4

Description

I'm trying to deploy one of the pre-trained TensorFlow object detection models (faster_rcnn_inception_v2_coco_2018_01_28) in TF Serving. I follow these steps:

  1. Clone https://github.com/tensorflow/models
  2. Download pre-trained Inception model checkpoint
  3. Export Inception as described here with this command. Note I'm changing input_type to encoded_image_string_tensor so that I won't need to attach an input tensor to de-serialize the input string within the API.
  4. Convert Inception from the previous step to a servable with the code below
  5. Run TF Serving, pointing model_base_path to the folder created by the previous step.

Issue

When querying the API predict endpoint it fails, it seems the model has not been initialized properly:

{ "error": "Attempting to use uninitialized value SecondStageFeatureExtractor/InceptionV2/Mixed_5c/Branch_3/Conv2d_0b_1x1/BatchNorm/moving_mean\\n\\t [[Node: SecondStageFeatureExtractor/InceptionV2/Mixed_5c/Branch_3/Conv2d_0b_1x1/BatchNorm/moving_mean/read = Identity[T=DT_FLOAT, _output_shapes=[[128]], _device=\\"/job:localhost/replica:0/task:0/device:CPU:0\\"](SecondStageFeatureExtractor/InceptionV2/Mixed_5c/Branch_3/Conv2d_0b_1x1/BatchNorm/moving_mean)]]" }

Indeed the TF Serving log warned The specified SavedModel has no variables; no checkpoints were restored. (see Appendix for Step 5).

What am I missing in the model export phase? From older issues (this and this), I see this might be due to step (3) freezing variables as constants into the saved model. Is this the case?

Appendix for Step 3 - Re-export Inception

python object_detection/export_inference_graph.py \
  --input_type encoded_image_string_tensor \
  --pipeline_config_path ${MODEL_DIR}/pipeline.config   \
  --trained_checkpoint_prefix ${MODEL_DIR}/model.ckpt   \
  --inference_graph_path ${MODEL_DIR} \
  --export_as_saved_model=True \
  --write_inference_graph=True  \
  --output_directory ${OUTPUT_DIR}

This command produces this model

MetaGraphDef with tag-set: 'serve' contains the following SignatureDefs:

signature_def['serving_default']:
  The given SavedModel SignatureDef contains the following input(s):
    inputs['inputs'] tensor_info:
        dtype: DT_STRING
        shape: (-1)
        name: encoded_image_string_tensor:0
  The given SavedModel SignatureDef contains the following output(s):
    outputs['detection_boxes'] tensor_info:
        dtype: DT_FLOAT
        shape: (-1, 100, 4)
        name: detection_boxes:0
    outputs['detection_classes'] tensor_info:
        dtype: DT_FLOAT
        shape: (-1, 100)
        name: detection_classes:0
    outputs['detection_scores'] tensor_info:
        dtype: DT_FLOAT
        shape: (-1, 100)
        name: detection_scores:0
    outputs['num_detections'] tensor_info:
        dtype: DT_FLOAT
        shape: (-1)
        name: num_detections:0
  Method name is: tensorflow/serving/predict

Appendix for Step 4 - Saved model to servable

import os
import shutil
import tensorflow as tf

tf.app.flags.DEFINE_string('checkpoint_dir', '/tmp/inception_train',
                           """Directory where to read training checkpoints.""")
tf.app.flags.DEFINE_string('output_dir', '/tmp/faster_rcnn_inception_v2_coco_2018_01_28-export/',
                           """Directory where to export inference model.""")
tf.app.flags.DEFINE_integer('model_version', 1,
                            """Version number of the model.""")
tf.app.flags.DEFINE_string('summaries_dir', '/tmp/tensorboard_data',
                            """Directory where to store tensorboard data.""")

FLAGS = tf.app.flags.FLAGS


def main(_):
     with tf.Graph().as_default() as graph:
        saver = tf.train.import_meta_graph(meta_graph_or_file=os.path.join(FLAGS.checkpoint_dir, 'model.ckpt.meta'))

        with tf.Session(graph=graph) as sess:
            saver.restore(sess, tf.train.latest_checkpoint(FLAGS.checkpoint_dir))
            # (re-)create export directory
            export_path = os.path.join(
                tf.compat.as_bytes(FLAGS.output_dir),
                tf.compat.as_bytes(str(FLAGS.model_version)))
            if os.path.exists(export_path):
                shutil.rmtree(export_path)

            tf.global_variables_initializer().run()
            tf.local_variables_initializer().run()
            print("tf.global_variables()")
            print(sess.run(tf.global_variables()))
            print("tf.local_variables()")
            print(sess.run(tf.local_variables()))

