attention_ocr: added export for SavedModel format. (#8757)

* Added export for SavedModel format.

* Fixed some pylint errors.

research/attention_ocr/README.md

@@ -166,6 +166,14 @@ implement one in Python or C++.
 
 The recommended way is to use the [Serving infrastructure][serving].
 
+To export to SavedModel format:
+
+```
+python model_export.py \
+  --checkpoint=model.ckpt-399731 \
+  --export_dir=/tmp/attention_ocr_export
+```
+
 Alternatively you can:
 1. define a placeholder for images (or use directly an numpy array)
 2. [create a graph ](https://github.com/tensorflow/models/blob/master/research/attention_ocr/python/eval.py#L60)
@@ -188,7 +196,7 @@ other than a one time experiment please use the [TensorFlow Serving][serving].
 
 [1]: https://github.com/tensorflow/tensorflow/blob/aaf7adc/tensorflow/contrib/rnn/python/tools/checkpoint_convert.py
 [2]: https://www.tensorflow.org/api_docs/python/tf/contrib/framework/assign_from_checkpoint_fn
-[serving]: https://tensorflow.github.io/serving/serving_basic
+[serving]: https://www.tensorflow.org/tfx/serving/serving_basic
 
 ## Disclaimer
 

research/attention_ocr/python/common_flags.py

@@ -14,10 +14,10 @@
 # ==============================================================================
 
 """Define flags are common for both train.py and eval.py scripts."""
+import logging
 import sys
 
 from tensorflow.python.platform import flags
-import logging
 
 import datasets
 import model
@@ -35,9 +35,17 @@ logging.basicConfig(
     datefmt='%Y-%m-%d %H:%M:%S')
 
 
+_common_flags_defined = False
+
 def define():
   """Define common flags."""
   # yapf: disable
+  # common_flags.define() may be called multiple times in unit tests.
+  global _common_flags_defined
+  if _common_flags_defined:
+    return
+  _common_flags_defined = True
+
   flags.DEFINE_integer('batch_size', 32,
                        'Batch size.')
 

research/attention_ocr/python/data_provider.py

@@ -144,9 +144,6 @@ def preprocess_image(image, augment=False, central_crop_size=None,
         images = [augment_image(img) for img in images]
       image = tf.concat(images, 1)
 
-    image = tf.subtract(image, 0.5)
-    image = tf.multiply(image, 2.5)
-
   return image
 
 

research/attention_ocr/python/datasets/fsns.py

@@ -177,6 +177,8 @@ def get_split(split_name, dataset_dir=None, config=None):
       items_to_descriptions=config['items_to_descriptions'],
       #  additional parameters for convenience.
       charset=charset,
+      charset_file=charset_file,
+      image_shape=config['image_shape'],
       num_char_classes=len(charset),
       num_of_views=config['num_of_views'],
       max_sequence_length=config['max_sequence_length'],

research/attention_ocr/python/model.py

@@ -12,7 +12,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 # ==============================================================================
-
 """Functions to build the Attention OCR model.
 
 Usage example:
@@ -26,6 +25,7 @@ Usage example:
 import sys
 import collections
 import logging
+import numpy as np
 import tensorflow as tf
 from tensorflow.contrib import slim
 from tensorflow.contrib.slim.nets import inception
@@ -36,13 +36,13 @@ import utils
 
 OutputEndpoints = collections.namedtuple('OutputEndpoints', [
     'chars_logit', 'chars_log_prob', 'predicted_chars', 'predicted_scores',
-  'predicted_text'
+    'predicted_text', 'predicted_length', 'predicted_conf',
+    'normalized_seq_conf'
 ])
 
 # TODO(gorban): replace with tf.HParams when it is released.
-ModelParams = collections.namedtuple('ModelParams', [
-  'num_char_classes', 'seq_length', 'num_views', 'null_code'
-])
+ModelParams = collections.namedtuple(
+    'ModelParams', ['num_char_classes', 'seq_length', 'num_views', 'null_code'])
 
 ConvTowerParams = collections.namedtuple('ConvTowerParams', ['final_endpoint'])
 
@@ -51,13 +51,12 @@ SequenceLogitsParams = collections.namedtuple('SequenceLogitsParams', [
     'lstm_state_clip_value'
 ])
 
-SequenceLossParams = collections.namedtuple('SequenceLossParams', [
-  'label_smoothing', 'ignore_nulls', 'average_across_timesteps'
-])
+SequenceLossParams = collections.namedtuple(
+    'SequenceLossParams',
+    ['label_smoothing', 'ignore_nulls', 'average_across_timesteps'])
 
-EncodeCoordinatesParams = collections.namedtuple('EncodeCoordinatesParams', [
-  'enabled'
-])
+EncodeCoordinatesParams = collections.namedtuple('EncodeCoordinatesParams',
+                                                 ['enabled'])
 
 
 def _dict_to_array(id_to_char, default_character):
@@ -121,6 +120,142 @@ def get_softmax_loss_fn(label_smoothing):
   return loss_fn
 
