Support for model conversion to SavedModel and Tensorflow Model Servi…

README.md

@@ -173,6 +173,28 @@ Alternatively, you can run the program interactively by typing image
 paths in the terminal (one per line, type Control-D when you want the
 model to run the input entered so far).
 
+# Deploying model to Tensorflow Model Serving
+
+To host models trained by this platform using Tensorflow Model Serving, 
+you would first need to convert the given checkpoint into the SavedModel 
+format. The SavedModel conversion script (`export_to_saved_model.py`) 
+allows you to restore an existing checkpoint, and export it as the
+SavedModel, which is required format by Tensorflow Model Serving. 
+
+Please edit scripts parameters and execute the following:
+
+    cd src ; python export_to_saved_model.py
+
+# Validation with Tensorflow Model Serving
+
+We have created the example script (`tf_serving_predict.py`) , which 
+will use the specified image file and run it against the model hosted by 
+the Tensorflow Model Serving.
+
+Please edit scripts parameters and execute the following:
+
+    cd src ; python tf_serving_predict.py
+
 # Configuration
 
 There are many command-line options to configure training
@@ -240,7 +262,7 @@ Please cite the following [paper](https://weinman.cs.grinnell.edu/pubs/weinman19
 
 ```text
 @inproceedings{ weinman19deep,
-    author = {Jerod Weinman and Ziwen Chen and Ben Gafford and Nathan Gifford and Abyaya Lamsal and Liam Niehus-Staab},
+    author = {Jerod Weinman and Ziwen Chen and Ben Gafford and Nathan Gifford and Abyaya Lamsal and Liam Niehus-Staab and Igor Vishnevskiy},
     title = {Deep Neural Networks for Text Detection and Recognition in Historical Maps},
     booktitle = {Proc. IAPR International Conference on Document Analysis and Recognition},
     month = {Sep.},

src/export_to_saved_model.py

@@ -0,0 +1,93 @@
+# CNN-LSTM-CTC-OCR
+# Copyright (C) 2017,2018 Jerod Weinman, Abyaya Lamsal, Benjamin Gafford
+#
+# This program is free software: you can redistribute it and/or modify
+# it under the terms of the GNU General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# This program is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU General Public License for more details.
+#
+# You should have received a copy of the GNU General Public License
+# along with this program.  If not, see <http://www.gnu.org/licenses/>.
+
+# validate.py - Run model directly on from paths to image filenames. 
+# NOTE: assumes mjsynth files are given by adding an extra row of padding
+
+import tensorflow as tf
+
+import model_fn as model_fn
+import os
+
+FLAGS = tf.app.flags.FLAGS
+
+tf.app.flags.DEFINE_string( 'ckpt_path','/media/path/to/models/',
+                            """Directory for model checkpoints""" )
+tf.app.flags.DEFINE_string( 'checkpoint','model.ckpt-73152',
+                            """The checkpoint to export. Just the name of the checkpoint, 
+                            no path needed. Exemple: 'model.ckpt-104976'""" )
+tf.app.flags.DEFINE_string( 'export_dir','/media/path/to/where/to/export/saved_model/',
+                            """Path to the directory where SavedModel is to be exported""" )
+tf.app.flags.DEFINE_string( 'lexicon','',
+                            """File containing lexicon of image words""" )
+tf.app.flags.DEFINE_float( 'lexicon_prior',None,
+                           """Prior bias [0,1] for lexicon word""" )
+
+IMAGE = 'image'
+WIDTH = 'width'
+LABELS = 'labels'
+LENGTH = 'length'
+TEXT = 'text'
+
+
+class ExportSavedModel():
+
+    def __get_config(self):
+        """Setup session config to soften device placement"""
+
+        device_config=tf.ConfigProto(
+            allow_soft_placement=True,
+            log_device_placement=False )
+
+        custom_config = tf.estimator.RunConfig( session_config=device_config )
+
+        return custom_config
+
+    def __serving_input_receiver(self):
+        """Placeholders function required by the Estimator to export SavedModel"""
+
+        image = tf.placeholder(dtype=tf.float32, shape=[1, 32, None, 1], name=IMAGE)
+        width = tf.placeholder(dtype=tf.int32, shape=[], name=WIDTH)
+        labels = tf.placeholder(dtype=tf.float32, shape=[None], name=LABELS)
+        length = tf.placeholder(dtype=tf.float32, shape=[], name=LENGTH)
+        text = tf.placeholder(dtype=tf.string, shape=[], name=TEXT)
+        features = {IMAGE: image,
+                    WIDTH: width,
+                    LABELS: labels,
+                    LENGTH: length,
+                    TEXT: text}
+        receiver_tensor = {IMAGE: image,
+                           WIDTH: width}
+        return tf.estimator.export.ServingInputReceiver(features, receiver_tensor)
+
+    def export_saved_model(self):
