def _init_summary_op(self, summary_op=USE_DEFAULT): """Initilizes summary_op. Args: summary_op: An Operation that returns a Summary for the event logs. If set to USE_DEFAULT, create an op that merges all the summaries. """ if summary_op is Supervisor.USE_DEFAULT: summary_op = self._get_first_op_from_collection(ops.GraphKeys.SUMMARY_OP) if summary_op is None: summary_op = logging_ops.merge_all_summaries() if summary_op is not None: ops.add_to_collection(ops.GraphKeys.SUMMARY_OP, summary_op) self._summary_op = summary_op
def evaluation_loop(master, checkpoint_dir, logdir, num_evals=1, initial_op=None, initial_op_feed_dict=None, eval_op=None, eval_op_feed_dict=None, final_op=None, final_op_feed_dict=None, summary_op=_USE_DEFAULT, summary_op_feed_dict=None, variables_to_restore=None, eval_interval_secs=60, max_number_of_evaluations=None, session_config=None, timeout=None): """Runs TF-Slim's Evaluation Loop. Args: master: The BNS address of the TensorFlow master. checkpoint_dir: The directory where checkpoints are stored. logdir: The directory where the TensorFlow summaries are written to. num_evals: The number of times to run `eval_op`. initial_op: An operation run at the beginning of evaluation. initial_op_feed_dict: A feed dictionary to use when executing `initial_op`. eval_op: A operation run `num_evals` times. eval_op_feed_dict: The feed dictionary to use when executing the `eval_op`. final_op: An operation to execute after all of the `eval_op` executions. The value of `final_op` is returned. final_op_feed_dict: A feed dictionary to use when executing `final_op`. summary_op: The summary_op to evaluate after running TF-Slims metric ops. By default the summary_op is set to tf.merge_all_summaries(). summary_op_feed_dict: An optional feed dictionary to use when running the `summary_op`. variables_to_restore: A list of TensorFlow variables to restore during evaluation. If the argument is left as `None` then slim.variables.GetVariablesToRestore() is used. eval_interval_secs: The minimum number of seconds between evaluations. max_number_of_evaluations: the max number of iterations of the evaluation. If the value is left as 'None', the evaluation continues indefinitely. session_config: An instance of `tf.ConfigProto` that will be used to configure the `Session`. If left as `None`, the default will be used. timeout: The maximum amount of time to wait between checkpoints. If left as `None`, then the process will wait indefinitely. Returns: The value of `final_op` or `None` if `final_op` is `None`. """ if summary_op == _USE_DEFAULT: summary_op = logging_ops.merge_all_summaries() global_step = variables.get_or_create_global_step() saver = tf_saver.Saver(variables_to_restore or variables.get_variables_to_restore()) summary_writer = summary_io.SummaryWriter(logdir) sv = supervisor.Supervisor(graph=ops.get_default_graph(), logdir=logdir, summary_op=None, summary_writer=None, global_step=None, saver=saver) number_of_evaluations = 0 for checkpoint_path in checkpoints_iterator(checkpoint_dir, eval_interval_secs, timeout): logging.info('Starting evaluation at ' + time.strftime('%Y-%m-%d-%H:%M:%S', time.gmtime())) with sv.managed_session( master, start_standard_services=False, config=session_config) as sess: sv.saver.restore(sess, checkpoint_path) sv.start_queue_runners(sess) final_op_value = evaluation(sess, num_evals=num_evals, initial_op=initial_op, initial_op_feed_dict=initial_op_feed_dict, eval_op=eval_op, eval_op_feed_dict=eval_op_feed_dict, final_op=final_op, final_op_feed_dict=final_op_feed_dict, summary_op=summary_op, summary_op_feed_dict=summary_op_feed_dict, summary_writer=summary_writer, global_step=global_step) logging.info('Finished evaluation at ' + time.strftime('%Y-%m-%d-%H:%M:%S', time.gmtime())) number_of_evaluations += 1 if (max_number_of_evaluations and number_of_evaluations >= max_number_of_evaluations): logging.info('Reached max_number_of_evaluations=%s. Exit', max_number_of_evaluations) return final_op_value logging.info( 'Timed-out waiting for new checkpoint file. Exiting evaluation loop.') return final_op_value