 
+def get_tensor_dimensions(tensor):
+  """Returns the shape components of a 4D tensor with variable batch size.
+
+  Args:
+    tensor : A 4D tensor, whose last 3 dimensions are known at graph
+      construction time.
+
+  Returns:
+    batch_size : The first dimension as a tensor object.
+    height : The second dimension as a scalar value.
+    width : The third dimension as a scalar value.
+    num_features : The forth dimension as a scalar value.
+
+  Raises:
+    ValueError: if input tensor does not have 4 dimensions.
+  """
+  if len(tensor.get_shape().dims) != 4:
+    raise ValueError(
+        'Incompatible shape: len(tensor.get_shape().dims) != 4 (%d != 4)' %
+        len(tensor.get_shape().dims))
+  batch_size = tf.shape(tensor)[0]
+  height = tensor.get_shape().dims[1].value
+  width = tensor.get_shape().dims[2].value
+  num_features = tensor.get_shape().dims[3].value
+  return batch_size, height, width, num_features
+
+
+def lookup_indexed_value(indices, row_vecs):
+  """Lookup values in each row of 'row_vecs' indexed by 'indices'.
+
+  For each sample in the batch, look up the element for the corresponding
+  index.
+
+  Args:
+    indices : A tensor of shape (batch, )
+    row_vecs : A tensor of shape [batch, depth]
+
+  Returns:
+    A tensor of shape (batch, ) formed by row_vecs[i, indices[i]].
+  """
+  gather_indices = tf.stack((tf.range(
+      tf.shape(row_vecs)[0], dtype=tf.int32), tf.cast(indices, tf.int32)),
+                            axis=1)
+  return tf.gather_nd(row_vecs, gather_indices)
+
+
+@utils.ConvertAllInputsToTensors
+def max_char_logprob_cumsum(char_log_prob):
+  """Computes the cumulative sum of character logprob for all sequence lengths.
+
+  Args:
+    char_log_prob: A tensor of shape [batch x seq_length x num_char_classes]
+      with log probabilities of a character.
+
+  Returns:
+    A tensor of shape [batch x (seq_length+1)] where each element x[_, j] is
+    the sum of the max char logprob for all positions upto j.
+    Note this duplicates the final column and produces (seq_length+1) columns
+    so the same function can be used regardless whether use_length_predictions
+    is true or false.
+  """
+  max_char_log_prob = tf.reduce_max(char_log_prob, reduction_indices=2)
+  # For an input array [a, b, c]) tf.cumsum returns [a, a + b, a + b + c] if
+  # exclusive set to False (default).
+  return tf.cumsum(max_char_log_prob, axis=1, exclusive=False)
+
+
+def find_length_by_null(predicted_chars, null_code):
+  """Determine sequence length by finding null_code among predicted char IDs.
+
+  Given the char class ID for each position, compute the sequence length.
+  Note that this function computes this based on the number of null_code,
+  instead of the position of the first null_code.
+
+  Args:
+    predicted_chars: A tensor of [batch x seq_length] where each element stores
+      the char class ID with max probability;
+    null_code: an int32, character id for the NULL.
+
+  Returns:
+    A [batch, ] tensor which stores the sequence length for each sample.
+  """
+  return tf.reduce_sum(
+      tf.cast(tf.not_equal(null_code, predicted_chars), tf.int32), axis=1)
+
+
+def axis_pad(tensor, axis, before=0, after=0, constant_values=0.0):
+  """Pad a tensor with the specified values along a single axis.
+
+  Args:
+    tensor: a Tensor;
+    axis: the dimension to add pad along to;
+    before: number of values to add before the contents of tensor in the
+      selected dimension;
+    after: number of values to add after the contents of tensor in the selected
+      dimension;
+    constant_values: the scalar pad value to use. Must be same type as tensor.
+
+  Returns:
+    A Tensor. Has the same type as the input tensor, but with a changed shape
+    along the specified dimension.
+  """
+  if before == 0 and after == 0:
+    return tensor
+  ndims = tensor.shape.ndims
+  padding_size = np.zeros((ndims, 2), dtype='int32')
+  padding_size[axis] = before, after
+  return tf.pad(
+      tensor=tensor,
+      paddings=tf.constant(padding_size),
+      constant_values=constant_values)
+
+
+def null_based_length_prediction(chars_log_prob, null_code):
+  """Computes length and confidence of prediction based on positions of NULLs.
+
+  Args:
+    chars_log_prob: A tensor of shape [batch x seq_length x num_char_classes]
+      with log probabilities of a character;
+    null_code: an int32, character id for the NULL.
+
+  Returns:
+    A tuple (text_log_prob, predicted_length), where
+    text_log_prob - is a tensor of the same shape as length_log_prob.
+    Element #0 of the output corresponds to probability of the empty string,
+    element #seq_length - is the probability of length=seq_length.
+    predicted_length is a tensor with shape [batch].
+  """
+  predicted_chars = tf.to_int32(tf.argmax(chars_log_prob, axis=2))
+  # We do right pad to support sequences with seq_length elements.
+  text_log_prob = max_char_logprob_cumsum(
+      axis_pad(chars_log_prob, axis=1, after=1))
+  predicted_length = find_length_by_null(predicted_chars, null_code)
+  return text_log_prob, predicted_length
+
+
 class Model(object):
   """Class to create the Attention OCR Model."""
 
@@ -137,13 +272,13 @@ class Model(object):
       num_char_classes: size of character set.
       seq_length: number of characters in a sequence.
       num_views: Number of views (conv towers) to use.
-      null_code: A character code corresponding to a character which
-        indicates end of a sequence.
-      mparams: a dictionary with hyper parameters for methods,  keys -
-        function names, values - corresponding namedtuples.
+      null_code: A character code corresponding to a character which indicates
+        end of a sequence.
+      mparams: a dictionary with hyper parameters for methods,  keys - function
+        names, values - corresponding namedtuples.
       charset: an optional dictionary with a mapping between character ids and
-        utf8 strings. If specified the OutputEndpoints.predicted_text will
-        utf8 encoded strings corresponding to the character ids returned by
+        utf8 strings. If specified the OutputEndpoints.predicted_text will utf8
+        encoded strings corresponding to the character ids returned by
         OutputEndpoints.predicted_chars (by default the predicted_text contains
         an empty vector).
         NOTE: Make sure you call tf.tables_initializer().run() if the charset
@@ -176,7 +311,8 @@ class Model(object):
                 label_smoothing=0.1,
                 ignore_nulls=True,
                 average_across_timesteps=False),
-      'encode_coordinates_fn': EncodeCoordinatesParams(enabled=False)
+        'encode_coordinates_fn':
+            EncodeCoordinatesParams(enabled=False)
     }
 
