture of base Resnet building block.')

parser.add_argument('--output_stride', type=int, default=16,
                    choices=[8, 16],
                    help='Output stride for DeepLab v3. Currently 8 or 16 is supported.')


_NUM_CLASSES = 21


def main(unused_argv):
  # Using the Winograd non-fused algorithms provides a small performance boost.
  os.environ['TF_ENABLE_WINOGRAD_NONFUSED'] = '1'

  model = tf.estimator.Estimator(
      model_fn=deeplab_model.deeplabv3_model_fn,
      model_dir=FLAGS.model_dir,
      params={
          'output_stride': FLAGS.output_stride,
          'batch_size': 1,  # Batch size must be 1 because the images' size may differ
          'base_architecture': FLAGS.base_architecture,
          'pre_trained_model': None,
          'batch_norm_decay': None,
          'num_classes': _NUM_CLASSES,
      })

  # Export the model
  def serving_input_receiver_fn():
    image = tf.placeholder(tf.float32, [None, None, None, 3], name='image_tensor')
    receiver_tensors = {'image': image}
    features = tf.map_fn(preprocessing.mean_image_subtraction, image)
    return tf.estimator.export.ServingInputReceiver(
        features=features,
        receiver_tensors=receiver_tensors)

  model.export_savedmodel(FLAGS.export_dir, serving_input_receiver_fn)


if __name__ == '__main__':
  tf.logging.set_verbosity(tf.logging.INFO)
  FLAGS, unparsed = parser.parse_known_args()
  tf.app.run(main=main, argv=[sys.argv[0]] + unparsed)