Merge remote-tracking branch 'upstream/master'

adversarial_text/train_classifier.py

+# Copyright 2017 Google Inc. 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.
+# ==============================================================================
+"""Trains LSTM text classification model.
+
+Model trains with adversarial or virtual adversarial training.
+
+Computational time:
+  1.8 hours to train 10000 steps without adversarial or virtual adversarial
+    training, on 1 layer 1024 hidden units LSTM, 256 embeddings, 400 truncated
+    BP, 64 minibatch and on single GPU (Pascal Titan X, cuDNNv5).
+
+  4 hours to train 10000 steps with adversarial or virtual adversarial
+    training, with above condition.
+
+To initialize embedding and LSTM cell weights from a pretrained model, set
+FLAGS.pretrained_model_dir to the pretrained model's checkpoint directory.
+"""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+# Dependency imports
+
+import tensorflow as tf
+
+import graphs
+import train_utils
+
+flags = tf.app.flags
+FLAGS = flags.FLAGS
+
+flags.DEFINE_string('pretrained_model_dir', None,
+                    'Directory path to pretrained model to restore from')
+
+
+def main(_):
+  """Trains LSTM classification model."""
+  tf.logging.set_verbosity(tf.logging.INFO)
+  with tf.device(tf.train.replica_device_setter(FLAGS.ps_tasks)):
+    model = graphs.get_model()
+    train_op, loss, global_step = model.classifier_training()
+    train_utils.run_training(
+        train_op,
+        loss,
+        global_step,
+        variables_to_restore=model.pretrained_variables,
+        pretrained_model_dir=FLAGS.pretrained_model_dir)
+
+
+if __name__ == '__main__':
+  tf.app.run()

adversarial_text/train_utils.py

+# Copyright 2017 Google Inc. 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.
+# ==============================================================================
+"""Utilities for training adversarial text models."""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import time
+
+# Dependency imports
+
+import numpy as np
+import tensorflow as tf
+
+flags = tf.app.flags
+FLAGS = flags.FLAGS
+
+flags.DEFINE_string('master', '', 'Master address.')
+flags.DEFINE_integer('task', 0, 'Task id of the replica running the training.')
+flags.DEFINE_integer('ps_tasks', 0, 'Number of parameter servers.')
+flags.DEFINE_string('train_dir', '/tmp/text_train',
+                    'Directory for logs and checkpoints.')
+flags.DEFINE_integer('max_steps', 1000000, 'Number of batches to run.')
+flags.DEFINE_boolean('log_device_placement', False,
+                     'Whether to log device placement.')
+
+
+def run_training(train_op,
+                 loss,
+                 global_step,
+                 variables_to_restore=None,
+                 pretrained_model_dir=None):
+  """Sets up and runs training loop."""
+  tf.gfile.MakeDirs(FLAGS.train_dir)
+
+  # Create pretrain Saver
+  if pretrained_model_dir:
+    assert variables_to_restore
+    tf.logging.info('Will attempt restore from %s: %s', pretrained_model_dir,
+                    variables_to_restore)
+    saver_for_restore = tf.train.Saver(variables_to_restore)
+
+  # Init ops
+  if FLAGS.sync_replicas:
+    local_init_op = tf.get_collection('local_init_op')[0]
+    ready_for_local_init_op = tf.get_collection('ready_for_local_init_op')[0]
+  else:
+    local_init_op = tf.train.Supervisor.USE_DEFAULT
+    ready_for_local_init_op = tf.train.Supervisor.USE_DEFAULT
+
+  is_chief = FLAGS.task == 0
+  sv = tf.train.Supervisor(
+      logdir=FLAGS.train_dir,
+      is_chief=is_chief,
+      save_summaries_secs=5 * 60,
+      save_model_secs=5 * 60,
+      local_init_op=local_init_op,
+      ready_for_local_init_op=ready_for_local_init_op,
+      global_step=global_step)
+
+  # Delay starting standard services to allow possible pretrained model restore.
+  with sv.managed_session(
+      master=FLAGS.master,
+      config=tf.ConfigProto(log_device_placement=FLAGS.log_device_placement),
+      start_standard_services=False) as sess:
+    # Initialization
+    if is_chief:
+      if pretrained_model_dir:
+        maybe_restore_pretrained_model(sess, saver_for_restore,
+                                       pretrained_model_dir)
+      if FLAGS.sync_replicas:
+        sess.run(tf.get_collection('chief_init_op')[0])
+      sv.start_standard_services(sess)
+
+    sv.start_queue_runners(sess)
+
+    # Training loop
+    global_step_val = 0
+    while not sv.should_stop() and global_step_val < FLAGS.max_steps:
+      global_step_val = train_step(sess, train_op, loss, global_step)
+    sv.stop()
+
+    # Final checkpoint
+    if is_chief:
+      sv.saver.save(sess, sv.save_path, global_step=global_step)
+
+
+def maybe_restore_pretrained_model(sess, saver_for_restore, model_dir):
+  """Restores pretrained model if there is no ckpt model."""
+  ckpt = tf.train.get_checkpoint_state(FLAGS.train_dir)
+  checkpoint_exists = ckpt and ckpt.model_checkpoint_path
+  if checkpoint_exists:
+    tf.logging.info('Checkpoint exists in FLAGS.train_dir; skipping '
+                    'pretraining restore')
+    return
+
+  pretrain_ckpt = tf.train.get_checkpoint_state(model_dir)
+  if not (pretrain_ckpt and pretrain_ckpt.model_checkpoint_path):
+    raise ValueError(
+        'Asked to restore model from %s but no checkpoint found.' % model_dir)
+  saver_for_restore.restore(sess, pretrain_ckpt.model_checkpoint_path)
+
+
+def train_step(sess, train_op, loss, global_step):
+  """Runs a single training step."""
+  start_time = time.time()
+  _, loss_val, global_step_val = sess.run([train_op, loss, global_step])
+  duration = time.time() - start_time
+
+  # Logging
+  if global_step_val % 10 == 0:
+    examples_per_sec = FLAGS.batch_size / duration
+    sec_per_batch = float(duration)
+
+    format_str = ('step %d, loss = %.2f (%.1f examples/sec; %.3f ' 'sec/batch)')
+    tf.logging.info(format_str % (global_step_val, loss_val, examples_per_sec,
+                                  sec_per_batch))
+
+  if np.isnan(loss_val):
+    raise OverflowError('Loss is nan')
+
+  return global_step_val

attention_ocr/README.md

+## Attention-based Extraction of Structured Information from Street View Imagery
+
+*A TensorFlow model for real-world image text extraction problems.*
+
+This folder contains the code needed to train a new Attention OCR model on the
+[FSNS dataset][FSNS] dataset to transcribe street names in France. You can
+also use it to train it on your own data.
+
+More details can be found in our paper:
+
+["Attention-based Extraction of Structured Information from Street View
+Imagery"](https://arxiv.org/abs/1704.03549)
+
+## Contacts
+
+Authors:
+Zbigniew Wojna <zbigniewwojna@gmail.com>,
+Alexander Gorban <gorban@google.com>
+
+Pull requests:
+[alexgorban](https://github.com/alexgorban)
+
+## Requirements
+
+1. Install the TensorFlow library ([instructions][TF]). For example:
+
+```
+virtualenv --system-site-packages ~/.tensorflow
+source ~/.tensorflow/bin/activate
+pip install --upgrade pip
+pip install --upgrade tensorflow_gpu
+```
+
+2. At least 158GB of free disk space to download the FSNS dataset:
+
+```
+cd models/attention_ocr/python/datasets
+aria2c -c -j 20 -i ../../../street/python/fsns_urls.txt
+cd ..
+```
+
+3. 16GB of RAM or more; 32GB is recommended.
+4. `train.py` works with both CPU and GPU, though using GPU is preferable. It has been tested with a Titan X and with a GTX980.
+
+[TF]: https://www.tensorflow.org/install/
+[FSNS]: https://github.com/tensorflow/models/tree/master/street
+
+## How to use this code
+
+To run all unit tests:
+
+```
+cd models/attention_ocr/python
+python -m unittest discover -p  '*_test.py'
+```
+
+To train from scratch:
+
+```
+python train.py
+```
+
+To train a model using pre-trained Inception weights as initialization:
+
+```
+wget http://download.tensorflow.org/models/inception_v3_2016_08_28.tar.gz
+tar xf inception_v3_2016_08_28.tar.gz
+python train.py --checkpoint_inception=inception_v3.ckpt
+```
+
+To fine tune the Attention OCR model using a checkpoint:
+
+```
+wget http://download.tensorflow.org/models/attention_ocr_2017_05_17.tar.gz
+tar xf attention_ocr_2017_05_17.tar.gz
+python train.py --checkpoint=model.ckpt-399731
+```
+
+## How to use your own image data to train the model
+
+You need to define a new dataset. There are two options:
+
+1. Store data in the same format as the FSNS dataset and just reuse the
+[python/datasets/fsns.py](https://github.com/tensorflow/models/blob/master/attention_ocr/python/datasets/fsns.py)
+module. E.g., create a file datasets/newtextdataset.py:
+```
+import fsns
+
+DEFAULT_DATASET_DIR = 'path/to/the/dataset'
+
+DEFAULT_CONFIG = {
+    'name':
+        'MYDATASET',
+    'splits': {
+        'train': {
+            'size': 123,
+            'pattern': 'tfexample_train*'
+        },
+        'test': {
+            'size': 123,
+            'pattern': 'tfexample_test*'
+        }
+    },
+    'charset_filename':
+        'charset_size.txt',
+    'image_shape': (150, 600, 3),
+    'num_of_views':
+        4,
+    'max_sequence_length':
+        37,
+    'null_code':
+        42,
+    'items_to_descriptions': {
+        'image':
+            'A [150 x 600 x 3] color image.',
+        'label':
+            'Characters codes.',
+        'text':
+            'A unicode string.',
+        'length':
+            'A length of the encoded text.',
+        'num_of_views':
+            'A number of different views stored within the image.'
+    }
+}
+
+
+def get_split(split_name, dataset_dir=None, config=None):
+  if not dataset_dir:
+    dataset_dir = DEFAULT_DATASET_DIR
+  if not config:
+    config = DEFAULT_CONFIG
+
+  return fsns.get_split(split_name, dataset_dir, config)
+```
+You will also need to include it into the `datasets/__init__.py` and specify the
+dataset name in the command line.
+
+```
+python train.py --dataset_name=newtextdataset
+```
+
+Please note that eval.py will also require the same flag.
+
+2. Define a new dataset format. The model needs the following data to train:
+
+- images: input images,  shape [batch_size x H x W x 3];
+- labels: ground truth label ids,  shape=[batch_size x seq_length];
+- labels_one_hot: labels in one-hot encoding,  shape [batch_size x seq_length x num_char_classes];
+
+Refer to [python/data_provider.py](https://github.com/tensorflow/models/blob/master/attention_ocr/python/data_provider.py#L33)
+for more details. You can use [python/datasets/fsns.py](https://github.com/tensorflow/models/blob/master/attention_ocr/python/datasets/fsns.py)
+as the example.
+
+## How to use a pre-trained model
+
+The inference part was not released yet, but it is pretty straightforward to
+implement one in Python or C++.
+
+The recommended way is to use the [Serving infrastructure](https://tensorflow.github.io/serving/serving_basic).
+
+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/attention_ocr/python/eval.py#L60)
+`endpoints = model.create_base(images_placeholder, labels_one_hot=None)`
+3. [load a pretrained model](https://github.com/tensorflow/models/blob/master/attention_ocr/python/model.py#L494)
+4. run computations through the graph:
+`predictions = sess.run(endpoints.predicted_chars, feed_dict={images_placeholder:images_actual_data})`
+5. Convert character IDs (predictions) to UTF8 using the provided charset file.
+
+## Disclaimer
+
+This code is a modified version of the internal model we used for our paper.
+Currently it reaches 83.79% full sequence accuracy after 400k steps of training.
+The main difference between this version and the version used in the paper - for
+the paper we used a distributed training with 50 GPU (K80) workers (asynchronous
+updates), the provided checkpoint was created using this code after ~6 days of
+training on a single GPU (Titan X) (it reached 81% after 24 hours of training),
+the coordinate encoding is missing TODO(alexgorban@).