   def set_mparam(self, function, **kwargs):
@@ -222,10 +358,10 @@ class Model(object):
     """
     num_features = net.get_shape().dims[1].value
     if num_features < self._params.seq_length:
-      raise AssertionError('Incorrect dimension #1 of input tensor'
+      raise AssertionError(
+          'Incorrect dimension #1 of input tensor'
           ' %d should be bigger than %d (shape=%s)' %
-                           (num_features, self._params.seq_length,
-                            net.get_shape()))
+          (num_features, self._params.seq_length, net.get_shape()))
     elif num_features > self._params.seq_length:
       logging.warning('Ignoring some features: use %d of %d (shape=%s)',
                       self._params.seq_length, num_features, net.get_shape())
@@ -279,16 +415,17 @@ class Model(object):
     """
     with tf.variable_scope('pool_views_fn/STCK'):
       net = tf.concat(nets, 1)
-      batch_size = net.get_shape().dims[0].value
+      batch_size = tf.shape(net)[0]
+      image_size = net.get_shape().dims[1].value * net.get_shape().dims[2].value
       feature_size = net.get_shape().dims[3].value
-      return tf.reshape(net, [batch_size, -1, feature_size])
+      return tf.reshape(net, tf.stack([batch_size, image_size, feature_size]))
 
   def char_predictions(self, chars_logit):
     """Returns confidence scores (softmax values) for predicted characters.
 
     Args:
-      chars_logit: chars logits, a tensor with shape
-        [batch_size x seq_length x num_char_classes]
+      chars_logit: chars logits, a tensor with shape [batch_size x seq_length x
+        num_char_classes]
 
     Returns:
       A tuple (ids, log_prob, scores), where:
@@ -306,7 +443,10 @@ class Model(object):
         slim.one_hot_encoding(ids, self._params.num_char_classes), tf.bool)
     all_scores = tf.nn.softmax(chars_logit)
     selected_scores = tf.boolean_mask(all_scores, mask, name='char_scores')
-    scores = tf.reshape(selected_scores, shape=(-1, self._params.seq_length))
+    scores = tf.reshape(
+        selected_scores,
+        shape=(-1, self._params.seq_length),
+        name='predicted_scores')
     return ids, log_prob, scores
 
   def encode_coordinates_fn(self, net):
@@ -323,12 +463,12 @@ class Model(object):
     """
     mparams = self._mparams['encode_coordinates_fn']
     if mparams.enabled:
-      batch_size, h, w, _ = net.shape.as_list()
+      batch_size, h, w, _ = get_tensor_dimensions(net)
       x, y = tf.meshgrid(tf.range(w), tf.range(h))
       w_loc = slim.one_hot_encoding(x, num_classes=w)
       h_loc = slim.one_hot_encoding(y, num_classes=h)
       loc = tf.concat([h_loc, w_loc], 2)
-      loc = tf.tile(tf.expand_dims(loc, 0), [batch_size, 1, 1, 1])
+      loc = tf.tile(tf.expand_dims(loc, 0), tf.stack([batch_size, 1, 1, 1]))
       return tf.concat([net, loc], 3)
     else:
       return net
@@ -341,7 +481,8 @@ class Model(object):
     """Creates a base part of the Model (no gradients, losses or summaries).
 
     Args:
-      images: A tensor of shape [batch_size, height, width, channels].
+      images: A tensor of shape [batch_size, height, width, channels] with pixel
+        values in the range [0.0, 1.0].
       labels_one_hot: Optional (can be None) one-hot encoding for ground truth
         labels. If provided the function will create a model for training.
       scope: Optional variable_scope.
@@ -353,6 +494,11 @@ class Model(object):
     """
     logging.debug('images: %s', images)
     is_training = labels_one_hot is not None
+
+    # Normalize image pixel values to have a symmetrical range around zero.
+    images = tf.subtract(images, 0.5)
+    images = tf.multiply(images, 2.5)
+
     with tf.variable_scope(scope, reuse=reuse):
       views = tf.split(
           value=images, num_or_size_splits=self._params.num_views, axis=2)
@@ -380,12 +526,28 @@ class Model(object):
         predicted_text = character_mapper.get_text(predicted_chars)
       else:
         predicted_text = tf.constant([])
+
+      text_log_prob, predicted_length = null_based_length_prediction(
+          chars_log_prob, self._params.null_code)
+      predicted_conf = lookup_indexed_value(predicted_length, text_log_prob)
+      # Convert predicted confidence from sum of logs to geometric mean
+      normalized_seq_conf = tf.exp(
+          tf.divide(predicted_conf,
+                    tf.cast(predicted_length + 1, predicted_conf.dtype)),
+          name='normalized_seq_conf')
+      predicted_conf = tf.identity(predicted_conf, name='predicted_conf')
+      predicted_text = tf.identity(predicted_text, name='predicted_text')
+      predicted_length = tf.identity(predicted_length, name='predicted_length')
+
     return OutputEndpoints(
         chars_logit=chars_logit,
         chars_log_prob=chars_log_prob,
         predicted_chars=predicted_chars,
         predicted_scores=predicted_scores,
-      predicted_text=predicted_text)
+        predicted_length=predicted_length,
+        predicted_text=predicted_text,
+        predicted_conf=predicted_conf,
+        normalized_seq_conf=normalized_seq_conf)
 
   def create_loss(self, data, endpoints):
     """Creates all losses required to train the model.
@@ -413,8 +575,8 @@ class Model(object):
     Uses the same method as in https://arxiv.org/abs/1512.00567.
 
     Args:
-      chars_labels: ground truth ids of charactes,
-        shape=[batch_size, seq_length];
+      chars_labels: ground truth ids of charactes, shape=[batch_size,
+        seq_length];
       weight: label-smoothing regularization weight.
 
     Returns:
@@ -433,10 +595,10 @@ class Model(object):
     also ignore all null chars after the first one.
 
     Args:
-      chars_logits: logits for predicted characters,
-        shape=[batch_size, seq_length, num_char_classes];
-      chars_labels: ground truth ids of characters,
-        shape=[batch_size, seq_length];
+      chars_logits: logits for predicted characters, shape=[batch_size,
+        seq_length, num_char_classes];
+      chars_labels: ground truth ids of characters, shape=[batch_size,
+        seq_length];
       mparams: method hyper parameters.
 
