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"""Implementation of Cluster Resolvers for TF_CONFIG Environment Variables.""" 16 17 18import json 19import os 20 21from tensorflow.python.distribute.cluster_resolver.cluster_resolver import ClusterResolver 22from tensorflow.python.training.server_lib import ClusterSpec 23from tensorflow.python.util.tf_export import tf_export 24 25_TF_CONFIG_ENV = 'TF_CONFIG' 26_SESSION_MASTER_KEY = 'session_master' 27_RPC_LAYER_KEY = 'rpc_layer' 28_TASK_KEY = 'task' 29 30 31def format_master_url(master, rpc_layer=None): 32 if rpc_layer: 33 return '%s://%s' % (rpc_layer, master) 34 else: 35 return master 36 37 38def _load_tf_config(): 39 return json.loads(os.environ.get(_TF_CONFIG_ENV, '{}')) 40 41 42def _get_value_in_tfconfig(key, default=None): 43 tf_config = _load_tf_config() 44 return tf_config[key] if key in tf_config else default 45 46 47@tf_export('distribute.cluster_resolver.TFConfigClusterResolver') 48class TFConfigClusterResolver(ClusterResolver): 49 """Implementation of a ClusterResolver which reads the TF_CONFIG EnvVar. 50 51 This is an implementation of cluster resolvers when using TF_CONFIG to set 52 information about the cluster. The cluster spec returned will be 53 initialized from the TF_CONFIG environment variable. 54 55 An example to set TF_CONFIG is: 56 57 ```Python 58 os.environ['TF_CONFIG'] = json.dumps({ 59 'cluster': { 60 'worker': ["localhost:12345", "localhost:23456"] 61 }, 62 'task': {'type': 'worker', 'index': 0} 63 }) 64 ``` 65 66 However, sometimes the container orchestration framework will set TF_CONFIG 67 for you. In this case, you can just create an instance without passing in any 68 arguments. You can find an example here to let Kuburnetes set TF_CONFIG for 69 you: https://github.com/tensorflow/ecosystem/tree/master/kubernetes. Then you 70 can use it with `tf.distribute.Strategy` as: 71 72 ```Python 73 # `TFConfigClusterResolver` is already the default one in the following 74 # strategy. 75 strategy = tf.distribute.experimental.MultiWorkerMirroredStrategy( 76 cluster_resolver=TFConfigClusterResolver()) 77 ``` 78 """ 79 80 def __init__(self, 81 task_type=None, 82 task_id=None, 83 rpc_layer=None, 84 environment=None): 85 """Creates a new TFConfigClusterResolver. 86 87 Args: 88 task_type: (String, optional) Overrides the task type specified in the 89 TF_CONFIG environment variable. 90 task_id: (Integer, optional) Overrides the task index specified in the 91 TF_CONFIG environment variable. 92 rpc_layer: (String, optional) Overrides the rpc layer TensorFlow uses. 93 environment: (String, optional) Overrides the environment TensorFlow 94 operates in. 95 """ 96 self._task_type = task_type 97 self._task_id = task_id 98 self._rpc_layer = rpc_layer 99 self._environment = environment 100 101 @property 102 def task_type(self): 103 if self._task_type is None: 104 task_info = _get_value_in_tfconfig(_TASK_KEY, {}) 105 return str(task_info['type']) if 'type' in task_info else None 106 else: 107 return str(self._task_type) 108 109 @property 110 def task_id(self): 111 if self._task_id is None: 112 task_info = _get_value_in_tfconfig(_TASK_KEY, {}) 113 return int(task_info['index']) if 'index' in task_info else None 114 else: 115 return int(self._task_id) 116 117 @task_type.setter 118 def task_type(self, task_type): 119 self._task_type = task_type 120 121 @task_id.setter 122 def task_id(self, task_id): 123 self._task_id = task_id 124 125 @property 126 def environment(self): 127 return self._environment 128 129 @property 130 def rpc_layer(self): 131 if self._rpc_layer is None: 132 return _get_value_in_tfconfig(_RPC_LAYER_KEY) 133 else: 134 return self._rpc_layer 135 136 @rpc_layer.setter 137 def rpc_layer(self, rpc_layer): 138 self._rpc_layer = rpc_layer 139 140 def num_accelerators(self, 141 task_type=None, 142 task_id=None, 143 config_proto=None): 144 task_type = self.task_type if task_type is None else task_type 145 task_id = self.task_id if task_id is None else task_id 146 return super(TFConfigClusterResolver, self).num_accelerators( 147 task_type, task_id, config_proto) 148 149 def cluster_spec(self): 150 """Returns a ClusterSpec based on the TF_CONFIG environment variable. 151 152 Returns: 153 A ClusterSpec with information from the TF_CONFIG environment variable. 154 """ 155 tf_config = _load_tf_config() 156 if 'cluster' not in tf_config: 157 return ClusterSpec({}) 158 return ClusterSpec(tf_config['cluster']) 159 160 def master(self, task_type=None, task_id=None, rpc_layer=None): 161 """Returns the master address to use when creating a TensorFlow session. 162 163 Note: this is only useful for TensorFlow 1.x. 164 165 Args: 166 task_type: (String, optional) Overrides and sets the task_type of the 167 master. 168 task_id: (Integer, optional) Overrides and sets the task id of the 169 master. 170 rpc_layer: (String, optional) Overrides and sets the protocol over which 171 TensorFlow nodes communicate with each other. 172 173 Returns: 174 The address of the master. 175 176 Raises: 177 RuntimeError: If the task_type or task_id is not specified and the 178 `TF_CONFIG` environment variable does not contain a task section. 179 """ 180 181 # If `session_master` is set, just use that. 182 session_master = _get_value_in_tfconfig(_SESSION_MASTER_KEY) 183 if session_master is not None: 184 return session_master 185 186 # Return an empty string if we are the only job in the ClusterSpec. 187 cluster_spec = self.cluster_spec() 188 if (not cluster_spec.jobs or 189 (len(cluster_spec.jobs) == 1 and 190 len(cluster_spec.job_tasks(cluster_spec.jobs[0])) == 1)): 191 return '' 192 193 # We try to auto-detect the task type and id, but uses the user-supplied one 194 # where available 195 task_type = task_type if task_type is not None else self.task_type 196 task_id = task_id if task_id is not None else self.task_id 197 rpc_layer = rpc_layer if rpc_layer is not None else self.rpc_layer 198 199 return format_master_url(cluster_spec.task_address(task_type, task_id), 200 rpc_layer) 201