attention_ocr/python/all_jobs.screenrc

+# A GPU/screen config to run all jobs for training and evaluation in parallel.
+# Execute:
+# source /path/to/your/virtualenv/bin/activate
+# screen -R TF -c all_jobs.screenrc
+
+screen -t train 0 python train.py --train_log_dir=workdir/train
+screen -t eval_train 1 python eval.py --split_name=train --train_log_dir=workdir/train --eval_log_dir=workdir/eval_train
+screen -t eval_test 2 python eval.py --split_name=test --train_log_dir=workdir/train --eval_log_dir=workdir/eval_test
+screen -t tensorboard 3 tensorboard --logdir=workdir

attention_ocr/python/common_flags.py

+# Copyright 2017 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.
+# ==============================================================================
+
+"""Define flags are common for both train.py and eval.py scripts."""
+import sys
+
+from tensorflow.python.platform import flags
+import logging
+
+import datasets
+import model
+
+FLAGS = flags.FLAGS
+
+logging.basicConfig(
+    level=logging.DEBUG,
+    stream=sys.stderr,
+    format='%(levelname)s '
+    '%(asctime)s.%(msecs)06d: '
+    '%(filename)s: '
+    '%(lineno)d '
+    '%(message)s',
+    datefmt='%Y-%m-%d %H:%M:%S')
+
+
+def define():
+  """Define common flags."""
+  # yapf: disable
+  flags.DEFINE_integer('batch_size', 32,
+                       'Batch size.')
+
+  flags.DEFINE_integer('crop_width', None,
+                       'Width of the central crop for images.')
+
+  flags.DEFINE_integer('crop_height', None,
+                       'Height of the central crop for images.')
+
+  flags.DEFINE_string('train_log_dir', '/tmp/attention_ocr/train',
+                      'Directory where to write event logs.')
+
+  flags.DEFINE_string('dataset_name', 'fsns',
+                      'Name of the dataset. Supported: fsns')
+
+  flags.DEFINE_string('split_name', 'train',
+                      'Dataset split name to run evaluation for: test,train.')
+
+  flags.DEFINE_string('dataset_dir', None,
+                      'Dataset root folder.')
+
+  flags.DEFINE_string('checkpoint', '',
+                      'Path for checkpoint to restore weights from.')
+
+  flags.DEFINE_string('master',
+                      '',
+                      'BNS name of the TensorFlow master to use.')
+
+  # Model hyper parameters
+  flags.DEFINE_float('learning_rate', 0.004,
+                     'learning rate')
+
+  flags.DEFINE_string('optimizer', 'momentum',
+                      'the optimizer to use')
+
+  flags.DEFINE_string('momentum', 0.9,
+                      'momentum value for the momentum optimizer if used')
+
+  flags.DEFINE_bool('use_augment_input', True,
+                    'If True will use image augmentation')
+
+  # Method hyper parameters
+  # conv_tower_fn
+  flags.DEFINE_string('final_endpoint', 'Mixed_5d',
+                      'Endpoint to cut inception tower')
+
+  # sequence_logit_fn
+  flags.DEFINE_bool('use_attention', True,
+                    'If True will use the attention mechanism')
+
+  flags.DEFINE_bool('use_autoregression', True,
+                    'If True will use autoregression (a feedback link)')
+
+  flags.DEFINE_integer('num_lstm_units', 256,
+                       'number of LSTM units for sequence LSTM')
+
+  flags.DEFINE_float('weight_decay', 0.00004,
+                     'weight decay for char prediction FC layers')
+
+  flags.DEFINE_float('lstm_state_clip_value', 10.0,
+                     'cell state is clipped by this value prior to the cell'
+                     ' output activation')
+
+  # 'sequence_loss_fn'
+  flags.DEFINE_float('label_smoothing', 0.1,
+                     'weight for label smoothing')
+
+  flags.DEFINE_bool('ignore_nulls', True,
+                    'ignore null characters for computing the loss')
+
+  flags.DEFINE_bool('average_across_timesteps', False,
+                    'divide the returned cost by the total label weight')
+  # yapf: enable
+
+
+def get_crop_size():
+  if FLAGS.crop_width and FLAGS.crop_height:
+    return (FLAGS.crop_width, FLAGS.crop_height)
+  else:
+    return None
+
+
+def create_dataset(split_name):
+  ds_module = getattr(datasets, FLAGS.dataset_name)
+  return ds_module.get_split(split_name, dataset_dir=FLAGS.dataset_dir)
+
+
+def create_mparams():
+  return {
+      'conv_tower_fn':
+      model.ConvTowerParams(final_endpoint=FLAGS.final_endpoint),
+      'sequence_logit_fn':
+      model.SequenceLogitsParams(
+          use_attention=FLAGS.use_attention,
+          use_autoregression=FLAGS.use_autoregression,
+          num_lstm_units=FLAGS.num_lstm_units,
+          weight_decay=FLAGS.weight_decay,
+          lstm_state_clip_value=FLAGS.lstm_state_clip_value),
+      'sequence_loss_fn':
+      model.SequenceLossParams(
+          label_smoothing=FLAGS.label_smoothing,
+          ignore_nulls=FLAGS.ignore_nulls,
+          average_across_timesteps=FLAGS.average_across_timesteps)
+  }
+
+
+def create_model(*args, **kwargs):
+  ocr_model = model.Model(mparams=create_mparams(), *args, **kwargs)
+  return ocr_model