     Returns:
@@ -482,8 +644,8 @@ class Model(object):
     Args:
       data: InputEndpoints namedtuple.
       endpoints: OutputEndpoints namedtuple.
-      charset: A dictionary with mapping between character codes and
-        unicode characters. Use the one provided by a dataset.charset.
+      charset: A dictionary with mapping between character codes and unicode
+        characters. Use the one provided by a dataset.charset.
       is_training: If True will create summary prefixes for training job,
         otherwise - for evaluation.
 
@@ -520,14 +682,16 @@ class Model(object):
         names_to_values[name] = value_update_tuple[0]
         names_to_updates[name] = value_update_tuple[1]
 
-      use_metric('CharacterAccuracy',
+      use_metric(
+          'CharacterAccuracy',
           metrics.char_accuracy(
               endpoints.predicted_chars,
               data.labels,
               streaming=True,
               rej_char=self._params.null_code))
       # Sequence accuracy computed by cutting sequence at the first null char
-      use_metric('SequenceAccuracy',
+      use_metric(
+          'SequenceAccuracy',
           metrics.sequence_accuracy(
               endpoints.predicted_chars,
               data.labels,
@@ -539,13 +703,14 @@ class Model(object):
         tf.summary.scalar(summary_name, tf.Print(value, [value], summary_name))
       return list(names_to_updates.values())
 
-  def create_init_fn_to_restore(self, master_checkpoint,
+  def create_init_fn_to_restore(self,
+                                master_checkpoint,
                                 inception_checkpoint=None):
     """Creates an init operations to restore weights from various checkpoints.
 
     Args:
-      master_checkpoint: path to a checkpoint which contains all weights for
-        the whole model.
+      master_checkpoint: path to a checkpoint which contains all weights for the
+        whole model.
       inception_checkpoint: path to a checkpoint which contains weights for the
         inception part only.
 
@@ -556,8 +721,8 @@ class Model(object):
     all_feed_dict = {}
 
     def assign_from_checkpoint(variables, checkpoint):
-      logging.info('Request to re-store %d weights from %s',
-                   len(variables), checkpoint)
+      logging.info('Request to re-store %d weights from %s', len(variables),
+                   checkpoint)
       if not variables:
         logging.error('Can\'t find any variables to restore.')
         sys.exit(1)
@@ -565,9 +730,12 @@ class Model(object):
       all_assign_ops.append(assign_op)
       all_feed_dict.update(feed_dict)
 
-    logging.info('variables_to_restore:\n%s' % utils.variables_to_restore().keys())
-    logging.info('moving_average_variables:\n%s' % [v.op.name for v in tf.moving_average_variables()])
-    logging.info('trainable_variables:\n%s' % [v.op.name for v in tf.trainable_variables()])
+    logging.info('variables_to_restore:\n%s',
+                 utils.variables_to_restore().keys())
+    logging.info('moving_average_variables:\n%s',
+                 [v.op.name for v in tf.moving_average_variables()])
+    logging.info('trainable_variables:\n%s',
+                 [v.op.name for v in tf.trainable_variables()])
     if master_checkpoint:
       assign_from_checkpoint(utils.variables_to_restore(), master_checkpoint)
 

