1# Copyright 2018 The TensorFlow Authors. All Rights Reserved. 2# 3# Licensed under the Apache License, Version 2.0 (the "License"); 4# you may not use this file except in compliance with the License. 5# You may obtain a copy of the License at 6# 7# http://www.apache.org/licenses/LICENSE-2.0 8# 9# Unless required by applicable law or agreed to in writing, software 10# distributed under the License is distributed on an "AS IS" BASIS, 11# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12# See the License for the specific language governing permissions and 13# limitations under the License. 14# ============================================================================== 15"""Utilities for API compatibility between TensorFlow release versions. 16 17See [Version 18Compatibility](https://tensorflow.org/guide/version_compat#backward_forward) 19""" 20 21from __future__ import absolute_import 22from __future__ import division 23from __future__ import print_function 24 25import datetime 26 27from tensorflow.python.util import tf_contextlib 28from tensorflow.python.util.tf_export import tf_export 29 30_FORWARD_COMPATIBILITY_HORIZON = datetime.date(2019, 3, 15) 31 32 33@tf_export("compat.forward_compatible") 34def forward_compatible(year, month, day): 35 """Return true if the forward compatibility window has expired. 36 37 See [Version 38 compatibility](https://tensorflow.org/guide/version_compat#backward_forward). 39 40 Forward-compatibility refers to scenarios where the producer of a TensorFlow 41 model (a GraphDef or SavedModel) is compiled against a version of the 42 TensorFlow library newer than what the consumer was compiled against. The 43 "producer" is typically a Python program that constructs and trains a model 44 while the "consumer" is typically another program that loads and serves the 45 model. 46 47 TensorFlow has been supporting a 3 week forward-compatibility window for 48 programs compiled from source at HEAD. 49 50 For example, consider the case where a new operation `MyNewAwesomeAdd` is 51 created with the intent of replacing the implementation of an existing Python 52 wrapper - `tf.add`. The Python wrapper implementation should change from 53 something like: 54 55 ```python 56 def add(inputs, name=None): 57 return gen_math_ops.add(inputs, name) 58 ``` 59 60 to: 61 62 ```python 63 from tensorflow.python.compat import compat 64 65 def add(inputs, name=None): 66 if compat.forward_compatible(year, month, day): 67 # Can use the awesome new implementation. 68 return gen_math_ops.my_new_awesome_add(inputs, name) 69 # To maintain forward compatibiltiy, use the old implementation. 70 return gen_math_ops.add(inputs, name) 71 ``` 72 73 Where `year`, `month`, and `day` specify the date beyond which binaries 74 that consume a model are expected to have been updated to include the 75 new operations. This date is typically at least 3 weeks beyond the date 76 the code that adds the new operation is committed. 77 78 Args: 79 year: A year (e.g., 2018). 80 month: A month (1 <= month <= 12) in year. 81 day: A day (1 <= day <= 31, or 30, or 29, or 28) in month. 82 83 Returns: 84 True if the caller can expect that serialized TensorFlow graphs produced 85 can be consumed by programs that are compiled with the TensorFlow library 86 source code after (year, month, day). 87 """ 88 return _FORWARD_COMPATIBILITY_HORIZON > datetime.date(year, month, day) 89 90 91@tf_export("compat.forward_compatibility_horizon") 92@tf_contextlib.contextmanager 93def forward_compatibility_horizon(year, month, day): 94 """Context manager for testing forward compatibility of generated graphs. 95 96 See [Version 97 compatibility](https://tensorflow.org/guide/version_compat#backward_forward). 98 99 To ensure forward compatibility of generated graphs (see `forward_compatible`) 100 with older binaries, new features can be gated with: 101 102 ```python 103 if compat.forward_compatible(year=2018, month=08, date=01): 104 generate_graph_with_new_features() 105 else: 106 generate_graph_so_older_binaries_can_consume_it() 107 ``` 108 109 However, when adding new features, one may want to unittest it before 110 the forward compatibility window expires. This context manager enables 111 such tests. For example: 112 113 ```python 114 from tensorflow.python.compat import compat 115 116 def testMyNewFeature(self): 117 with compat.forward_compatibility_horizon(2018, 08, 02): 118 # Test that generate_graph_with_new_features() has an effect 119 ``` 120 121 Args : 122 year: A year (e.g. 2018). 123 month: A month (1 <= month <= 12) in year. 124 day: A day (1 <= day <= 31, or 30, or 29, or 28) in month. 125 126 Yields: 127 Nothing. 128 """ 129 global _FORWARD_COMPATIBILITY_HORIZON 130 try: 131 old_compat_date = _FORWARD_COMPATIBILITY_HORIZON 132 _FORWARD_COMPATIBILITY_HORIZON = datetime.date(year, month, day) 133 yield 134 finally: 135 _FORWARD_COMPATIBILITY_HORIZON = old_compat_date 136