attention_ocr/python/data_provider.py

+# Copyright 2017 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.
+# ==============================================================================
+
+"""Functions to read, decode and pre-process input data for the Model.
+"""
+import collections
+import functools
+import tensorflow as tf
+from tensorflow.contrib import slim
+
+import inception_preprocessing
+
+# Tuple to store input data endpoints for the Model.
+# It has following fields (tensors):
+#    images: input images,
+#      shape [batch_size x H x W x 3];
+#    labels: ground truth label ids,
+#      shape=[batch_size x seq_length];
+#    labels_one_hot: labels in one-hot encoding,
+#      shape [batch_size x seq_length x num_char_classes];
+InputEndpoints = collections.namedtuple(
+    'InputEndpoints', ['images', 'images_orig', 'labels', 'labels_one_hot'])
+
+# A namedtuple to define a configuration for shuffled batch fetching.
+#   num_batching_threads: A number of parallel threads to fetch data.
+#   queue_capacity: a max number of elements in the batch shuffling queue.
+#   min_after_dequeue: a min number elements in the queue after a dequeue, used
+#     to ensure a level of mixing of elements.
+ShuffleBatchConfig = collections.namedtuple('ShuffleBatchConfig', [
+    'num_batching_threads', 'queue_capacity', 'min_after_dequeue'
+])
+
+DEFAULT_SHUFFLE_CONFIG = ShuffleBatchConfig(
+    num_batching_threads=8, queue_capacity=3000, min_after_dequeue=1000)
+
+
+def augment_image(image):
+  """Augmentation the image with a random modification.
+
+  Args:
+    image: input Tensor image of rank 3, with the last dimension
+           of size 3.
+
+  Returns:
+    Distorted Tensor image of the same shape.
+  """
+  with tf.variable_scope('AugmentImage'):
+    height = image.get_shape().dims[0].value
+    width = image.get_shape().dims[1].value
+
+    # Random crop cut from the street sign image, resized to the same size.
+    # Assures that the crop is covers at least 0.8 area of the input image.
+    bbox_begin, bbox_size, _ = tf.image.sample_distorted_bounding_box(
+        tf.shape(image),
+        bounding_boxes=tf.zeros([0, 0, 4]),
+        min_object_covered=0.8,
+        aspect_ratio_range=[0.8, 1.2],
+        area_range=[0.8, 1.0],
+        use_image_if_no_bounding_boxes=True)
+    distorted_image = tf.slice(image, bbox_begin, bbox_size)
+
+    # Randomly chooses one of the 4 interpolation methods
+    distorted_image = inception_preprocessing.apply_with_random_selector(
+        distorted_image,
+        lambda x, method: tf.image.resize_images(x, [height, width], method),
+        num_cases=4)
+    distorted_image.set_shape([height, width, 3])
+
+    # Color distortion
+    distorted_image = inception_preprocessing.apply_with_random_selector(
+        distorted_image,
+        functools.partial(
+            inception_preprocessing.distort_color, fast_mode=False),
+        num_cases=4)
+    distorted_image = tf.clip_by_value(distorted_image, -1.5, 1.5)
+
+  return distorted_image
+
+
+def central_crop(image, crop_size):
+  """Returns a central crop for the specified size of an image.
+
+  Args:
+    image: A tensor with shape [height, width, channels]
+    crop_size: A tuple (crop_width, crop_height)
+
+  Returns:
+    A tensor of shape [crop_height, crop_width, channels].
+  """
+  with tf.variable_scope('CentralCrop'):
+    target_width, target_height = crop_size
+    image_height, image_width = tf.shape(image)[0], tf.shape(image)[1]
+    assert_op1 = tf.Assert(
+        tf.greater_equal(image_height, target_height),
+        ['image_height < target_height', image_height, target_height])
+    assert_op2 = tf.Assert(
+        tf.greater_equal(image_width, target_width),
+        ['image_width < target_width', image_width, target_width])
+    with tf.control_dependencies([assert_op1, assert_op2]):
+      offset_width = (image_width - target_width) / 2
+      offset_height = (image_height - target_height) / 2
+      return tf.image.crop_to_bounding_box(image, offset_height, offset_width,
+                                           target_height, target_width)
+
+
+def preprocess_image(image, augment=False, central_crop_size=None,
+                     num_towers=4):
+  """Normalizes image to have values in a narrow range around zero.
+
+  Args:
+    image: a [H x W x 3] uint8 tensor.
+    augment: optional, if True do random image distortion.
+    central_crop_size: A tuple (crop_width, crop_height).
+    num_towers: optional, number of shots of the same image in the input image.
+
+  Returns:
+    A float32 tensor of shape [H x W x 3] with RGB values in the required
+    range.
+  """
+  with tf.variable_scope('PreprocessImage'):
+    image = tf.image.convert_image_dtype(image, dtype=tf.float32)
+    if augment or central_crop_size:
+      if num_towers == 1:
+        images = [image]
+      else:
+        images = tf.split(value=image, num_or_size_splits=num_towers, axis=1)
+      if central_crop_size:
+        view_crop_size = (central_crop_size[0] / num_towers,
+                          central_crop_size[1])
+        images = [central_crop(img, view_crop_size) for img in images]
+      if augment:
+        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
+
+
+def get_data(dataset,
+             batch_size,
+             augment=False,
+             central_crop_size=None,
+             shuffle_config=None,
+             shuffle=True):
+  """Wraps calls to DatasetDataProviders and shuffle_batch.
+
+  For more details about supported Dataset objects refer to datasets/fsns.py.
+
+  Args:
+    dataset: a slim.data.dataset.Dataset object.
+    batch_size: number of samples per batch.
+    augment: optional, if True does random image distortion.
+    central_crop_size: A CharLogittuple (crop_width, crop_height).
+    shuffle_config: A namedtuple ShuffleBatchConfig.
+    shuffle: if True use data shuffling.
+
+  Returns:
+
+  """
+  if not shuffle_config:
+    shuffle_config = DEFAULT_SHUFFLE_CONFIG
+
+  provider = slim.dataset_data_provider.DatasetDataProvider(
+      dataset,
+      shuffle=shuffle,
+      common_queue_capacity=2 * batch_size,
+      common_queue_min=batch_size)
+  image_orig, label = provider.get(['image', 'label'])
+
+  image = preprocess_image(
+      image_orig, augment, central_crop_size, num_towers=dataset.num_of_views)
+  label_one_hot = slim.one_hot_encoding(label, dataset.num_char_classes)
+
+  images, images_orig, labels, labels_one_hot = (tf.train.shuffle_batch(
+      [image, image_orig, label, label_one_hot],
+      batch_size=batch_size,
+      num_threads=shuffle_config.num_batching_threads,
+      capacity=shuffle_config.queue_capacity,
+      min_after_dequeue=shuffle_config.min_after_dequeue))
+
+  return InputEndpoints(
+      images=images,
+      images_orig=images_orig,
+      labels=labels,
+      labels_one_hot=labels_one_hot)

attention_ocr/python/data_provider_test.py

+# Copyright 2017 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 data_provider."""
+
+import numpy as np
+import tensorflow as tf
+from tensorflow.contrib.slim import queues
+
+import datasets
+import data_provider
+
+
+class DataProviderTest(tf.test.TestCase):
+  def setUp(self):
+    tf.test.TestCase.setUp(self)
+
+  def test_preprocessed_image_values_are_in_range(self):
+    image_shape = (5, 4, 3)
+    fake_image = np.random.randint(low=0, high=255, size=image_shape)
+    image_tf = data_provider.preprocess_image(fake_image)
+
+    with self.test_session() as sess:
+      image_np = sess.run(image_tf)
+
+    self.assertEqual(image_np.shape, image_shape)
+    min_value, max_value = np.min(image_np), np.max(image_np)
+    self.assertTrue((-1.28 < min_value) and (min_value < 1.27))
+    self.assertTrue((-1.28 < max_value) and (max_value < 1.27))
+
+  def test_provided_data_has_correct_shape(self):
+    batch_size = 4
+    data = data_provider.get_data(
+        dataset=datasets.fsns_test.get_test_split(),
+        batch_size=batch_size,
+        augment=True,
+        central_crop_size=None)
+
+    with self.test_session() as sess, queues.QueueRunners(sess):
+      images_np, labels_np = sess.run([data.images, data.labels_one_hot])
+
+    self.assertEqual(images_np.shape, (batch_size, 150, 600, 3))
+    self.assertEqual(labels_np.shape, (batch_size, 37, 134))
+
+  def test_optionally_applies_central_crop(self):
+    batch_size = 4
+    data = data_provider.get_data(
+        dataset=datasets.fsns_test.get_test_split(),
+        batch_size=batch_size,
+        augment=True,
+        central_crop_size=(500, 100))
+
+    with self.test_session() as sess, queues.QueueRunners(sess):
+      images_np = sess.run(data.images)
+
+    self.assertEqual(images_np.shape, (batch_size, 100, 500, 3))
+
+
+if __name__ == '__main__':
+  tf.test.main()

attention_ocr/python/datasets/__init__.py

+# Copyright 2017 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.
+# ==============================================================================
+
+import fsns
+import fsns_test
+
+__all__ = [fsns, fsns_test]