def evaluate_once(master, checkpoint_path, logdir, num_evals=1, initial_op=None, initial_op_feed_dict=None, eval_op=None, eval_op_feed_dict=None, final_op=None, final_op_feed_dict=None, summary_op=_USE_DEFAULT, summary_op_feed_dict=None, variables_to_restore=None, session_config=None): """Evaluates the model at the given checkpoint path. Args: master: The BNS address of the TensorFlow master. checkpoint_path: The path to a checkpoint to use for evaluation. logdir: The directory where the TensorFlow summaries are written to. num_evals: The number of times to run `eval_op`. initial_op: An operation run at the beginning of evaluation. initial_op_feed_dict: A feed dictionary to use when executing `initial_op`. eval_op: A operation run `num_evals` times. eval_op_feed_dict: The feed dictionary to use when executing the `eval_op`. final_op: An operation to execute after all of the `eval_op` executions. The value of `final_op` is returned. final_op_feed_dict: A feed dictionary to use when executing `final_op`. summary_op: The summary_op to evaluate after running TF-Slims metric ops. By default the summary_op is set to tf.merge_all_summaries(). summary_op_feed_dict: An optional feed dictionary to use when running the `summary_op`. variables_to_restore: A list of TensorFlow variables to restore during evaluation. If the argument is left as `None` then slim.variables.GetVariablesToRestore() is used. session_config: An instance of `tf.ConfigProto` that will be used to configure the `Session`. If left as `None`, the default will be used. Returns: The value of `final_op` or `None` if `final_op` is `None`. """ if summary_op == _USE_DEFAULT: summary_op = logging_ops.merge_all_summaries() global_step = variables.get_or_create_global_step() saver = tf_saver.Saver( variables_to_restore or variables.get_variables_to_restore(), write_version=saver_pb2.SaverDef.V1) summary_writer = summary_io.SummaryWriter(logdir) sv = supervisor.Supervisor(graph=ops.get_default_graph(), logdir=logdir, summary_op=None, summary_writer=None, global_step=None, saver=None) logging.info('Starting evaluation at ' + time.strftime('%Y-%m-%d-%H:%M:%S', time.gmtime())) with sv.managed_session( master, start_standard_services=False, config=session_config) as sess: saver.restore(sess, checkpoint_path) sv.start_queue_runners(sess) final_op_value = evaluation(sess, num_evals=num_evals, initial_op=initial_op, initial_op_feed_dict=initial_op_feed_dict, eval_op=eval_op, eval_op_feed_dict=eval_op_feed_dict, final_op=final_op, final_op_feed_dict=final_op_feed_dict, summary_op=summary_op, summary_op_feed_dict=summary_op_feed_dict, summary_writer=summary_writer, global_step=global_step) logging.info('Finished evaluation at ' + time.strftime('%Y-%m-%d-%H:%M:%S', time.gmtime())) return final_op_value
def _setup_training(self): """Sets up graph, model and trainer.""" # Create config if not given. if self._config is None: self._config = RunConfig(verbose=self.verbose) # Create new graph. self._graph = ops.Graph() self._graph.add_to_collection("IS_TRAINING", True) with self._graph.as_default(): random_seed.set_random_seed(self._config.tf_random_seed) self._global_step = variables.Variable(0, name="global_step", trainable=False) # Setting up inputs and outputs. self._inp, self._out = self._data_feeder.input_builder() # If class weights are provided, add them to the graph. # Different loss functions can use this tensor by name. if self.class_weight: self._class_weight_node = constant_op.constant( self.class_weight, name='class_weight') # Add histograms for X and y if they are floats. if self._data_feeder.input_dtype in (np.float32, np.float64): logging_ops.histogram_summary("X", self._inp) if self._data_feeder.output_dtype in (np.float32, np.float64): logging_ops.histogram_summary("y", self._out) # Create model's graph. self._model_predictions, self._model_loss = self.model_fn( self._inp, self._out) # Set up a single operator to merge all the summaries self._summaries = logging_ops.merge_all_summaries() # Create trainer and augment graph with gradients and optimizer. # Additionally creates initialization ops. learning_rate = self.learning_rate optimizer = self.optimizer if callable(learning_rate): learning_rate = learning_rate(self._global_step) if callable(optimizer): optimizer = optimizer(learning_rate) self._train = optimizers.optimize_loss( self._model_loss, self._global_step, learning_rate=learning_rate, optimizer=optimizer, clip_gradients=self.clip_gradients) # Update ops during training, e.g. batch_norm_ops self._train = control_flow_ops.group( self._train, *ops.get_collection('update_ops')) # Get all initializers for all trainable variables. self._initializers = variables.initialize_all_variables() # Create model's saver capturing all the nodes created up until now. self._saver = train.Saver( max_to_keep=self._config.keep_checkpoint_max, keep_checkpoint_every_n_hours=self._config. keep_checkpoint_every_n_hours) # Enable monitor to create validation data dict with appropriate tf placeholders self._monitor.create_val_feed_dict(self._inp, self._out) # Create session to run model with. self._session = session.Session(self._config.tf_master, config=self._config.tf_config) # Run parameter initializers. self._session.run(self._initializers)
def _setup_training(self): """Sets up graph, model and trainer.""" # Create config if not given. if self._config is None: self._config = RunConfig(verbose=self.verbose) # Create new graph. self._graph = ops.Graph() self._graph.add_to_collection("IS_TRAINING", True) with self._graph.as_default(): random_seed.set_random_seed(self._config.tf_random_seed) self._global_step = variables.Variable( 0, name="global_step", trainable=False) # Setting up inputs and outputs. self._inp, self._out = self._data_feeder.input_builder() # If class weights are provided, add them to the graph. # Different loss functions can use this tensor by name. if self.class_weight: self._class_weight_node = constant_op.constant( self.class_weight, name='class_weight') # Add histograms for X and y if they are floats. if self._data_feeder.input_dtype in (np.float32, np.float64): logging_ops.histogram_summary("X", self._inp) if self._data_feeder.output_dtype in (np.float32, np.float64): logging_ops.histogram_summary("y", self._out) # Create model's graph. self._model_predictions, self._model_loss = self.model_fn( self._inp, self._out) # Create trainer and augment graph with gradients and optimizer. # Additionally creates initialization ops. learning_rate = self.learning_rate optimizer = self.optimizer if callable(learning_rate): learning_rate = learning_rate(self._global_step) if callable(optimizer): optimizer = optimizer(learning_rate) self._train = optimizers.optimize_loss(self._model_loss, self._global_step, learning_rate=learning_rate, optimizer=optimizer, clip_gradients=self.clip_gradients) # Update ops during training, e.g. batch_norm_ops self._train = control_flow_ops.group(self._train, *ops.get_collection('update_ops')) # Merge all summaries into single tensor. self._summaries = logging_ops.merge_all_summaries() # Get all initializers for all trainable variables. self._initializers = variables.initialize_all_variables() # Create model's saver capturing all the nodes created up until now. self._saver = train.Saver( max_to_keep=self._config.keep_checkpoint_max, keep_checkpoint_every_n_hours=self._config.keep_checkpoint_every_n_hours) # Enable monitor to create validation data dict with appropriate tf placeholders self._monitor.create_val_feed_dict(self._inp, self._out) # Create session to run model with. self._session = session.Session(self._config.tf_master, config=self._config.tf_config) # Run parameter initializers. self._session.run(self._initializers)