            # create model builder
            builder = tf.saved_model.builder.SavedModelBuilder(export_path)

            # Build the signature_def_map.
            predict_inputs_tensor_info = tf.saved_model.utils.build_tensor_info(graph.get_tensor_by_name('encoded_image_string_tensor:0'))
            boxes_output_tensor_info = tf.saved_model.utils.build_tensor_info(graph.get_tensor_by_name('detection_boxes:0'))

            prediction_signature = (
                tf.saved_model.signature_def_utils.build_signature_def(
                    inputs={
                        'images': predict_inputs_tensor_info
                    },
                    outputs={
                        'classes': boxes_output_tensor_info
                    },
                    method_name=tf.saved_model.signature_constants.
                        PREDICT_METHOD_NAME
                )
            )

            builder.add_meta_graph_and_variables(sess,
                [tf.saved_model.tag_constants.SERVING],
                signature_def_map={
                    tf.saved_model.signature_constants.
                    DEFAULT_SERVING_SIGNATURE_DEF_KEY:
                        prediction_signature,
                },
                legacy_init_op=None)

            builder.save(as_text=False)

            merged = tf.summary.merge_all()
            train_writer = tf.summary.FileWriter(FLAGS.summaries_dir + '/' + str(FLAGS.model_version),
                                      sess.graph)

        print("Successfully exported Faster RCNN Inception model version '{}' into '{}'".format(
        FLAGS.model_version, FLAGS.output_dir))

if __name__ == '__main__':
    tf.app.run()

This produces the following servable:

MetaGraphDef with tag-set: 'serve' contains the following SignatureDefs:

signature_def['serving_default']:
  The given SavedModel SignatureDef contains the following input(s):
    inputs['images'] tensor_info:
        dtype: DT_STRING
        shape: (-1)
        name: encoded_image_string_tensor:0
  The given SavedModel SignatureDef contains the following output(s):
    outputs['classes'] tensor_info:
        dtype: DT_FLOAT
        shape: unknown_rank
        name: detection_boxes:0
  Method name is: tensorflow/serving/predict

Appendix for Step 5 - docker-compose log for deploying the servable

tf-serving-faster_rcnn_inception | 2018-08-16 16:05:28.191941: I tensorflow_serving/model_servers/main.cc:153] Building single TensorFlow model file config:  model_name: faster_rcnn_inception model_base_path: /tmp/faster_rcnn_inception_v2_coco_2018_01_28_string_input_version-export/
tf-serving-faster_rcnn_inception | 2018-08-16 16:05:28.192341: I tensorflow_serving/model_servers/server_core.cc:459] Adding/updating models.
tf-serving-faster_rcnn_inception | 2018-08-16 16:05:28.192465: I tensorflow_serving/model_servers/server_core.cc:514]  (Re-)adding model: faster_rcnn_inception
tf-serving-faster_rcnn_inception | 2018-08-16 16:05:28.195056: I tensorflow_serving/core/basic_manager.cc:716] Successfully reserved resources to load servable {name: faster_rcnn_inception version: 1}
tf-serving-faster_rcnn_inception | 2018-08-16 16:05:28.195241: I tensorflow_serving/core/loader_harness.cc:66] Approving load for servable version {name: faster_rcnn_inception version: 1}
tf-serving-faster_rcnn_inception | 2018-08-16 16:05:28.195404: I tensorflow_serving/core/loader_harness.cc:74] Loading servable version {name: faster_rcnn_inception version: 1}
tf-serving-faster_rcnn_inception | 2018-08-16 16:05:28.195652: I external/org_tensorflow/tensorflow/contrib/session_bundle/bundle_shim.cc:360] Attempting to load native SavedModelBundle in bundle-shim from: /tmp/faster_rcnn_inception_v2_coco_2018_01_28_string_input_version-export/1
tf-serving-faster_rcnn_inception | 2018-08-16 16:05:28.195829: I external/org_tensorflow/tensorflow/cc/saved_model/loader.cc:242] Loading SavedModel with tags: { serve }; from: /tmp/faster_rcnn_inception_v2_coco_2018_01_28_string_input_version-export/1
tf-serving-base exited with code 0
tf-serving-faster_rcnn_inception | 2018-08-16 16:05:28.313633: I external/org_tensorflow/tensorflow/core/platform/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
tf-serving-faster_rcnn_inception | 2018-08-16 16:05:28.492904: I external/org_tensorflow/tensorflow/cc/saved_model/loader.cc:161] Restoring SavedModel bundle.
tf-serving-faster_rcnn_inception | 2018-08-16 16:05:28.493224: I external/org_tensorflow/tensorflow/cc/saved_model/loader.cc:171] The specified SavedModel has no variables; no checkpoints were restored.
tf-serving-faster_rcnn_inception | 2018-08-16 16:05:28.493329: I external/org_tensorflow/tensorflow/cc/saved_model/loader.cc:196] Running LegacyInitOp on SavedModel bundle.
tf-serving-faster_rcnn_inception | 2018-08-16 16:05:28.512745: I external/org_tensorflow/tensorflow/cc/saved_model/loader.cc:291] SavedModel load for tags { serve }; Status: success. Took 316614 microseconds.
tf-serving-faster_rcnn_inception | 2018-08-16 16:05:28.513085: I tensorflow_serving/servables/tensorflow/saved_model_warmup.cc:83] No warmup data file found at /tmp/faster_rcnn_inception_v2_coco_2018_01_28_string_input_version-export/1/assets.extra/tf_serving_warmup_requests
tf-serving-faster_rcnn_inception | 2018-08-16 16:05:28.513927: I tensorflow_serving/core/loader_harness.cc:86] Successfully loaded servable version {name: faster_rcnn_inception version: 1}
tf-serving-faster_rcnn_inception | 2018-08-16 16:05:28.516875: I tensorflow_serving/model_servers/main.cc:323] Running ModelServer at 0.0.0.0:8500 ...
tf-serving-faster_rcnn_inception | [warn] getaddrinfo: address family for nodename not supported
tf-serving-faster_rcnn_inception | 2018-08-16 16:05:28.518018: I tensorflow_serving/model_servers/main.cc:333] Exporting HTTP/REST API at:localhost:8501 ...
tf-serving-faster_rcnn_inception | [evhttp_server.cc : 235] RAW: Entering the event loop ...