attention_ocr/python/datasets/fsns.py

+# -*- coding: utf-8 -*-
+# Copyright 2017 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.
+# ==============================================================================
+
+"""Configuration to read FSNS dataset https://goo.gl/3Ldm8v."""
+
+import os
+import re
+import tensorflow as tf
+from tensorflow.contrib import slim
+import logging
+
+DEFAULT_DATASET_DIR = os.path.join(os.path.dirname(__file__), 'data/fsns')
+
+# The dataset configuration, should be used only as a default value.
+DEFAULT_CONFIG = {
+    'name': 'FSNS',
+    'splits': {
+        'train': {
+            'size': 1044868,
+            'pattern': 'train/train*'
+        },
+        'test': {
+            'size': 20404,
+            'pattern': 'test/test*'
+        },
+        'validation': {
+            'size': 16150,
+            'pattern': 'validation/validation*'
+        }
+    },
+    'charset_filename': 'charset_size=134.txt',
+    'image_shape': (150, 600, 3),
+    'num_of_views': 4,
+    'max_sequence_length': 37,
+    'null_code': 133,
+    'items_to_descriptions': {
+        'image': 'A [150 x 600 x 3] color image.',
+        'label': 'Characters codes.',
+        'text': 'A unicode string.',
+        'length': 'A length of the encoded text.',
+        'num_of_views': 'A number of different views stored within the image.'
+    }
+}
+
+
+def read_charset(filename, null_character=u'\u2591'):
+  """Reads a charset definition from a tab separated text file.
+
+  charset file has to have format compatible with the FSNS dataset.
+
+  Args:
+    filename: a path to the charset file.
+    null_character: a unicode character used to replace '<null>' character. the
+      default value is a light shade block '░'.
+
+  Returns:
+    a dictionary with keys equal to character codes and values - unicode
+    characters.
+  """
+  pattern = re.compile(r'(\d+)\t(.+)')
+  charset = {}
+  with tf.gfile.GFile(filename) as f:
+    for i, line in enumerate(f):
+      m = pattern.match(line)
+      if m is None:
+        logging.warning('incorrect charset file. line #%d: %s', i, line)
+        continue
+      code = int(m.group(1))
+      char = m.group(2).decode('utf-8')
+      if char == '<nul>':
+        char = null_character
+      charset[code] = char
+  return charset
+
+
+class _NumOfViewsHandler(slim.tfexample_decoder.ItemHandler):
+  """Convenience handler to determine number of views stored in an image."""
+
+  def __init__(self, width_key, original_width_key, num_of_views):
+    super(_NumOfViewsHandler, self).__init__([width_key, original_width_key])
+    self._width_key = width_key
+    self._original_width_key = original_width_key
+    self._num_of_views = num_of_views
+
+  def tensors_to_item(self, keys_to_tensors):
+    return tf.to_int64(
+        self._num_of_views * keys_to_tensors[self._original_width_key] /
+        keys_to_tensors[self._width_key])
+
+
+def get_split(split_name, dataset_dir=None, config=None):
+  """Returns a dataset tuple for FSNS dataset.
+
+  Args:
+    split_name: A train/test split name.
+    dataset_dir: The base directory of the dataset sources, by default it uses
+      a predefined CNS path (see DEFAULT_DATASET_DIR).
+    config: A dictionary with dataset configuration. If None - will use the
+      DEFAULT_CONFIG.
+
+  Returns:
+    A `Dataset` namedtuple.
+
+  Raises:
+    ValueError: if `split_name` is not a valid train/test split.
+  """
+  if not dataset_dir:
+    dataset_dir = DEFAULT_DATASET_DIR
+
+  if not config:
+    config = DEFAULT_CONFIG
+
+  if split_name not in config['splits']:
+    raise ValueError('split name %s was not recognized.' % split_name)
+
+  logging.info('Using %s dataset split_name=%s dataset_dir=%s', config['name'],
+               split_name, dataset_dir)
+
+  # Ignores the 'image/height' feature.
+  zero = tf.zeros([1], dtype=tf.int64)
+  keys_to_features = {
+      'image/encoded':
+      tf.FixedLenFeature((), tf.string, default_value=''),
+      'image/format':
+      tf.FixedLenFeature((), tf.string, default_value='png'),
+      'image/width':
+      tf.FixedLenFeature([1], tf.int64, default_value=zero),
+      'image/orig_width':
+      tf.FixedLenFeature([1], tf.int64, default_value=zero),
+      'image/class':
+      tf.FixedLenFeature([config['max_sequence_length']], tf.int64),
+      'image/unpadded_class':
+      tf.VarLenFeature(tf.int64),
+      'image/text':
+      tf.FixedLenFeature([1], tf.string, default_value=''),
+  }
+  items_to_handlers = {
+      'image':
+      slim.tfexample_decoder.Image(
+          shape=config['image_shape'],
+          image_key='image/encoded',
+          format_key='image/format'),
+      'label':
+      slim.tfexample_decoder.Tensor(tensor_key='image/class'),
+      'text':
+      slim.tfexample_decoder.Tensor(tensor_key='image/text'),
+      'num_of_views':
+      _NumOfViewsHandler(
+          width_key='image/width',
+          original_width_key='image/orig_width',
+          num_of_views=config['num_of_views'])
+  }
+  decoder = slim.tfexample_decoder.TFExampleDecoder(keys_to_features,
+                                                    items_to_handlers)
+  charset_file = os.path.join(dataset_dir, config['charset_filename'])
+  charset = read_charset(charset_file)
+  file_pattern = os.path.join(dataset_dir,
+                              config['splits'][split_name]['pattern'])
+  return slim.dataset.Dataset(
+      data_sources=file_pattern,
+      reader=tf.TFRecordReader,
+      decoder=decoder,
+      num_samples=config['splits'][split_name]['size'],
+      items_to_descriptions=config['items_to_descriptions'],
+      #  additional parameters for convenience.
+      charset=charset,
+      num_char_classes=len(charset),
+      num_of_views=config['num_of_views'],
+      max_sequence_length=config['max_sequence_length'],
+      null_code=config['null_code'])

attention_ocr/python/datasets/fsns_test.py

+# Copyright 2017 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 FSNS datasets module."""
+
+import collections
+import os
+import tensorflow as tf
+from tensorflow.contrib import slim
+
+import fsns
+import unittest_utils
+
+FLAGS = tf.flags.FLAGS
+
+
+def get_test_split():
+  config = fsns.DEFAULT_CONFIG.copy()
+  config['splits'] = {'test': {'size': 50, 'pattern': 'fsns-00000-of-00001'}}
+  return fsns.get_split('test', dataset_dir(), config)
+
+
+def dataset_dir():
+  return os.path.join(os.path.dirname(__file__), 'testdata/fsns')
+
+
+class FsnsTest(tf.test.TestCase):
+  def test_decodes_example_proto(self):
+    expected_label = range(37)
+    expected_image, encoded = unittest_utils.create_random_image(
+        'PNG', shape=(150, 600, 3))
+    serialized = unittest_utils.create_serialized_example({
+        'image/encoded': [encoded],
+        'image/format': ['PNG'],
+        'image/class':
+        expected_label,
+        'image/unpadded_class':
+        range(10),
+        'image/text': ['Raw text'],
+        'image/orig_width': [150],
+        'image/width': [600]
+    })
+
+    decoder = fsns.get_split('train', dataset_dir()).decoder
+    with self.test_session() as sess:
+      data_tuple = collections.namedtuple('DecodedData', decoder.list_items())
+      data = sess.run(data_tuple(*decoder.decode(serialized)))
+
+    self.assertAllEqual(expected_image, data.image)
+    self.assertAllEqual(expected_label, data.label)
+    self.assertEqual(['Raw text'], data.text)
+    self.assertEqual([1], data.num_of_views)
+
+  def test_label_has_shape_defined(self):
+    serialized = 'fake'
+    decoder = fsns.get_split('train', dataset_dir()).decoder
+
+    [label_tf] = decoder.decode(serialized, ['label'])
+
+    self.assertEqual(label_tf.get_shape().dims[0], 37)
+
+  def test_dataset_tuple_has_all_extra_attributes(self):
+    dataset = fsns.get_split('train', dataset_dir())
+
+    self.assertTrue(dataset.charset)
+    self.assertTrue(dataset.num_char_classes)
+    self.assertTrue(dataset.num_of_views)
+    self.assertTrue(dataset.max_sequence_length)
+    self.assertTrue(dataset.null_code)
+
+  def test_can_use_the_test_data(self):
+    batch_size = 1
+    dataset = get_test_split()
+    provider = slim.dataset_data_provider.DatasetDataProvider(
+        dataset,
+        shuffle=True,
+        common_queue_capacity=2 * batch_size,
+        common_queue_min=batch_size)
+    image_tf, label_tf = provider.get(['image', 'label'])
+
+    with self.test_session() as sess:
+      sess.run(tf.global_variables_initializer())
+      with slim.queues.QueueRunners(sess):
+        image_np, label_np = sess.run([image_tf, label_tf])
+
+    self.assertEqual((150, 600, 3), image_np.shape)
+    self.assertEqual((37, ), label_np.shape)
+
+
+if __name__ == '__main__':
+  tf.test.main()

attention_ocr/python/datasets/testdata/fsns/charset_size=134.txt

+0	 
+133	<nul>
+1	l
+2	’
+3	é
+4	t
+5	e
+6	i
+7	n
+8	s
+9	x
+10	g
+11	u
+12	o
+13	1
+14	8
+15	7
+16	0
+17	-
+18	.
+19	p
+20	a
+21	r
+22	è
+23	d
+24	c
+25	V
+26	v
+27	b
+28	m
+29	)
+30	C
+31	z
+32	S
+33	y
+34	,
+35	k
+36	É
+37	A
+38	h
+39	E
+40	»
+41	D
+42	/
+43	H
+44	M
+45	(
+46	G
+47	P
+48	ç
+2	'
+49	R
+50	f
+51	"
+52	2
+53	j
+54	|
+55	N
+56	6
+57	°
+58	5
+59	T
+60	O
+61	U
+62	3
+63	%
+64	9
+65	q
+66	Z
+67	B
+68	K
+69	w
+70	W
+71	:
+72	4
+73	L
+74	F
+75	]
+76	ï
+2	‘
+77	I
+78	J
+79	ä
+80	î
+81	;
+82	à
+83	ê
+84	X
+85	ü
+86	Y
+87	ô
+88	=
+89	+
+90	\
+91	{
+92	}
+93	_
+94	Q
+95	œ
+96	ñ
+97	*
+98	!
+99	Ü
+51	“
+100	â
+101	Ç
+102	Œ
+103	û
+104	?
+105	$
+106	ë
+107	«
+108	€
+109	&
+110	<
+51	”
+111	æ
+112	#
+113	®
+114	Â
+115	È
+116	>
+117	[
+17	—
+118	Æ
+119	ù
+120	Î
+121	Ô
+122	ÿ
+123	À
+124	Ê
+125	@
+126	Ï
+127	©
+128	Ë
+129	Ù
+130	£
+131	Ÿ
+132	Û

attention_ocr/python/datasets/testdata/fsns/links.txt

+http://download.tensorflow.org/data/fsns-20160927/testdata/fsns-00000-of-00001

attention_ocr/python/datasets/unittest_utils.py

+# Copyright 2017 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.
+# ==============================================================================
+
+"""Functions to make unit testing easier."""
+
+import StringIO
+import numpy as np
+from PIL import Image as PILImage
+import tensorflow as tf
+
+
+def create_random_image(image_format, shape):
+  """Creates an image with random values.
+
+  Args:
+    image_format: An image format (PNG or JPEG).
+    shape: A tuple with image shape (including channels).
+
+  Returns:
+    A tuple (<numpy ndarray>, <a string with encoded image>)
+  """
+  image = np.random.randint(low=0, high=255, size=shape, dtype='uint8')
+  io = StringIO.StringIO()
+  image_pil = PILImage.fromarray(image)
+  image_pil.save(io, image_format, subsampling=0, quality=100)
+  return image, io.getvalue()
+
+
+def create_serialized_example(name_to_values):
+  """Creates a tf.Example proto using a dictionary.
+
+  It automatically detects type of values and define a corresponding feature.
+
+  Args:
+    name_to_values: A dictionary.
+
+  Returns:
+    tf.Example proto.
+  """
+  example = tf.train.Example()
+  for name, values in name_to_values.items():
+    feature = example.features.feature[name]
+    if isinstance(values[0], str):
+      add = feature.bytes_list.value.extend
+    elif isinstance(values[0], float):
+      add = feature.float32_list.value.extend
+    elif isinstance(values[0], int):
+      add = feature.int64_list.value.extend
+    else:
+      raise AssertionError('Unsupported type: %s' % type(values[0]))
+    add(values)
+  return example.SerializeToString()