research/attention_ocr/python/model_export.py

+# Lint as: python3
+# Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+#
+# 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.
+# ==============================================================================
+r"""Converts existing checkpoint into a SavedModel.
+
+Usage example:
+python model_export.py \
+  --logtostderr --checkpoint=model.ckpt-399731 \
+  --export_dir=/tmp/attention_ocr_export
+"""
+import os
+
+import tensorflow as tf
+from tensorflow import app
+from tensorflow.contrib import slim
+from tensorflow.python.platform import flags
+
+import common_flags
+import model_export_lib
+
+FLAGS = flags.FLAGS
+common_flags.define()
+
+flags.DEFINE_string('export_dir', None, 'Directory to export model files to.')
+flags.DEFINE_integer(
+    'image_width', None,
+    'Image width used during training (or crop width if used)'
+    ' If not set, the dataset default is used instead.')
+flags.DEFINE_integer(
+    'image_height', None,
+    'Image height used during training(or crop height if used)'
+    ' If not set, the dataset default is used instead.')
+flags.DEFINE_string('work_dir', '/tmp', 'A directory to store temporary files.')
+flags.DEFINE_integer('version_number', 1, 'Version number of the model')
+flags.DEFINE_bool(
+    'export_for_serving', True,
+    'Whether the exported model accepts serialized tf.Example '
+    'protos as input')
+
+
+def get_checkpoint_path():
+  """Returns a path to a checkpoint based on specified commandline flags.
+
+  In order to specify a full path to a checkpoint use --checkpoint flag.
+  Alternatively, if --train_log_dir was specified it will return a path to the
+  most recent checkpoint.
+
+  Raises:
+    ValueError: in case it can't find a checkpoint.
+
+  Returns:
+    A string.
+  """
+  if FLAGS.checkpoint:
+    return FLAGS.checkpoint
+  else:
+    model_save_path = tf.train.latest_checkpoint(FLAGS.train_log_dir)
+    if not model_save_path:
+      raise ValueError('Can\'t find a checkpoint in: %s' % FLAGS.train_log_dir)
+    return model_save_path
+
+
+def export_model(export_dir,
+                 export_for_serving,
+                 batch_size=None,
+                 crop_image_width=None,
+                 crop_image_height=None):
+  """Exports a model to the named directory.
+
+  Note that --datatset_name and --checkpoint are required and parsed by the
+  underlying module common_flags.
+
+  Args:
+    export_dir: The output dir where model is exported to.
+    export_for_serving: If True, expects a serialized image as input and attach
+      image normalization as part of exported graph.
+    batch_size: For non-serving export, the input batch_size needs to be
+      specified.
+    crop_image_width: Width of the input image. Uses the dataset default if
+      None.
+    crop_image_height: Height of the input image. Uses the dataset default if
+      None.
+
+  Returns:
+    Returns the model signature_def.
+  """
+  # Dataset object used only to get all parameters for the model.
+  dataset = common_flags.create_dataset(split_name='test')
+  model = common_flags.create_model(
+      dataset.num_char_classes,
+      dataset.max_sequence_length,
+      dataset.num_of_views,
+      dataset.null_code,
+      charset=dataset.charset)
+  dataset_image_height, dataset_image_width, image_depth = dataset.image_shape
+
+  # Add check for charmap file
+  if not os.path.exists(dataset.charset_file):
+    raise ValueError('No charset defined at {}: export will fail'.format(
+        dataset.charset))
+
+  # Default to dataset dimensions, otherwise use provided dimensions.
+  image_width = crop_image_width or dataset_image_width
+  image_height = crop_image_height or dataset_image_height
+
+  if export_for_serving:
+    images_orig = tf.placeholder(
+        tf.string, shape=[batch_size], name='tf_example')
+    images_orig_float = model_export_lib.generate_tfexample_image(
+        images_orig,
+        image_height,
+        image_width,
+        image_depth,
+        name='float_images')
+  else:
+    images_shape = (batch_size, image_height, image_width, image_depth)
+    images_orig = tf.placeholder(
+        tf.uint8, shape=images_shape, name='original_image')
+    images_orig_float = tf.image.convert_image_dtype(
+        images_orig, dtype=tf.float32, name='float_images')
+
+  endpoints = model.create_base(images_orig_float, labels_one_hot=None)
+
+  sess = tf.Session()
+  saver = tf.train.Saver(slim.get_variables_to_restore(), sharded=True)
+  saver.restore(sess, get_checkpoint_path())
+  tf.logging.info('Model restored successfully.')
+
+  # Create model signature.
+  if export_for_serving:
+    input_tensors = {
+        tf.saved_model.signature_constants.CLASSIFY_INPUTS: images_orig
+    }
+  else:
+    input_tensors = {'images': images_orig}
+  signature_inputs = model_export_lib.build_tensor_info(input_tensors)
+  # NOTE: Tensors 'image_float' and 'chars_logit' are used by the inference
+  # or to compute saliency maps.
+  output_tensors = {
+      'images_float': images_orig_float,
+      'predictions': endpoints.predicted_chars,
+      'scores': endpoints.predicted_scores,
+      'chars_logit': endpoints.chars_logit,
+      'predicted_length': endpoints.predicted_length,
+      'predicted_text': endpoints.predicted_text,
+      'predicted_conf': endpoints.predicted_conf,
+      'normalized_seq_conf': endpoints.normalized_seq_conf
+  }
+  for i, t in enumerate(
+      model_export_lib.attention_ocr_attention_masks(
+          dataset.max_sequence_length)):
+    output_tensors['attention_mask_%d' % i] = t
+  signature_outputs = model_export_lib.build_tensor_info(output_tensors)
+  signature_def = tf.saved_model.signature_def_utils.build_signature_def(
+      signature_inputs, signature_outputs,
+      tf.saved_model.signature_constants.CLASSIFY_METHOD_NAME)
+  # Save model.
+  builder = tf.saved_model.builder.SavedModelBuilder(export_dir)
+  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:
+              signature_def
+      },
+      main_op=tf.tables_initializer(),
+      strip_default_attrs=True)
+  builder.save()
+  tf.logging.info('Model has been exported to %s' % export_dir)
+
+  return signature_def
+
+
+def main(unused_argv):
+  if os.path.exists(FLAGS.export_dir):
+    raise ValueError('export_dir already exists: exporting will fail')
+
+  export_model(FLAGS.export_dir, FLAGS.export_for_serving, FLAGS.batch_size,
+               FLAGS.image_width, FLAGS.image_height)
+
+
+if __name__ == '__main__':
+  flags.mark_flag_as_required('dataset_name')
+  flags.mark_flag_as_required('export_dir')
+  app.run(main)

research/attention_ocr/python/model_export_lib.py

+# Lint as: python3
+# Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+#
+# 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.
+# ==============================================================================
+"""Utility functions for exporting Attention OCR model."""
+
+import tensorflow as tf
+
+
+# Function borrowed from research/object_detection/core/preprocessor.py
+def normalize_image(image, original_minval, original_maxval, target_minval,
+                    target_maxval):
+  """Normalizes pixel values in the image.
+
+  Moves the pixel values from the current [original_minval, original_maxval]
+  range to a the [target_minval, target_maxval] range.
+  Args:
+    image: rank 3 float32 tensor containing 1 image -> [height, width,
+      channels].
+    original_minval: current image minimum value.
+    original_maxval: current image maximum value.
+    target_minval: target image minimum value.
+    target_maxval: target image maximum value.
+
+  Returns:
+    image: image which is the same shape as input image.
+  """
+  with tf.name_scope('NormalizeImage', values=[image]):
+    original_minval = float(original_minval)
+    original_maxval = float(original_maxval)
+    target_minval = float(target_minval)
+    target_maxval = float(target_maxval)
+    image = tf.cast(image, dtype=tf.float32)
+    image = tf.subtract(image, original_minval)
+    image = tf.multiply(image, (target_maxval - target_minval) /
+                        (original_maxval - original_minval))
+    image = tf.add(image, target_minval)
+    return image
+
+
+def generate_tfexample_image(input_example_strings,
+                             image_height,
+                             image_width,
+                             image_channels,
+                             name=None):
+  """Parses a 1D tensor of serialized tf.Example protos and returns image batch.
+
+  Args:
+    input_example_strings: A 1-Dimensional tensor of size [batch_size] and type
+      tf.string containing a serialized Example proto per image.
+    image_height: First image dimension.
+    image_width: Second image dimension.
+    image_channels: Third image dimension.
+    name: optional tensor name.
+
+  Returns:
+    A tensor with shape [batch_size, height, width, channels] of type float32
+    with values in the range [0..1]
+  """
+  batch_size = tf.shape(input_example_strings)[0]
+  images_shape = tf.stack(
+      [batch_size, image_height, image_width, image_channels])
+  tf_example_image_key = 'image/encoded'
+  feature_configs = {
+      tf_example_image_key:
+          tf.FixedLenFeature(
+              image_height * image_width * image_channels, dtype=tf.float32)
+  }
+  feature_tensors = tf.parse_example(input_example_strings, feature_configs)
+  float_images = tf.reshape(
+      normalize_image(
+          feature_tensors[tf_example_image_key],
+          original_minval=0.0,
+          original_maxval=255.0,
+          target_minval=0.0,
+          target_maxval=1.0),
+      images_shape,
+      name=name)
+  return float_images
+
+
+def attention_ocr_attention_masks(num_characters):
+  # TODO(gorban): use tensors directly after replacing LSTM unroll methods.
+  prefix = ('AttentionOcr_v1/'
+            'sequence_logit_fn/SQLR/LSTM/attention_decoder/Attention_0')
+  names = ['%s/Softmax:0' % (prefix)]
+  for i in range(1, num_characters):
+    names += ['%s_%d/Softmax:0' % (prefix, i)]
+  return [tf.get_default_graph().get_tensor_by_name(n) for n in names]
+
+
+def build_tensor_info(tensor_dict):
+  return {
+      k: tf.saved_model.utils.build_tensor_info(t)
+      for k, t in tensor_dict.items()
+  }