def evaluation_loop(master, checkpoint_dir, logdir, num_evals=1, eval_op=None, eval_op_feed_dict=None, final_op=None, final_op_feed_dict=None, summary_op=_USE_DEFAULT, summary_op_feed_dict=None, variables_to_restore=None, eval_interval_secs=60, max_number_of_evaluations=None): """Runs TF-Slim's Evaluation Loop. Args: master: The BNS address of the TensorFlow master. checkpoint_dir: The directory where checkpoints are stored. logdir: The directory where the TensorFlow summaries are written to. num_evals: The number of times to run `eval_op`. eval_op: A operation run `num_evals` times. eval_op_feed_dict: The feed dictionary to use when executing the `eval_op`. final_op: An operation to execute after all of the `eval_op` executions. The value of `final_op` is returned. final_op_feed_dict: A feed dictionary to use when executing `final_op`. summary_op: The summary_op to evaluate after running TF-Slims metric ops. By default the summary_op is set to tf.merge_all_summaries(). summary_op_feed_dict: An optional feed dictionary to use when running the `summary_op`. variables_to_restore: A list of TensorFlow variables to restore during evaluation. If the argument is left as `None` then slim.variables.GetVariablesToRestore() is used. eval_interval_secs: The minimum number of seconds between evaluations. max_number_of_evaluations: the max number of iterations of the evaluation. If the value is left as 'None', the evaluation continues indefinitely. """ if summary_op == _USE_DEFAULT: summary_op = logging_ops.merge_all_summaries() global_step = variables.get_or_create_global_step() init_op = control_flow_ops.group(tf_variables.initialize_all_variables(), tf_variables.initialize_local_variables(), data_flow_ops.initialize_all_tables()) saver = tf_saver.Saver(variables_to_restore or variables.get_variables_to_restore()) summary_writer = summary_io.SummaryWriter(logdir) sv = supervisor.Supervisor(graph=ops.get_default_graph(), logdir=logdir, init_op=init_op, summary_op=None, summary_writer=None, global_step=None, saver=saver) last_checkpoint = None number_of_evaluations = 0 while True: last_checkpoint = wait_for_new_checkpoint(checkpoint_dir, last_checkpoint) start = time.time() logging.info('Starting evaluation at ' + time.strftime('%Y-%m-%d-%H:%M:%S', time.gmtime())) with sv.managed_session(master, start_standard_services=False) as sess: sv.saver.restore(sess, last_checkpoint) sv.start_queue_runners(sess) evaluation(sess, num_evals=num_evals, eval_op=eval_op, eval_op_feed_dict=eval_op_feed_dict, final_op=final_op, final_op_feed_dict=final_op_feed_dict, summary_op=summary_op, summary_op_feed_dict=summary_op_feed_dict, summary_writer=summary_writer, global_step=global_step) logging.info('Finished evaluation at ' + time.strftime('%Y-%m-%d-%H:%M:%S', time.gmtime())) number_of_evaluations += 1 if (max_number_of_evaluations and number_of_evaluations >= max_number_of_evaluations): logging.info('Reached max_number_of_evaluations=%s. Exit', max_number_of_evaluations) break time_to_next_eval = start + eval_interval_secs - time.time() if time_to_next_eval > 0: time.sleep(time_to_next_eval)
def train( train_op, logdir, log_every_n_steps=1, graph=None, master='', is_chief=True, global_step=None, number_of_steps=None, init_op=_USE_DEFAULT, init_feed_dict=None, init_fn=None, summary_op=_USE_DEFAULT, save_summaries_secs=600, startup_delay_steps=0, saver=None, save_interval_secs=600, sync_optimizer=None): """Runs a training loop using a TensorFlow supervisor. When the sync_optimizer is supplied, gradient updates are applied synchronously. Otherwise, gradient updates are applied asynchronous. Args: train_op: A `Tensor` that, when executed, will apply the gradients and return the loss value. logdir: the directory where training logs are written to. log_every_n_steps: The frequency, in terms of global steps, that the loss and global step and logged. graph: The graph to pass to the supervisor. If no graph is supplied the default graph is used. master: The BNS name of the tensorflow master. is_chief: Specifies whether or not the training is being run by the primary replica during replica training. global_step: The `Tensor` representing the global step. If left as `None`, then slim.variables.get_or_create_global_step() is used. number_of_steps: The max number of gradient steps to take during