attention_ocr/python/datasets/unittest_utils_test.py

+# Copyright 2017 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 unittest_utils."""
+import StringIO
+
+import numpy as np
+from PIL import Image as PILImage
+import tensorflow as tf
+
+import unittest_utils
+
+
+class UnittestUtilsTest(tf.test.TestCase):
+  def test_creates_an_image_of_specified_shape(self):
+    image, _ = unittest_utils.create_random_image('PNG', (10, 20, 3))
+    self.assertEqual(image.shape, (10, 20, 3))
+
+  def test_encoded_image_corresponds_to_numpy_array(self):
+    image, encoded = unittest_utils.create_random_image('PNG', (20, 10, 3))
+    pil_image = PILImage.open(StringIO.StringIO(encoded))
+    self.assertAllEqual(image, np.array(pil_image))
+
+  def test_created_example_has_correct_values(self):
+    example_serialized = unittest_utils.create_serialized_example({
+        'labels': [1, 2, 3],
+        'data': ['FAKE']
+    })
+    example = tf.train.Example()
+    example.ParseFromString(example_serialized)
+    self.assertProtoEquals("""
+      features {
+        feature {
+          key: "labels"
+           value { int64_list {
+             value: 1
+             value: 2
+             value: 3
+           }}
+         }
+         feature {
+           key: "data"
+           value { bytes_list {
+             value: "FAKE"
+           }}
+         }
+      }
+    """, example)
+
+
+if __name__ == '__main__':
+  tf.test.main()

attention_ocr/python/eval.py

+# Copyright 2017 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.
+# ==============================================================================
+
+"""Script to evaluate a trained Attention OCR model.
+
+A simple usage example:
+python eval.py
+"""
+import tensorflow as tf
+from tensorflow.contrib import slim
+from tensorflow import app
+from tensorflow.python.platform import flags
+
+import data_provider
+import common_flags
+
+FLAGS = flags.FLAGS
+common_flags.define()
+
+# yapf: disable
+flags.DEFINE_integer('num_batches', 100,
+                     'Number of batches to run eval for.')
+
+flags.DEFINE_string('eval_log_dir', '/tmp/attention_ocr/eval',
+                    'Directory where the evaluation results are saved to.')
+
+flags.DEFINE_integer('eval_interval_secs', 60,
+                     'Frequency in seconds to run evaluations.')
+
+flags.DEFINE_integer('number_of_steps', None,
+                     'Number of times to run evaluation.')
+# yapf: enable
+
+
+def main(_):
+  if not tf.gfile.Exists(FLAGS.eval_log_dir):
+    tf.gfile.MakeDirs(FLAGS.eval_log_dir)
+
+  dataset = common_flags.create_dataset(split_name=FLAGS.split_name)
+  model = common_flags.create_model(dataset.num_char_classes,
+                                    dataset.max_sequence_length,
+                                    dataset.num_of_views, dataset.null_code)
+  data = data_provider.get_data(
+      dataset,
+      FLAGS.batch_size,
+      augment=False,
+      central_crop_size=common_flags.get_crop_size())
+  endpoints = model.create_base(data.images, labels_one_hot=None)
+  model.create_loss(data, endpoints)
+  eval_ops = model.create_summaries(
+      data, endpoints, dataset.charset, is_training=False)
+  slim.get_or_create_global_step()
+  session_config = tf.ConfigProto(device_count={"GPU": 0})
+  slim.evaluation.evaluation_loop(
+      master=FLAGS.master,
+      checkpoint_dir=FLAGS.train_log_dir,
+      logdir=FLAGS.eval_log_dir,
+      eval_op=eval_ops,
+      num_evals=FLAGS.num_batches,
+      eval_interval_secs=FLAGS.eval_interval_secs,
+      max_number_of_evaluations=FLAGS.number_of_steps,
+      session_config=session_config)
+
+
+if __name__ == '__main__':
+  app.run()