Notes:

  • python 2.7
  • tensorflow 1.10.0 in both steps (3), (4) and (5). Installed via pip.
w00dy
  • 748
  • 1
  • 6
  • 23

1 Answers1

2

Copying this from my answer here

I was also struggling with this, and was able to export a model (include the variables file) by updating @lionel92's suggestion for the current OD API code version (as of July 2, 2019). It mostly involves changing the write_saved_model function in models/research/object_detection/exporter.py

Update write_saved_model in exporter.py

def write_saved_model(saved_model_path,
                      trained_checkpoint_prefix,
                      inputs,
                      outputs):

  saver = tf.train.Saver()
  with tf.Session() as sess:
    saver.restore(sess, trained_checkpoint_prefix)

    builder = tf.saved_model.builder.SavedModelBuilder(saved_model_path)

    tensor_info_inputs = {
        'inputs': tf.saved_model.utils.build_tensor_info(inputs)}
    tensor_info_outputs = {}
    for k, v in outputs.items():
      tensor_info_outputs[k] = tf.saved_model.utils.build_tensor_info(v)

    detection_signature = (
        tf.saved_model.signature_def_utils.build_signature_def(
            inputs=tensor_info_inputs,
            outputs=tensor_info_outputs,
            method_name=tf.saved_model.signature_constants.PREDICT_METHOD_NAME
        ))

    builder.add_meta_graph_and_variables(
        sess,
        [tf.saved_model.tag_constants.SERVING],
        signature_def_map={
            tf.saved_model.signature_constants
            .DEFAULT_SERVING_SIGNATURE_DEF_KEY:
                detection_signature,
        },
    )
    builder.save()

Update _export_inference_graph in exporter.py

Then, within the _export_inference_graph function, update the final line to pass the checkpoint prefix like:

  write_saved_model(saved_model_path, trained_checkpoint_prefix,
                    placeholder_tensor, outputs)

Call export script

Call models/research/object_detection/export_inference_graph.py normally. For me, that looked similar to this:

INPUT_TYPE=encoded_image_string_tensor
PIPELINE_CONFIG_PATH=/path/to/model.config
TRAINED_CKPT_PREFIX=/path/to/model.ckpt-50000
EXPORT_DIR=/path/to/export/dir/001/

python $BUILDS_DIR/models/research/object_detection/export_inference_graph.py \
    --input_type=${INPUT_TYPE} \
    --pipeline_config_path=${PIPELINE_CONFIG_PATH} \
    --trained_checkpoint_prefix=${TRAINED_CKPT_PREFIX} \
    --output_directory=${EXPORT_DIR}

If it works, you should see a directory structure like this. This is ready to be dropped into a TF Serving Docker image for scaled inference.

image

wronk
  • 1,511
  • 2
  • 9
  • 8