research/attention_ocr/python/model_export_test.py

+# Lint as: python3
+# Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+#
+# 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.
+# ==============================================================================
+"""Tests for model_export."""
+import os
+
+import numpy as np
+from absl.testing import flagsaver
+import tensorflow as tf
+
+import common_flags
+import model_export
+
+_CHECKPOINT = 'model.ckpt-399731'
+_CHECKPOINT_URL = (
+    'http://download.tensorflow.org/models/attention_ocr_2017_08_09.tar.gz')
+
+
+def _clean_up():
+  tf.gfile.DeleteRecursively(tf.test.get_temp_dir())
+
+
+def _create_tf_example_string(image):
+  """Create a serialized tf.Example proto for feeding the model."""
+  example = tf.train.Example()
+  example.features.feature['image/encoded'].float_list.value.extend(
+      list(np.reshape(image, (-1))))
+  return example.SerializeToString()
+
+
+class AttentionOcrExportTest(tf.test.TestCase):
+  """Tests for model_export.export_model."""
+
+  def setUp(self):
+    for suffix in ['.meta', '.index', '.data-00000-of-00001']:
+      filename = _CHECKPOINT + suffix
+      self.assertTrue(
+          tf.gfile.Exists(filename),
+          msg='Missing checkpoint file %s. '
+          'Please download and extract it from %s' %
+          (filename, _CHECKPOINT_URL))
+    tf.flags.FLAGS.dataset_name = 'fsns'
+    tf.flags.FLAGS.checkpoint = _CHECKPOINT
+    tf.flags.FLAGS.dataset_dir = os.path.join(
+        os.path.dirname(__file__), 'datasets/testdata/fsns')
+    tf.test.TestCase.setUp(self)
+    _clean_up()
+    self.export_dir = os.path.join(tf.test.get_temp_dir(), 'exported_model')
+    self.minimal_output_signature = {
+        'predictions': 'AttentionOcr_v1/predicted_chars:0',
+        'scores': 'AttentionOcr_v1/predicted_scores:0',
+        'predicted_length': 'AttentionOcr_v1/predicted_length:0',
+        'predicted_text': 'AttentionOcr_v1/predicted_text:0',
+        'predicted_conf': 'AttentionOcr_v1/predicted_conf:0',
+        'normalized_seq_conf': 'AttentionOcr_v1/normalized_seq_conf:0'
+    }
+
+  def create_input_feed(self, graph_def, serving):
+    """Returns the input feed for the model.
+
+    Creates random images, according to the size specified by dataset_name,
+    format it in the correct way depending on whether the model was exported
+    for serving, and return the correctly keyed feed_dict for inference.
+
+    Args:
+      graph_def: Graph definition of the loaded model.
+      serving: Whether the model was exported for Serving.
+
+    Returns:
+      The feed_dict suitable for model inference.
+    """
+    # Creates a dataset based on FLAGS.dataset_name.
+    self.dataset = common_flags.create_dataset('test')
+    # Create some random images to test inference for any dataset.
+    self.images = {
+        'img1':
+            np.random.uniform(low=64, high=192,
+                              size=self.dataset.image_shape).astype('uint8'),
+        'img2':
+            np.random.uniform(low=32, high=224,
+                              size=self.dataset.image_shape).astype('uint8'),
+    }
+    signature_def = graph_def.signature_def[
+        tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY]
+    if serving:
+      input_name = signature_def.inputs[
+          tf.saved_model.signature_constants.CLASSIFY_INPUTS].name
+      # Model for serving takes input: inputs['inputs'] = 'tf_example:0'
+      feed_dict = {
+          input_name: [
+              _create_tf_example_string(self.images['img1']),
+              _create_tf_example_string(self.images['img2'])
+          ]
+      }
+    else:
+      input_name = signature_def.inputs['images'].name
+      # Model for direct use takes input: inputs['images'] = 'original_image:0'
+      feed_dict = {
+          input_name: np.stack([self.images['img1'], self.images['img2']])
+      }
+    return feed_dict
+
+  def verify_export_load_and_inference(self, export_for_serving=False):
+    """Verify exported model can be loaded and inference can run successfully.
+
+    This function will load the exported model in self.export_dir, then create
+    some fake images according to the specification of FLAGS.dataset_name.
+    It then feeds the input through the model, and verify the minimal set of
+    output signatures are present.
+    Note: Model and dataset creation in the underlying library depends on the
+          following commandline flags:
+            FLAGS.dataset_name
+    Args:
+      export_for_serving: True if the model was exported for Serving. This
+        affects how input is fed into the model.
+    """
+    tf.reset_default_graph()
+    sess = tf.Session()
+    graph_def = tf.saved_model.loader.load(
+        sess=sess,
+        tags=[tf.saved_model.tag_constants.SERVING],
+        export_dir=self.export_dir)
+    feed_dict = self.create_input_feed(graph_def, export_for_serving)
+    results = sess.run(self.minimal_output_signature, feed_dict=feed_dict)
+
+    out_shape = (2,)
+    self.assertEqual(np.shape(results['predicted_conf']), out_shape)
+    self.assertEqual(np.shape(results['predicted_text']), out_shape)
+    self.assertEqual(np.shape(results['predicted_length']), out_shape)
+    self.assertEqual(np.shape(results['normalized_seq_conf']), out_shape)
+    out_shape = (2, self.dataset.max_sequence_length)
+    self.assertEqual(np.shape(results['scores']), out_shape)
+    self.assertEqual(np.shape(results['predictions']), out_shape)
+
+  @flagsaver.flagsaver
+  def test_fsns_export_for_serving_and_load_inference(self):
+    model_export.export_model(self.export_dir, True)
+    self.verify_export_load_and_inference(True)
+
+  @flagsaver.flagsaver
+  def test_fsns_export_and_load_inference(self):
+    model_export.export_model(self.export_dir, False, batch_size=2)
+    self.verify_export_load_and_inference(False)
+
+
+if __name__ == '__main__':
+  tf.test.main()