attention_ocr/python/inception_preprocessing.py

+# Copyright 2017 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.
+# ==============================================================================
+"""Provides utilities to preprocess images for the Inception networks."""
+
+# TODO(gorban): add as a dependency, when slim or tensorflow/models are pipfied
+# Source:
+# https://raw.githubusercontent.com/tensorflow/models/a9d0e6e8923a4/slim/preprocessing/inception_preprocessing.py
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import tensorflow as tf
+
+from tensorflow.python.ops import control_flow_ops
+
+
+def apply_with_random_selector(x, func, num_cases):
+  """Computes func(x, sel), with sel sampled from [0...num_cases-1].
+
+  Args:
+    x: input Tensor.
+    func: Python function to apply.
+    num_cases: Python int32, number of cases to sample sel from.
+
+  Returns:
+    The result of func(x, sel), where func receives the value of the
+    selector as a python integer, but sel is sampled dynamically.
+  """
+  sel = tf.random_uniform([], maxval=num_cases, dtype=tf.int32)
+  # Pass the real x only to one of the func calls.
+  return control_flow_ops.merge([
+      func(control_flow_ops.switch(x, tf.equal(sel, case))[1], case)
+      for case in range(num_cases)
+  ])[0]
+
+
+def distort_color(image, color_ordering=0, fast_mode=True, scope=None):
+  """Distort the color of a Tensor image.
+
+  Each color distortion is non-commutative and thus ordering of the color ops
+  matters. Ideally we would randomly permute the ordering of the color ops.
+  Rather then adding that level of complication, we select a distinct ordering
+  of color ops for each preprocessing thread.
+
+  Args:
+    image: 3-D Tensor containing single image in [0, 1].
+    color_ordering: Python int, a type of distortion (valid values: 0-3).
+    fast_mode: Avoids slower ops (random_hue and random_contrast)
+    scope: Optional scope for name_scope.
+  Returns:
+    3-D Tensor color-distorted image on range [0, 1]
+  Raises:
+    ValueError: if color_ordering not in [0, 3]
+  """
+  with tf.name_scope(scope, 'distort_color', [image]):
+    if fast_mode:
+      if color_ordering == 0:
+        image = tf.image.random_brightness(image, max_delta=32. / 255.)
+        image = tf.image.random_saturation(image, lower=0.5, upper=1.5)
+      else:
+        image = tf.image.random_saturation(image, lower=0.5, upper=1.5)
+        image = tf.image.random_brightness(image, max_delta=32. / 255.)
+    else:
+      if color_ordering == 0:
+        image = tf.image.random_brightness(image, max_delta=32. / 255.)
+        image = tf.image.random_saturation(image, lower=0.5, upper=1.5)
+        image = tf.image.random_hue(image, max_delta=0.2)
+        image = tf.image.random_contrast(image, lower=0.5, upper=1.5)
+      elif color_ordering == 1:
+        image = tf.image.random_saturation(image, lower=0.5, upper=1.5)
+        image = tf.image.random_brightness(image, max_delta=32. / 255.)
+        image = tf.image.random_contrast(image, lower=0.5, upper=1.5)
+        image = tf.image.random_hue(image, max_delta=0.2)
+      elif color_ordering == 2:
+        image = tf.image.random_contrast(image, lower=0.5, upper=1.5)
+        image = tf.image.random_hue(image, max_delta=0.2)
+        image = tf.image.random_brightness(image, max_delta=32. / 255.)
+        image = tf.image.random_saturation(image, lower=0.5, upper=1.5)
+      elif color_ordering == 3:
+        image = tf.image.random_hue(image, max_delta=0.2)
+        image = tf.image.random_saturation(image, lower=0.5, upper=1.5)
+        image = tf.image.random_contrast(image, lower=0.5, upper=1.5)
+        image = tf.image.random_brightness(image, max_delta=32. / 255.)
+      else:
+        raise ValueError('color_ordering must be in [0, 3]')
+
+    # The random_* ops do not necessarily clamp.
+    return tf.clip_by_value(image, 0.0, 1.0)
+
+
+def distorted_bounding_box_crop(image,
+                                bbox,
+                                min_object_covered=0.1,
+                                aspect_ratio_range=(0.75, 1.33),
+                                area_range=(0.05, 1.0),
+                                max_attempts=100,
+                                scope=None):
+  """Generates cropped_image using a one of the bboxes randomly distorted.
+
+  See `tf.image.sample_distorted_bounding_box` for more documentation.
+
+  Args:
+    image: 3-D Tensor of image (it will be converted to floats in [0, 1]).
+    bbox: 3-D float Tensor of bounding boxes arranged [1, num_boxes, coords]
+      where each coordinate is [0, 1) and the coordinates are arranged
+      as [ymin, xmin, ymax, xmax]. If num_boxes is 0 then it would use the
+      whole image.
+    min_object_covered: An optional `float`. Defaults to `0.1`. The cropped
+      area of the image must contain at least this fraction of any bounding box
+      supplied.
+    aspect_ratio_range: An optional list of `floats`. The cropped area of the
+      image must have an aspect ratio = width / height within this range.
+    area_range: An optional list of `floats`. The cropped area of the image
+      must contain a fraction of the supplied image within in this range.
+    max_attempts: An optional `int`. Number of attempts at generating a cropped
+      region of the image of the specified constraints. After `max_attempts`
+      failures, return the entire image.
+    scope: Optional scope for name_scope.
+  Returns:
+    A tuple, a 3-D Tensor cropped_image and the distorted bbox
+  """
+  with tf.name_scope(scope, 'distorted_bounding_box_crop', [image, bbox]):
+    # Each bounding box has shape [1, num_boxes, box coords] and
+    # the coordinates are ordered [ymin, xmin, ymax, xmax].
+
+    # A large fraction of image datasets contain a human-annotated bounding
+    # box delineating the region of the image containing the object of interest.
+    # We choose to create a new bounding box for the object which is a randomly
+    # distorted version of the human-annotated bounding box that obeys an
+    # allowed range of aspect ratios, sizes and overlap with the human-annotated
+    # bounding box. If no box is supplied, then we assume the bounding box is
+    # the entire image.
+    sample_distorted_bounding_box = tf.image.sample_distorted_bounding_box(
+        tf.shape(image),
+        bounding_boxes=bbox,
+        min_object_covered=min_object_covered,
+        aspect_ratio_range=aspect_ratio_range,
+        area_range=area_range,
+        max_attempts=max_attempts,
+        use_image_if_no_bounding_boxes=True)
+    bbox_begin, bbox_size, distort_bbox = sample_distorted_bounding_box
+
+    # Crop the image to the specified bounding box.
+    cropped_image = tf.slice(image, bbox_begin, bbox_size)
+    return cropped_image, distort_bbox
+
+
+def preprocess_for_train(image,
+                         height,
+                         width,
+                         bbox,
+                         fast_mode=True,
+                         scope=None):
+  """Distort one image for training a network.
+
+  Distorting images provides a useful technique for augmenting the data
+  set during training in order to make the network invariant to aspects
+  of the image that do not effect the label.
+
+  Additionally it would create image_summaries to display the different
+  transformations applied to the image.
+
+  Args:
+    image: 3-D Tensor of image. If dtype is tf.float32 then the range should be
+      [0, 1], otherwise it would converted to tf.float32 assuming that the range
+      is [0, MAX], where MAX is largest positive representable number for
+      int(8/16/32) data type (see `tf.image.convert_image_dtype` for details).
+    height: integer
+    width: integer
+    bbox: 3-D float Tensor of bounding boxes arranged [1, num_boxes, coords]
+      where each coordinate is [0, 1) and the coordinates are arranged
+      as [ymin, xmin, ymax, xmax].
+    fast_mode: Optional boolean, if True avoids slower transformations (i.e.
+      bi-cubic resizing, random_hue or random_contrast).
+    scope: Optional scope for name_scope.
+  Returns:
+    3-D float Tensor of distorted image used for training with range [-1, 1].
+  """
+  with tf.name_scope(scope, 'distort_image', [image, height, width, bbox]):
+    if bbox is None:
+      bbox = tf.constant(
+          [0.0, 0.0, 1.0, 1.0], dtype=tf.float32, shape=[1, 1, 4])
+    if image.dtype != tf.float32:
+      image = tf.image.convert_image_dtype(image, dtype=tf.float32)
+    # Each bounding box has shape [1, num_boxes, box coords] and
+    # the coordinates are ordered [ymin, xmin, ymax, xmax].
+    image_with_box = tf.image.draw_bounding_boxes(
+        tf.expand_dims(image, 0), bbox)
+    tf.summary.image('image_with_bounding_boxes', image_with_box)
+
+    distorted_image, distorted_bbox = distorted_bounding_box_crop(image, bbox)
+    # Restore the shape since the dynamic slice based upon the bbox_size loses
+    # the third dimension.
+    distorted_image.set_shape([None, None, 3])
+    image_with_distorted_box = tf.image.draw_bounding_boxes(
+        tf.expand_dims(image, 0), distorted_bbox)
+    tf.summary.image('images_with_distorted_bounding_box',
+                     image_with_distorted_box)
+
+    # This resizing operation may distort the images because the aspect
+    # ratio is not respected. We select a resize method in a round robin
+    # fashion based on the thread number.
+    # Note that ResizeMethod contains 4 enumerated resizing methods.
+
+    # We select only 1 case for fast_mode bilinear.
+    num_resize_cases = 1 if fast_mode else 4
+    distorted_image = apply_with_random_selector(
+        distorted_image,
+        lambda x, method: tf.image.resize_images(x, [height, width], method=method),
+        num_cases=num_resize_cases)
+
+    tf.summary.image('cropped_resized_image',
+                     tf.expand_dims(distorted_image, 0))
+
+    # Randomly flip the image horizontally.
+    distorted_image = tf.image.random_flip_left_right(distorted_image)
+
+    # Randomly distort the colors. There are 4 ways to do it.
+    distorted_image = apply_with_random_selector(
+        distorted_image,
+        lambda x, ordering: distort_color(x, ordering, fast_mode),
+        num_cases=4)
+
+    tf.summary.image('final_distorted_image',
+                     tf.expand_dims(distorted_image, 0))
+    distorted_image = tf.subtract(distorted_image, 0.5)
+    distorted_image = tf.multiply(distorted_image, 2.0)
+    return distorted_image
+
+
+def preprocess_for_eval(image,
+                        height,
+                        width,
+                        central_fraction=0.875,
+                        scope=None):
+  """Prepare one image for evaluation.
+
+  If height and width are specified it would output an image with that size by
+  applying resize_bilinear.
+
+  If central_fraction is specified it would cropt the central fraction of the
+  input image.
+
+  Args:
+    image: 3-D Tensor of image. If dtype is tf.float32 then the range should be
+      [0, 1], otherwise it would converted to tf.float32 assuming that the range
+      is [0, MAX], where MAX is largest positive representable number for
+      int(8/16/32) data type (see `tf.image.convert_image_dtype` for details)
+    height: integer
+    width: integer
+    central_fraction: Optional Float, fraction of the image to crop.
+    scope: Optional scope for name_scope.
+  Returns:
+    3-D float Tensor of prepared image.
+  """
+  with tf.name_scope(scope, 'eval_image', [image, height, width]):
+    if image.dtype != tf.float32:
+      image = tf.image.convert_image_dtype(image, dtype=tf.float32)
+    # Crop the central region of the image with an area containing 87.5% of
+    # the original image.
+    if central_fraction:
+      image = tf.image.central_crop(image, central_fraction=central_fraction)
+
+    if height and width:
+      # Resize the image to the specified height and width.
+      image = tf.expand_dims(image, 0)
+      image = tf.image.resize_bilinear(
+          image, [height, width], align_corners=False)
+      image = tf.squeeze(image, [0])
+    image = tf.subtract(image, 0.5)
+    image = tf.multiply(image, 2.0)
+    return image
+
+
+def preprocess_image(image,
+                     height,
+                     width,
+                     is_training=False,
+                     bbox=None,
+                     fast_mode=True):
+  """Pre-process one image for training or evaluation.
+
+  Args:
+    image: 3-D Tensor [height, width, channels] with the image.
+    height: integer, image expected height.
+    width: integer, image expected width.
+    is_training: Boolean. If true it would transform an image for train,
+      otherwise it would transform it for evaluation.
+    bbox: 3-D float Tensor of bounding boxes arranged [1, num_boxes, coords]
+      where each coordinate is [0, 1) and the coordinates are arranged as
+      [ymin, xmin, ymax, xmax].
+    fast_mode: Optional boolean, if True avoids slower transformations.
+
+  Returns:
+    3-D float Tensor containing an appropriately scaled image
+
+  Raises:
+    ValueError: if user does not provide bounding box
+  """
+  if is_training:
+    return preprocess_for_train(image, height, width, bbox, fast_mode)
+  else:
+    return preprocess_for_eval(image, height, width)

attention_ocr/python/metrics.py

+# Copyright 2017 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.
+# ==============================================================================
+
+"""Quality metrics for the model."""
+
+import tensorflow as tf
+
+
+def char_accuracy(predictions, targets, rej_char, streaming=False):
+  """Computes character level accuracy.
+
+  Both predictions and targets should have the same shape
+  [batch_size x seq_length].
+
+  Args:
+    predictions: predicted characters ids.
+    targets: ground truth character ids.
+    rej_char: the character id used to mark an empty element (end of sequence).
+    streaming: if True, uses the streaming mean from the slim.metric module.
+
+  Returns:
+    a update_ops for execution and value tensor whose value on evaluation
+    returns the total character accuracy.
+  """
+  with tf.variable_scope('CharAccuracy'):
+    predictions.get_shape().assert_is_compatible_with(targets.get_shape())
+
+    targets = tf.to_int32(targets)
+    const_rej_char = tf.constant(rej_char, shape=targets.get_shape())
+    weights = tf.to_float(tf.not_equal(targets, const_rej_char))
+    correct_chars = tf.to_float(tf.equal(predictions, targets))
+    accuracy_per_example = tf.div(
+        tf.reduce_sum(tf.multiply(correct_chars, weights), 1),
+        tf.reduce_sum(weights, 1))
+    if streaming:
+      return tf.contrib.metrics.streaming_mean(accuracy_per_example)
+    else:
+      return tf.reduce_mean(accuracy_per_example)
+
+
+def sequence_accuracy(predictions, targets, rej_char, streaming=False):
+  """Computes sequence level accuracy.
+
+  Both input tensors should have the same shape: [batch_size x seq_length].
+
+  Args:
+    predictions: predicted character classes.
+    targets: ground truth character classes.
+    rej_char: the character id used to mark empty element (end of sequence).
+    streaming: if True, uses the streaming mean from the slim.metric module.
+
+  Returns:
+    a update_ops for execution and value tensor whose value on evaluation
+    returns the total sequence accuracy.
+  """
+
+  with tf.variable_scope('SequenceAccuracy'):
+    predictions.get_shape().assert_is_compatible_with(targets.get_shape())
+
+    targets = tf.to_int32(targets)
+    const_rej_char = tf.constant(
+        rej_char, shape=targets.get_shape(), dtype=tf.int32)
+    include_mask = tf.not_equal(targets, const_rej_char)
+    include_predictions = tf.to_int32(
+        tf.where(include_mask, predictions,
+                 tf.zeros_like(predictions) + rej_char))
+    correct_chars = tf.to_float(tf.equal(include_predictions, targets))
+    correct_chars_counts = tf.cast(
+        tf.reduce_sum(correct_chars, reduction_indices=[1]), dtype=tf.int32)
+    target_length = targets.get_shape().dims[1].value
+    target_chars_counts = tf.constant(
+        target_length, shape=correct_chars_counts.get_shape())
+    accuracy_per_example = tf.to_float(
+        tf.equal(correct_chars_counts, target_chars_counts))
+    if streaming:
+      return tf.contrib.metrics.streaming_mean(accuracy_per_example)
+    else:
+      return tf.reduce_mean(accuracy_per_example)