+        """Will restore the given checkpoint and export it into the SavedModel format"""
+
+        classifier = tf.estimator.Estimator(config=self.__get_config(),
+                                            model_fn=model_fn.predict_fn(
+                                                 FLAGS.lexicon,
+                                                 FLAGS.lexicon_prior),
+                                            model_dir=FLAGS.ckpt_path)
+
+        checkpoint = os.path.join(FLAGS.ckpt_path, FLAGS.checkpoint)
+
+        classifier.export_saved_model(FLAGS.export_dir,
+                                      serving_input_receiver_fn=self.__serving_input_receiver,
+                                      checkpoint_path=checkpoint)
+
+if __name__ == '__main__':
+    export = ExportSavedModel()
+    export.export_saved_model()

src/tf_serving_predict.py

@@ -0,0 +1,107 @@
+# CNN-LSTM-CTC-OCR
+# Copyright (C) 2017,2018 Jerod Weinman, Abyaya Lamsal, Benjamin Gafford
+#
+# This program is free software: you can redistribute it and/or modify
+# it under the terms of the GNU General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# This program is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU General Public License for more details.
+#
+# You should have received a copy of the GNU General Public License
+# along with this program.  If not, see <http://www.gnu.org/licenses/>.
+
+# validate.py - Run model directly on from paths to image filenames. 
+# NOTE: assumes mjsynth files are given by adding an extra row of padding
+
+import numpy as np
+from PIL import Image
+
+from tensorflow import saved_model as sm
+
+import charset as charset
+import mjsynth
+
+import datetime
+
+import grpc
+import tensorflow as tf
+from tensorflow_serving.apis import predict_pb2, prediction_service_pb2_grpc
+
+global ctc_graph
+ctc_graph = tf.compat.v1.get_default_graph
+
+FLAGS = tf.app.flags.FLAGS
+
+tf.app.flags.DEFINE_string( 'server_host','localhost:8500',
+                            """IP:PORT of the Tensorflow Model Serving""" )
+tf.app.flags.DEFINE_string( 'model_spec_name','crnn',
+                            """Name of the saved models batch as it is referenced in the "tf_server.conf""" )
+tf.app.flags.DEFINE_integer( 'model_version',1616119579,
+                            """SavedModel version to call""" )
+tf.app.flags.DEFINE_string( 'image_path','/media/test/images/test_image.png',
+                            """Image to be used as input for inference/prediction""" )
+tf.app.flags.DEFINE_integer( 'image_base_height',32,
+                            """The height for all tensors to be reshaped to""" )
+
+tf.logging.set_verbosity( tf.logging.INFO )
+
+server = FLAGS.server_host
+channel = grpc.insecure_channel(server)
+stub = prediction_service_pb2_grpc.PredictionServiceStub(channel)
+
+request = predict_pb2.PredictRequest()
+request.model_spec.signature_name = sm.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY
+request.model_spec.name = FLAGS.model_spec_name
+request.model_spec.version.value = FLAGS.model_version
+
+IMAGE = 'image'
+WIDTH = 'width'
+LABELS = 'labels'
+
+class Predict():
+
+    def __get_image(self, filename):
+        """Load image data for placement in graph"""
+
+        pil_image = Image.open(filename)
+        width, height = pil_image.size
+        new_width = FLAGS.image_base_height * width / height
+        pil_image = pil_image.resize((int(new_width), FLAGS.image_base_height), Image.ANTIALIAS)
+        image = np.array(pil_image)
+
+        # in mjsynth, all three channels are the same in these grayscale-cum-RGB data
+        image = image[:, :, :1]  # so just extract first channel, preserving 3D shape
+        image = mjsynth.preprocess_image(image)
+        image = image.eval(session=tf.compat.v1.Session())
+
+        return image
+
+    def predict(self):
+        """ Will call Tensorflow Model Server for Prediction, using gRPC API. """
+
+        print("Prediction Start Time: ", datetime.datetime.now())
+        image = self.__get_image(FLAGS.image_path)
+
+        h, w, c = image.shape
+        image = image[np.newaxis, :, :, :]
+
+        request.inputs[IMAGE].CopyFrom(tf.compat.v1.make_tensor_proto(image, dtype=None, shape=image.shape))
+        request.inputs[WIDTH].CopyFrom(tf.compat.v1.make_tensor_proto(w))
+        result = stub.Predict(request, 10)
+
+        protobuf_response = result.outputs[LABELS]
+        ndarray_response = tf.make_ndarray(protobuf_response)
+        label = charset.label_to_string(ndarray_response[0])
+
+        print("Prediction End Time: ", datetime.datetime.now())
+        print("Predicted Text: ", label)
+
+        return label
+
+if __name__ == '__main__':
+    pr = Predict()
+    pr.predict()