training. If the value is left as None, training proceeds indefinitely. init_op: The initialization operation. init_feed_dict: A feed dictionary to use when executing the `init_op`. init_fn: An optional callable to be executed after `init_op` is called. The callable must accept one argument, the session being initialized. summary_op: The summary operation. save_summaries_secs: How often, in seconds, to save summaries. startup_delay_steps: The number of steps to wait for before beginning. Note that this must be 0 if a sync_optimizer is supplied. saver: Saver to save checkpoints. If none, a default one will be created and used. save_interval_secs: How often, in seconds, to save the model to `logdir`. sync_optimizer: an instance of tf.train.SyncReplicasOptimizer. If the argument is supplied, gradient updates will be synchronous. If left as `None`, gradient updates will be asynchronous. Returns: the value of the loss function after training. Raises: ValueError: if `train_op` is empty or if `startup_delay_steps` is non-zero when `sync_optimizer` is supplied, or if `number_of_steps` is negative. """ if train_op is None: raise ValueError('train_op cannot be None.') if sync_optimizer and startup_delay_steps > 0: raise ValueError( 'startup_delay_steps must be zero when sync_optimizer is supplied.') if number_of_steps is not None and number_of_steps <= 0: raise ValueError( '`number_of_steps` must be either None or a positive number.') graph = graph or ops.get_default_graph() if global_step is None: global_step = variables.get_or_create_global_step() saver = saver or tf_saver.Saver() if init_op is None: init_op = control_flow_ops.group( tf_variables.initialize_all_variables(), tf_variables.initialize_local_variables(), tf_variables.initialize_all_tables()) if summary_op == _USE_DEFAULT: summary_op = logging_ops.merge_all_summaries() local_init_op = None cleanup_op = None if is_chief and sync_optimizer: if not isinstance(sync_optimizer, sync_replicas_optimizer.SyncReplicasOptimizer): raise ValueError( '`sync_optimizer` must be a tf.train.SyncReplicasOptimizer') # Need to create these BEFORE the supervisor finalizes the graph: local_init_op = sync_optimizer.get_init_tokens_op() chief_queue_runner = sync_optimizer.get_chief_queue_runner() cleanup_op = sync_optimizer.get_clean_up_op() if number_of_steps: should_stop_op = math_ops.greater_equal(global_step, number_of_steps) else: should_stop_op = constant_op.constant(False) should_log_op = math_ops.equal( math_ops.mod(global_step, log_every_n_steps), 0) sv = supervisor.Supervisor( graph=graph, is_chief=is_chief, logdir=logdir, init_op=init_op, init_feed_dict=init_feed_dict, local_init_op=local_init_op, summary_op=summary_op, global_step=global_step, saver=saver, save_summaries_secs=save_summaries_secs, save_model_secs=save_interval_secs, init_fn=init_fn) with sv.managed_session(master, start_standard_services=False) as sess: if is_chief: sv.start_standard_services(sess) elif not is_chief and startup_delay_steps > 0: _wait_for_step(sess, global_step, min(startup_delay_steps, number_of_steps or sys.maxint)) sv.start_queue_runners(sess) if is_chief and sync_optimizer: sv.start_queue_runners(sess, [chief_queue_runner]) total_loss = train_loop( sv, sess, train_op, should_stop_op, should_log_op, global_step, cleanup_op) # This waits for service threads to finish. sv.Stop() if sv.is_chief: logging.info('Finished training! Saving model to disk.') sv.saver.save(sess, sv.save_path, global_step=sv.global_step) return total_loss