attention_ocr/python/metrics_test.py

+# Copyright 2017 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 the metrics module."""
+import contextlib
+import numpy as np
+import tensorflow as tf
+
+import metrics
+
+
+class AccuracyTest(tf.test.TestCase):
+  def setUp(self):
+    tf.test.TestCase.setUp(self)
+    self.rng = np.random.RandomState([11, 23, 50])
+    self.num_char_classes = 3
+    self.batch_size = 4
+    self.seq_length = 5
+    self.rej_char = 42
+
+  @contextlib.contextmanager
+  def initialized_session(self):
+    """Wrapper for test session context manager with required initialization.
+
+    Yields:
+      A session object that should be used as a context manager.
+    """
+    with self.test_session() as sess:
+      sess.run(tf.global_variables_initializer())
+      sess.run(tf.local_variables_initializer())
+      yield sess
+
+  def _fake_labels(self):
+    return self.rng.randint(
+        low=0,
+        high=self.num_char_classes,
+        size=(self.batch_size, self.seq_length),
+        dtype='int32')
+
+  def _incorrect_copy(self, values, bad_indexes):
+    incorrect = np.copy(values)
+    incorrect[bad_indexes] = values[bad_indexes] + 1
+    return incorrect
+
+  def test_sequence_accuracy_identical_samples(self):
+    labels_tf = tf.convert_to_tensor(self._fake_labels())
+
+    accuracy_tf = metrics.sequence_accuracy(labels_tf, labels_tf,
+                                            self.rej_char)
+    with self.initialized_session() as sess:
+      accuracy_np = sess.run(accuracy_tf)
+
+    self.assertAlmostEqual(accuracy_np, 1.0)
+
+  def test_sequence_accuracy_one_char_difference(self):
+    ground_truth_np = self._fake_labels()
+    ground_truth_tf = tf.convert_to_tensor(ground_truth_np)
+    prediction_tf = tf.convert_to_tensor(
+        self._incorrect_copy(ground_truth_np, bad_indexes=((0, 0))))
+
+    accuracy_tf = metrics.sequence_accuracy(prediction_tf, ground_truth_tf,
+                                            self.rej_char)
+    with self.initialized_session() as sess:
+      accuracy_np = sess.run(accuracy_tf)
+
+    # 1 of 4 sequences is incorrect.
+    self.assertAlmostEqual(accuracy_np, 1.0 - 1.0 / self.batch_size)
+
+  def test_char_accuracy_one_char_difference_with_padding(self):
+    ground_truth_np = self._fake_labels()
+    ground_truth_tf = tf.convert_to_tensor(ground_truth_np)
+    prediction_tf = tf.convert_to_tensor(
+        self._incorrect_copy(ground_truth_np, bad_indexes=((0, 0))))
+
+    accuracy_tf = metrics.char_accuracy(prediction_tf, ground_truth_tf,
+                                        self.rej_char)
+    with self.initialized_session() as sess:
+      accuracy_np = sess.run(accuracy_tf)
+
+    chars_count = self.seq_length * self.batch_size
+    self.assertAlmostEqual(accuracy_np, 1.0 - 1.0 / chars_count)
+
+
+if __name__ == '__main__':
+  tf.test.main()