research/attention_ocr/python/model_test.py

@@ -12,11 +12,10 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 # ==============================================================================
-
 """Tests for the model."""
+import string
 
 import numpy as np
-import string
 import tensorflow as tf
 from tensorflow.contrib import slim
 
@@ -32,6 +31,7 @@ def create_fake_charset(num_char_classes):
 
 
 class ModelTest(tf.test.TestCase):
+
   def setUp(self):
     tf.test.TestCase.setUp(self)
 
@@ -51,18 +51,21 @@ class ModelTest(tf.test.TestCase):
     self.chars_logit_shape = (self.batch_size, self.seq_length,
                               self.num_char_classes)
     self.length_logit_shape = (self.batch_size, self.seq_length + 1)
+    # Placeholder knows image dimensions, but not batch size.
+    self.input_images = tf.placeholder(
+        tf.float32,
+        shape=(None, self.image_height, self.image_width, 3),
+        name='input_node')
 
     self.initialize_fakes()
 
   def initialize_fakes(self):
     self.images_shape = (self.batch_size, self.image_height, self.image_width,
                          3)
-    self.fake_images = tf.constant(
-        self.rng.randint(low=0, high=255,
-                         size=self.images_shape).astype('float32'),
-        name='input_node')
-    self.fake_conv_tower_np = self.rng.randn(
-        *self.conv_tower_shape).astype('float32')
+    self.fake_images = self.rng.randint(
+        low=0, high=255, size=self.images_shape).astype('float32')
+    self.fake_conv_tower_np = self.rng.randn(*self.conv_tower_shape).astype(
+        'float32')
     self.fake_conv_tower = tf.constant(self.fake_conv_tower_np)
     self.fake_logits = tf.constant(
         self.rng.randn(*self.chars_logit_shape).astype('float32'))
@@ -74,33 +77,44 @@ class ModelTest(tf.test.TestCase):
 
   def create_model(self, charset=None):
     return model.Model(
-        self.num_char_classes, self.seq_length, num_views=4, null_code=62,
+        self.num_char_classes,
+        self.seq_length,
+        num_views=4,
+        null_code=62,
         charset=charset)
 
   def test_char_related_shapes(self):
-    ocr_model = self.create_model()
+    charset = create_fake_charset(self.num_char_classes)
+    ocr_model = self.create_model(charset=charset)
     with self.test_session() as sess:
       endpoints_tf = ocr_model.create_base(
-          images=self.fake_images, labels_one_hot=None)
-
+          images=self.input_images, labels_one_hot=None)
       sess.run(tf.global_variables_initializer())
-      endpoints = sess.run(endpoints_tf)
-
-      self.assertEqual((self.batch_size, self.seq_length,
-                        self.num_char_classes), endpoints.chars_logit.shape)
-      self.assertEqual((self.batch_size, self.seq_length,
-                        self.num_char_classes), endpoints.chars_log_prob.shape)
+      tf.tables_initializer().run()
+      endpoints = sess.run(
+          endpoints_tf, feed_dict={self.input_images: self.fake_images})
+
+      self.assertEqual(
+          (self.batch_size, self.seq_length, self.num_char_classes),
+          endpoints.chars_logit.shape)
+      self.assertEqual(
+          (self.batch_size, self.seq_length, self.num_char_classes),
+          endpoints.chars_log_prob.shape)
       self.assertEqual((self.batch_size, self.seq_length),
                        endpoints.predicted_chars.shape)
       self.assertEqual((self.batch_size, self.seq_length),
                        endpoints.predicted_scores.shape)
+      self.assertEqual((self.batch_size,), endpoints.predicted_text.shape)
+      self.assertEqual((self.batch_size,), endpoints.predicted_conf.shape)
+      self.assertEqual((self.batch_size,), endpoints.normalized_seq_conf.shape)
 
   def test_predicted_scores_are_within_range(self):
     ocr_model = self.create_model()
 
     _, _, scores = ocr_model.char_predictions(self.fake_logits)
     with self.test_session() as sess:
-      scores_np = sess.run(scores)
+      scores_np = sess.run(
+          scores, feed_dict={self.input_images: self.fake_images})
 
     values_in_range = (scores_np >= 0.0) & (scores_np <= 1.0)
     self.assertTrue(
@@ -111,10 +125,11 @@ class ModelTest(tf.test.TestCase):
   def test_conv_tower_shape(self):
     with self.test_session() as sess:
       ocr_model = self.create_model()
-      conv_tower = ocr_model.conv_tower_fn(self.fake_images)
+      conv_tower = ocr_model.conv_tower_fn(self.input_images)
 
       sess.run(tf.global_variables_initializer())
-      conv_tower_np = sess.run(conv_tower)
+      conv_tower_np = sess.run(
+          conv_tower, feed_dict={self.input_images: self.fake_images})
 
       self.assertEqual(self.conv_tower_shape, conv_tower_np.shape)
 
