Add Apache Arrow dataset support

BUILD

@@ -5,7 +5,8 @@ sh_binary(
         "LICENSE",
         "MANIFEST.in",
         "setup.py",
-	"tensorflow_io/__init__.py",
+        "tensorflow_io/__init__.py",
+        "//tensorflow_io/arrow:arrow_py",
         "//tensorflow_io/hadoop:hadoop_py",
         "//tensorflow_io/ignite:ignite_py",
         "//tensorflow_io/kafka:kafka_py",

README.md

@@ -5,11 +5,12 @@
 TensorFlow I/O is a collection of file systems and file formats that are not
 available in TensorFlow's built-in support.
 
-At the moment TensorFlow I/O supports 4 data sources:
+At the moment TensorFlow I/O supports 5 data sources:
 - `tensorflow_io.ignite`: Data source for Apache Ignite and Ignite File System (IGFS).
 - `tensorflow_io.kafka`: Apache Kafka stream-processing support.
 - `tensorflow_io.kinesis`: Amazon Kinesis data streams support.
 - `tensorflow_io.hadoop`: Hadoop SequenceFile format support.
+- `tensorflow_io.arrow`: Apache Arrow data format support. Usage guide [here](tensorflow_io/arrow/README.md).
 
 ## Installation
 

tensorflow_io/arrow/BUILD

@@ -0,0 +1,56 @@
+licenses(["notice"])  # Apache 2.0
+
+package(default_visibility = ["//visibility:public"])
+
+cc_binary(
+    name = 'python/ops/_arrow_ops.so',
+    srcs = [
+        "kernels/arrow_dataset_ops.cc",
+        "kernels/arrow_stream_client.h",
+        "kernels/arrow_stream_client_unix.cc",
+        "ops/dataset_ops.cc",
+    ],
+    linkshared = 1,
+    deps = [
+        "@local_config_tf//:libtensorflow_framework",
+        "@local_config_tf//:tf_header_lib",
+        "@arrow//:arrow",
+    ],
+    copts = ["-pthread", "-std=c++11", "-D_GLIBCXX_USE_CXX11_ABI=0", "-DNDEBUG"]
+)
+
+py_library(
+    name = "arrow_ops_py",
+    srcs = [
+        "python/ops/arrow_dataset_ops.py",
+    ],
+    data = [
+        ":python/ops/_arrow_ops.so"
+    ],
+    srcs_version = "PY2AND3",
+)
+
+py_test(
+    name = "arrow_py_test",
+    srcs = [
+        "python/kernel_tests/arrow_test.py"
+    ],
+    main = "python/kernel_tests/arrow_test.py",
+    deps = [
+        ":arrow_ops_py"
+    ],
+    srcs_version = "PY2AND3",
+)
+
+py_library(
+    name = "arrow_py",
+    srcs = [
+        "__init__.py",
+        "python/__init__.py",
+        "python/ops/__init__.py",
+    ],
+    deps = [
+        ":arrow_ops_py",
+    ],
+    srcs_version = "PY2AND3",
+)