attention_ocr/python/model.py

+# Copyright 2017 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.
+# ==============================================================================
+
+"""Functions to build the Attention OCR model.
+
+Usage example:
+  ocr_model = model.Model(num_char_classes, seq_length, num_of_views)
+
+  data = ... # create namedtuple InputEndpoints
+  endpoints = model.create_base(data.images, data.labels_one_hot)
+  # endpoints.predicted_chars is a tensor with predicted character codes.
+  total_loss = model.create_loss(data, endpoints)
+"""
+import sys
+import collections
+import logging
+import tensorflow as tf
+from tensorflow.contrib import slim
+from tensorflow.contrib.slim.nets import inception
+
+import metrics
+import sequence_layers
+import utils
+
+
+OutputEndpoints = collections.namedtuple('OutputEndpoints', [
+    'chars_logit', 'chars_log_prob', 'predicted_chars', 'predicted_scores'
+])
+
+# TODO(gorban): replace with tf.HParams when it is released.
+ModelParams = collections.namedtuple('ModelParams', [
+    'num_char_classes', 'seq_length', 'num_views', 'null_code'
+])
+
+ConvTowerParams = collections.namedtuple('ConvTowerParams', ['final_endpoint'])
+
+SequenceLogitsParams = collections.namedtuple('SequenceLogitsParams', [
+    'use_attention', 'use_autoregression', 'num_lstm_units', 'weight_decay',
+    'lstm_state_clip_value'
+])
+
+SequenceLossParams = collections.namedtuple('SequenceLossParams', [
+    'label_smoothing', 'ignore_nulls', 'average_across_timesteps'
+])
+
+
+def _dict_to_array(id_to_char, default_character):
+  num_char_classes = max(id_to_char.keys()) + 1
+  array = [default_character] * num_char_classes
+  for k, v in id_to_char.iteritems():
+    array[k] = v
+  return array
+
+
+class CharsetMapper(object):
+  """A simple class to map tensor ids into strings.
+
+    It works only when the character set is 1:1 mapping between individual
+    characters and individual ids.
+
+    Make sure you call tf.tables_initializer().run() as part of the init op.
+    """
+
+  def __init__(self, charset, default_character='?'):
+    """Creates a lookup table.
+
+    Args:
+      charset: a dictionary with id-to-character mapping.
+    """
+    mapping_strings = tf.constant(_dict_to_array(charset, default_character))
+    self.table = tf.contrib.lookup.index_to_string_table_from_tensor(
+        mapping=mapping_strings, default_value=default_character)
+
+  def get_text(self, ids):
+    """Returns a string corresponding to a sequence of character ids.
+
+        Args:
+          ids: a tensor with shape [batch_size, max_sequence_length]
+        """
+    return tf.reduce_join(
+        self.table.lookup(tf.to_int64(ids)), reduction_indices=1)
+
+
+def get_softmax_loss_fn(label_smoothing):
+  """Returns sparse or dense loss function depending on the label_smoothing.
+
+    Args:
+      label_smoothing: weight for label smoothing
+
+    Returns:
+      a function which takes labels and predictions as arguments and returns
+      a softmax loss for the selected type of labels (sparse or dense).
+    """
+  if label_smoothing > 0:
+
+    def loss_fn(labels, logits):
+      return (tf.nn.softmax_cross_entropy_with_logits(
+          logits=logits, labels=labels))
+  else:
+
+    def loss_fn(labels, logits):
+      return tf.nn.sparse_softmax_cross_entropy_with_logits(
+          logits=logits, labels=labels)
+
+  return loss_fn
+
+
+class Model(object):
+  """Class to create the Attention OCR Model."""
+
+  def __init__(self,
+               num_char_classes,
+               seq_length,
+               num_views,
+               null_code,
+               mparams=None):
+    """Initialized model parameters.
+
+    Args:
+      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.
+    """
+    super(Model, self).__init__()
+    self._params = ModelParams(
+        num_char_classes=num_char_classes,
+        seq_length=seq_length,
+        num_views=num_views,
+        null_code=null_code)
+    self._mparams = self.default_mparams()
+    if mparams:
+      self._mparams.update(mparams)
+
+  def default_mparams(self):
+    return {
+        'conv_tower_fn':
+        ConvTowerParams(final_endpoint='Mixed_5d'),
+        'sequence_logit_fn':
+        SequenceLogitsParams(
+            use_attention=True,
+            use_autoregression=True,
+            num_lstm_units=256,
+            weight_decay=0.00004,
+            lstm_state_clip_value=10.0),
+        'sequence_loss_fn':
+        SequenceLossParams(
+            label_smoothing=0.1,
+            ignore_nulls=True,
+            average_across_timesteps=False)
+    }
+
+  def set_mparam(self, function, **kwargs):
+    self._mparams[function] = self._mparams[function]._replace(**kwargs)
+
+  def conv_tower_fn(self, images, is_training=True, reuse=None):
+    """Computes convolutional features using the InceptionV3 model.
+
+    Args:
+      images: A tensor of shape [batch_size, height, width, channels].
+      is_training: whether is training or not.
+      reuse: whether or not the network and its variables should be reused. To
+        be able to reuse 'scope' must be given.
+
+    Returns:
+      A tensor of shape [batch_size, OH, OW, N], where OWxOH is resolution of
+      output feature map and N is number of output features (depends on the
+      network architecture).
+    """
+    mparams = self._mparams['conv_tower_fn']
+    logging.debug('Using final_endpoint=%s', mparams.final_endpoint)
+    with tf.variable_scope('conv_tower_fn/INCE'):
+      if reuse:
+        tf.get_variable_scope().reuse_variables()
+      with slim.arg_scope(inception.inception_v3_arg_scope()):
+        net, _ = inception.inception_v3_base(
+            images, final_endpoint=mparams.final_endpoint)
+      return net
+
+  def _create_lstm_inputs(self, net):
+    """Splits an input tensor into a list of tensors (features).
+
+    Args:
+      net: A feature map of shape [batch_size, num_features, feature_size].
+
+    Raises:
+      AssertionError: if num_features is less than seq_length.
+
+    Returns:
+      A list with seq_length tensors of shape [batch_size, feature_size]
+    """
+    num_features = net.get_shape().dims[1].value
+    if num_features < self._params.seq_length:
+      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()))
+    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())
+      net = tf.slice(net, [0, 0, 0], [-1, self._params.seq_length, -1])
+
+    return tf.unstack(net, axis=1)
+
+  def sequence_logit_fn(self, net, labels_one_hot):
+    mparams = self._mparams['sequence_logit_fn']
+    # TODO(gorban): remove /alias suffixes from the scopes.
+    with tf.variable_scope('sequence_logit_fn/SQLR'):
+      layer_class = sequence_layers.get_layer_class(mparams.use_attention,
+                                                    mparams.use_autoregression)
+      layer = layer_class(net, labels_one_hot, self._params, mparams)
+      return layer.create_logits()
+
+  def max_pool_views(self, nets_list):
+    """Max pool across all nets in spatial dimensions.
+
+    Args:
+      nets_list: A list of 4D tensors with identical size.
+
+    Returns:
+      A tensor with the same size as any input tensors.
+    """
+    batch_size, height, width, num_features = [
+        d.value for d in nets_list[0].get_shape().dims
+    ]
+    xy_flat_shape = (batch_size, 1, height * width, num_features)
+    nets_for_merge = []
+    with tf.variable_scope('max_pool_views', values=nets_list):
+      for net in nets_list:
+        nets_for_merge.append(tf.reshape(net, xy_flat_shape))
+      merged_net = tf.concat(nets_for_merge, 1)
+      net = slim.max_pool2d(
+          merged_net, kernel_size=[len(nets_list), 1], stride=1)
+      net = tf.reshape(net, (batch_size, height, width, num_features))
+    return net
+
+  def pool_views_fn(self, nets):
+    """Combines output of multiple convolutional towers into a single tensor.
+
+    It stacks towers one on top another (in height dim) in a 4x1 grid.
+    The order is arbitrary design choice and shouldn't matter much.
+
+    Args:
+      nets: list of tensors of shape=[batch_size, height, width, num_features].
+
+    Returns:
+      A tensor of shape [batch_size, seq_length, features_size].
+    """
+    with tf.variable_scope('pool_views_fn/STCK'):
+      net = tf.concat(nets, 1)
+      batch_size = net.get_shape().dims[0].value
+      feature_size = net.get_shape().dims[3].value
+      return tf.reshape(net, [batch_size, -1, 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]
+
+    Returns:
+      A tuple (ids, log_prob, scores), where:
+        ids - predicted characters, a int32 tensor with shape
+          [batch_size x seq_length];
+        log_prob - a log probability of all characters, a float tensor with
+          shape [batch_size, seq_length, num_char_classes];
+        scores - corresponding confidence scores for characters, a float
+        tensor
+          with shape [batch_size x seq_length].
+    """
+    log_prob = utils.logits_to_log_prob(chars_logit)
+    ids = tf.to_int32(tf.argmax(log_prob, dimension=2), name='predicted_chars')
+    mask = tf.cast(
+        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))
+    return ids, log_prob, scores
+
+  def create_base(self,
+                  images,
+                  labels_one_hot,
+                  scope='AttentionOcr_v1',
+                  reuse=None):
+    """Creates a base part of the Model (no gradients, losses or summaries).
+
+    Args:
+      images: A tensor of shape [batch_size, height, width, channels].
+      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.
+      reuse: whether or not the network and its variables should be reused. To
+        be able to reuse 'scope' must be given.
+
+    Returns:
+      A named tuple OutputEndpoints.
+    """
+    logging.debug('images: %s', images)
+    is_training = labels_one_hot is not None
+    with tf.variable_scope(scope, reuse=reuse):
+      views = tf.split(
+          value=images, num_or_size_splits=self._params.num_views, axis=2)
+      logging.debug('Views=%d single view: %s', len(views), views[0])
+
+      nets = [
+          self.conv_tower_fn(v, is_training, reuse=(i != 0))
+          for i, v in enumerate(views)
+      ]
+      logging.debug('Conv tower: %s', nets[0])
+
+      net = self.pool_views_fn(nets)
+      logging.debug('Pooled views: %s', net)
+
+      chars_logit = self.sequence_logit_fn(net, labels_one_hot)
+      logging.debug('chars_logit: %s', chars_logit)
+
+      predicted_chars, chars_log_prob, predicted_scores = (
+          self.char_predictions(chars_logit))
+
+    return OutputEndpoints(
+        chars_logit=chars_logit,
+        chars_log_prob=chars_log_prob,
+        predicted_chars=predicted_chars,
+        predicted_scores=predicted_scores)
+
+  def create_loss(self, data, endpoints):
+    """Creates all losses required to train the model.
+
+    Args:
+      data: InputEndpoints namedtuple.
+      endpoints: Model namedtuple.
+
+    Returns:
+      Total loss.
+    """
+    # NOTE: the return value of ModelLoss is not used directly for the
+    # gradient computation because under the hood it calls slim.losses.AddLoss,
+    # which registers the loss in an internal collection and later returns it
+    # as part of GetTotalLoss. We need to use total loss because model may have
+    # multiple losses including regularization losses.
+    self.sequence_loss_fn(endpoints.chars_logit, data.labels)
+    total_loss = slim.losses.get_total_loss()
+    tf.summary.scalar('TotalLoss', total_loss)
+    return total_loss
+
+  def label_smoothing_regularization(self, chars_labels, weight=0.1):
+    """Applies a label smoothing regularization.
+
+    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];
+      weight: label-smoothing regularization weight.
+
+    Returns:
+      A sensor with the same shape as the input.
+    """
+    one_hot_labels = tf.one_hot(
+        chars_labels, depth=self._params.num_char_classes, axis=-1)
+    pos_weight = 1.0 - weight
+    neg_weight = weight / self._params.num_char_classes
+    return one_hot_labels * pos_weight + neg_weight
+
+  def sequence_loss_fn(self, chars_logits, chars_labels):
+    """Loss function for char sequence.
+
+    Depending on values of hyper parameters it applies label smoothing and can
+    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];
+      mparams: method hyper parameters.
+
+    Returns:
+      A Tensor with shape [batch_size] - the log-perplexity for each sequence.
+    """
+    mparams = self._mparams['sequence_loss_fn']
+    with tf.variable_scope('sequence_loss_fn/SLF'):
+      if mparams.label_smoothing > 0:
+        smoothed_one_hot_labels = self.label_smoothing_regularization(
+            chars_labels, mparams.label_smoothing)
+        labels_list = tf.unstack(smoothed_one_hot_labels, axis=1)
+      else:
+        # NOTE: in case of sparse softmax we are not using one-hot
+        # encoding.
+        labels_list = tf.unstack(chars_labels, axis=1)
+
+      batch_size, seq_length, _ = chars_logits.shape.as_list()
+      if mparams.ignore_nulls:
+        weights = tf.ones((batch_size, seq_length), dtype=tf.float32)
+      else:
+        # Suppose that reject character is the last in the charset.
+        reject_char = tf.constant(
+            self._params.num_char_classes - 1,
+            shape=(batch_size, seq_length),
+            dtype=tf.int64)
+        known_char = tf.not_equal(chars_labels, reject_char)
+        weights = tf.to_float(known_char)
+
+      logits_list = tf.unstack(chars_logits, axis=1)
+      weights_list = tf.unstack(weights, axis=1)
+      loss = tf.contrib.legacy_seq2seq.sequence_loss(
+          logits_list,
+          labels_list,
+          weights_list,
+          softmax_loss_function=get_softmax_loss_fn(mparams.label_smoothing),
+          average_across_timesteps=mparams.average_across_timesteps)
+      tf.losses.add_loss(loss)
+      return loss
+
+  def create_summaries(self, data, endpoints, charset, is_training):
+    """Creates all summaries for the model.
+
+    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.
+      is_training: If True will create summary prefixes for training job,
+        otherwise - for evaluation.
+
+    Returns:
+      A list of evaluation ops
+    """
+
+    def sname(label):
+      prefix = 'train' if is_training else 'eval'
+      return '%s/%s' % (prefix, label)
+
+    max_outputs = 4
+    # TODO(gorban): uncomment, when tf.summary.text released.
+    # charset_mapper = CharsetMapper(charset)
+    # pr_text = charset_mapper.get_text(
+    #     endpoints.predicted_chars[:max_outputs,:])
+    # tf.summary.text(sname('text/pr'), pr_text)
+    # gt_text = charset_mapper.get_text(data.labels[:max_outputs,:])
+    # tf.summary.text(sname('text/gt'), gt_text)
+    tf.summary.image(sname('image'), data.images, max_outputs=max_outputs)
+
+    if is_training:
+      tf.summary.image(
+          sname('image/orig'), data.images_orig, max_outputs=max_outputs)
+      for var in tf.trainable_variables():
+        tf.summary.histogram(var.op.name, var)
+      return None
+
+    else:
+      names_to_values = {}
+      names_to_updates = {}
+
+      def use_metric(name, value_update_tuple):
+        names_to_values[name] = value_update_tuple[0]
+        names_to_updates[name] = value_update_tuple[1]
+
+      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',
+                 metrics.sequence_accuracy(
+                     endpoints.predicted_chars,
+                     data.labels,
+                     streaming=True,
+                     rej_char=self._params.null_code))
+
+      for name, value in names_to_values.iteritems():
+        summary_name = 'eval/' + name
+        tf.summary.scalar(summary_name, tf.Print(value, [value], summary_name))
+      return names_to_updates.values()
+
+  def create_init_fn_to_restore(self, master_checkpoint, inception_checkpoint):
+    """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.
+      inception_checkpoint: path to a checkpoint which contains weights for the
+        inception part only.
+
+    Returns:
+      a function to run initialization ops.
+    """
+    all_assign_ops = []
+    all_feed_dict = {}
+
+    def assign_from_checkpoint(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)
+      assign_op, feed_dict = slim.assign_from_checkpoint(checkpoint, variables)
+      all_assign_ops.append(assign_op)
+      all_feed_dict.update(feed_dict)
+
+    if master_checkpoint:
+      assign_from_checkpoint(utils.variables_to_restore(), master_checkpoint)
+
+    if inception_checkpoint:
+      variables = utils.variables_to_restore(
+          'AttentionOcr_v1/conv_tower_fn/INCE', strip_scope=True)
+      assign_from_checkpoint(variables, inception_checkpoint)
+
+    def init_assign_fn(sess):
+      logging.info('Restoring checkpoint(s)')
+      sess.run(all_assign_ops, all_feed_dict)
+
+    return init_assign_fn