@@ -124,11 +139,12 @@ class ModelTest(tf.test.TestCase):
     # updates, gradients and variances. It also depends on the type of used
     # optimizer.
     ocr_model = self.create_model()
-    ocr_model.create_base(images=self.fake_images, labels_one_hot=None)
+    ocr_model.create_base(images=self.input_images, labels_one_hot=None)
     with self.test_session() as sess:
       tfprof_root = tf.profiler.profile(
           sess.graph,
-          options=tf.profiler.ProfileOptionBuilder.trainable_variables_parameter())
+          options=tf.profiler.ProfileOptionBuilder
+          .trainable_variables_parameter())
 
       model_size_bytes = 4 * tfprof_root.total_parameters
       self.assertLess(model_size_bytes, 1 * 2**30)
@@ -158,7 +174,7 @@ class ModelTest(tf.test.TestCase):
 
     loss = model.sequence_loss_fn(self.fake_logits, self.fake_labels)
     with self.test_session() as sess:
-      loss_np = sess.run(loss)
+      loss_np = sess.run(loss, feed_dict={self.input_images: self.fake_images})
 
     # This test checks that the loss function is 'runnable'.
     self.assertEqual(loss_np.shape, tuple())
@@ -172,7 +188,8 @@ class ModelTest(tf.test.TestCase):
     Returns:
       a list of tensors with encoded image coordinates in them.
     """
-    batch_size, h, w, _ = net.shape.as_list()
+    batch_size = tf.shape(net)[0]
+    _, h, w, _ = net.shape.as_list()
     h_loc = [
         tf.tile(
             tf.reshape(
@@ -197,11 +214,12 @@ class ModelTest(tf.test.TestCase):
     conv_w_coords_tf = model.encode_coordinates_fn(self.fake_conv_tower)
 
     with self.test_session() as sess:
-      conv_w_coords = sess.run(conv_w_coords_tf)
+      conv_w_coords = sess.run(
+          conv_w_coords_tf, feed_dict={self.input_images: self.fake_images})
 
     batch_size, height, width, feature_size = self.conv_tower_shape
-    self.assertEqual(conv_w_coords.shape, (batch_size, height, width,
-                                           feature_size + height + width))
+    self.assertEqual(conv_w_coords.shape,
+                     (batch_size, height, width, feature_size + height + width))
 
   def test_disabled_coordinate_encoding_returns_features_unchanged(self):
     model = self.create_model()
@@ -209,7 +227,8 @@ class ModelTest(tf.test.TestCase):
     conv_w_coords_tf = model.encode_coordinates_fn(self.fake_conv_tower)
 
     with self.test_session() as sess:
-      conv_w_coords = sess.run(conv_w_coords_tf)
+      conv_w_coords = sess.run(
+          conv_w_coords_tf, feed_dict={self.input_images: self.fake_images})
 
     self.assertAllEqual(conv_w_coords, self.fake_conv_tower_np)
 
@@ -221,7 +240,8 @@ class ModelTest(tf.test.TestCase):
     conv_w_coords_tf = model.encode_coordinates_fn(fake_conv_tower)
 
     with self.test_session() as sess:
-      conv_w_coords = sess.run(conv_w_coords_tf)
+      conv_w_coords = sess.run(
+          conv_w_coords_tf, feed_dict={self.input_images: self.fake_images})
 
     # Original features
     self.assertAllEqual(conv_w_coords[0, :, :, :4],
@@ -261,10 +281,11 @@ class ModelTest(tf.test.TestCase):
 
 
 class CharsetMapperTest(tf.test.TestCase):
+
   def test_text_corresponds_to_ids(self):
     charset = create_fake_charset(36)
-    ids = tf.constant(
-        [[17, 14, 21, 21, 24], [32, 24, 27, 21, 13]], dtype=tf.int64)
+    ids = tf.constant([[17, 14, 21, 21, 24], [32, 24, 27, 21, 13]],
+                      dtype=tf.int64)
     charset_mapper = model.CharsetMapper(charset)
 
     with self.test_session() as sess:

research/attention_ocr/python/sequence_layers.py

@@ -111,7 +111,7 @@ class SequenceLayerBase(object):
     self._mparams = method_params
     self._net = net
     self._labels_one_hot = labels_one_hot
-    self._batch_size = net.get_shape().dims[0].value
+    self._batch_size = tf.shape(net)[0]
 
     # Initialize parameters for char logits which will be computed on the fly
     # inside an LSTM decoder.
@@ -275,7 +275,7 @@ class NetSlice(SequenceLayerBase):
   def __init__(self, *args, **kwargs):
     super(NetSlice, self).__init__(*args, **kwargs)
     self._zero_label = tf.zeros(
-        [self._batch_size, self._params.num_char_classes])
+        tf.stack([self._batch_size, self._params.num_char_classes]))
 
   def get_image_feature(self, char_index):
     """Returns a subset of image features for a character.
@@ -352,7 +352,7 @@ class Attention(SequenceLayerBase):
   def __init__(self, *args, **kwargs):
     super(Attention, self).__init__(*args, **kwargs)
     self._zero_label = tf.zeros(
-        [self._batch_size, self._params.num_char_classes])
+        tf.stack([self._batch_size, self._params.num_char_classes]))
 
   def get_eval_input(self, prev, i):
     """See SequenceLayerBase.get_eval_input for details."""

research/attention_ocr/python/utils.py

@@ -78,3 +78,20 @@ def variables_to_restore(scope=None, strip_scope=False):
     return variable_map
   else:
     return {v.op.name: v for v in slim.get_variables_to_restore()}
+
+
+def ConvertAllInputsToTensors(func):
+  """A decorator to convert all function's inputs into tensors.
+
+  Args:
+    func: a function to decorate.
+
+  Returns:
+    A decorated function.
+  """
+
+  def FuncWrapper(*args):
+    tensors = [tf.convert_to_tensor(a) for a in args]
+    return func(*tensors)
+
+  return FuncWrapper