tensorflow_io/arrow/README.md

@@ -0,0 +1,111 @@
+# TensorFlow I/O Apache Arrow Datasets
+
+Apache Arrow is a standard for in-memory columnar data, see [here](https://arrow.apache.org)
+for more information on the project. An Arrow dataset makes it easy to bring in
+column-oriented data from other systems to TensorFlow using the following
+sources:
+
+## From a Pandas DataFrame
+
+An `ArrowDataset` can be made directly from an existing Pandas DataFrame, or
+pyarrow record batches, in a Python process. Tensor types and shapes can be
+inferred from the DataFrame, although currently only scalar and vector values
+with primitive types are supported. PyArrow must be installed to use this
+Dataset. Example usage:
+
+```python
+import tensorflow as tf
+from tensorflow_io.arrow import ArrowDataset
+
+# Assume `df` is an existing Pandas DataFrame
+dataset = ArrowDataset.from_pandas(df)
+
+iterator = dataset.make_one_shot_iterator()
+next_element = iterator.get_next()
+
+with tf.Session() as sess:
+  for i in range(len(df)):
+    print(sess.run(next_element))
+```
+
+NOTE: The entire DataFrame will be serialized to the Dataset and is not
+recommended for use with large amounts of data
+
+## From Arrow Feather Files
+
+Feather is a light-weight file format that provides a simple and efficient way
+to write Pandas DataFrames to disk, see [here](https://arrow.apache.org/docs/python/ipc.html#feather-format)
+for more information and limitations of the format. An `ArrowFeatherDataset`
+can be created to read one or more Feather files from the given pathnames. The
+following example shows how to write a feather file from a Pandas DataFrame,
+then read multiple files back as an `ArrowFeatherDataset`:
+
+```python
+from pyarrow.feather import write_feather
+
+# Assume `df` is an existing Pandas DataFrame with dtypes=(int32, float32)
+write_feather(df, '/path/to/a.feather')
+```
+
+```python
+import tensorflow as tf
+from tensorflow_io.arrow import ArrowFeatherDataset
+
+# Each Feather file must have the same column types, here we use the above
+# DataFrame which has 2 columns with dtypes=(int32, float32)
+dataset = ArrowFeatherDataset(
+    ['/path/to/a.feather', '/path/to/b.feather'],
+    columns=(0, 1),
+    output_types=(tf.int32, tf.float32),
+    output_shapes=([], []))
+
+iterator = dataset.make_one_shot_iterator()
+next_element = iterator.get_next()
+
+# This will iterate over each row of each file provided
+with tf.Session() as sess:
+  while True:
+    try:
+      print(sess.run(next_element))
+    except tf.errors.OutOfRangeError:
+      break
+```
+
+An alternate constructor can also be used to infer output types and shapes from
+a given `pyarrow.Schema`, e.g. `dataset = ArrowFeatherDataset.from_schema(filenames, schema)`
+
+## From a Stream of Arrow Record Batches
+
+The `ArrowStreamDataset` provides a Dataset that will connect to a host over
+a socket that is serving Arrow record batches in the Arrow stream format. See
+[here](https://arrow.apache.org/docs/python/ipc.html#writing-and-reading-streams)
+for more on the stream format. The following example will create an
+`ArrowStreamDataset` that will connect to a host that is serving an Arrow
+stream of record batches with 2 columns of dtypes=(int32, float32):
+
+```python
+import tensorflow as tf
+from tensorflow_io.arrow import ArrowStreamDataset
+
+# The str `host` should be in the format '<HOSTNAME>:<PORT>'
+dataset = ArrowStreamDataset(
+    host,
+    columns=(0, 1),
+    output_types=(tf.int32, tf.float32),
+    output_shapes=([], []))
+
+iterator = dataset.make_one_shot_iterator()
+next_element = iterator.get_next()
+
+# The host connection is made when the Dataset op is run and will iterate over
+# each row of each record batch until the Arrow stream is finished
+with tf.Session() as sess:
+  while True:
+    try:
+      print(sess.run(next_element))
+    except tf.errors.OutOfRangeError:
+      break
+```
+
+An alternate constructor can also be used to infer output types and shapes from
+a given `pyarrow.Schema`, e.g. `dataset = ArrowStreamDataset.from_schema(host, schema)`

tensorflow_io/arrow/__init__.py

@@ -0,0 +1,36 @@
+# Copyright 2016 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.
+# ==============================================================================
+"""Arrow Dataset.
+
+@@ArrowDataset
+"""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from tensorflow_io.arrow.python.ops.arrow_dataset_ops import ArrowDataset
+from tensorflow_io.arrow.python.ops.arrow_dataset_ops import ArrowFeatherDataset
+from tensorflow_io.arrow.python.ops.arrow_dataset_ops import ArrowStreamDataset
+
+from tensorflow.python.util.all_util import remove_undocumented
+
+_allowed_symbols = [
+    "ArrowDataset",
+    "ArrowFeatherDataset",
+    "ArrowStreamDataset",
+]
+
+remove_undocumented(__name__)