def train(train_op, logdir, train_step_fn=train_step, train_step_kwargs=_USE_DEFAULT, log_every_n_steps=1, graph=None, master='', is_chief=True, global_step=None, number_of_steps=None, init_op=_USE_DEFAULT, init_feed_dict=None, local_init_op=_USE_DEFAULT, init_fn=None, ready_op=_USE_DEFAULT, summary_op=_USE_DEFAULT, save_summaries_secs=600, summary_writer=_USE_DEFAULT, startup_delay_steps=0, saver=None, save_interval_secs=600, sync_optimizer=None, session_config=None, trace_every_n_steps=None): """Runs a training loop using a TensorFlow supervisor. When the sync_optimizer is supplied, gradient updates are applied synchronously. Otherwise, gradient updates are applied asynchronous. Args: train_op: A `Tensor` that, when executed, will apply the gradients and return the loss value. logdir: The directory where training logs are written to. If None, model checkpoints and summaries will not be written. train_step_fn: The function to call in order to execute a single gradient step. The function must have take exactly four arguments: the current session, the `train_op` `Tensor`, a global step `Tensor` and a dictionary. train_step_kwargs: A dictionary which is passed to the `train_step_fn`. By default, two `Boolean`, scalar ops called "should_stop" and "should_log" are provided. log_every_n_steps: The frequency, in terms of global steps, that the loss and global step and logged. graph: The graph to pass to the supervisor. If no graph is supplied the default graph is used. master: The BNS name of the tensorflow master. is_chief: Specifies whether or not the training is being run by the primary replica during replica training. global_step: The `Tensor` representing the global step. If left as `None`, then slim.variables.get_or_create_global_step() is used. number_of_steps: The max number of gradient steps to take during training. If the value is left as None, training proceeds indefinitely. init_op: The initialization operation. If left to its default value, then the session is initialized by calling `tf.initialize_all_variables()`. init_feed_dict: A feed dictionary to use when executing the `init_op`. local_init_op: The local initialization operation. If left to its default value, then the session is initialized by calling `tf.initialize_local_variables()` and `tf.initialize_all_tables()`. init_fn: An optional callable to be executed after `init_op` is called. The callable must accept one argument, the session being initialized. ready_op: Operation to check if the model is ready to use. If left to its default value, then the session checks for readiness by calling `tf.report_uninitialized_variables()`. summary_op: The summary operation. save_summaries_secs: How often, in seconds, to save summaries. summary_writer: `SummaryWriter` to use. Can be `None` to indicate that no summaries should be written. If unset, we create a SummaryWriter. startup_delay_steps: The number of steps to wait for before beginning. Note that this must be 0 if a sync_optimizer is supplied. saver: Saver to save checkpoints. If None, a default one will be created and used. save_interval_secs: How often, in seconds, to save the model to `logdir`. sync_optimizer: an instance of tf.train.SyncReplicasOptimizer. If the argument is supplied, gradient updates will be synchronous. If left as `None`, gradient updates will be asynchronous. session_config: An instance of `tf.ConfigProto` that will be used to configure the `Session`. If left as `None`, the default will be used. trace_every_n_steps: produce and save a `Timeline` in Chrome trace format and add it to the summaries every `trace_every_n_steps`. If None, no trace information will be produced or saved. Returns: the value of the loss function after training. Raises: ValueError: if `train_op` is empty or if `startup_delay_steps` is non-zero when `sync_optimizer` is supplied, if `number_of_steps` is negative, or if `trace_every_n_steps` is not `None` and no `logdir` is provided. """ if train_op is None: raise ValueError('train_op cannot be None.') if logdir is None: if summary_op != _USE_DEFAULT: raise ValueError('Cannot provide summary_op because logdir=None') if saver is not None: raise ValueError('Cannot provide saver because logdir=None') if trace_every_n_steps is not None: raise ValueError('Cannot provide trace_every_n_steps because ' 'logdir=None') if sync_optimizer is not None and startup_delay_steps > 0: raise ValueError( 'startup_delay_steps must be zero when sync_optimizer is supplied.' ) if number_of_steps is not None and number_of_steps <= 0: raise ValueError( '`number_of_steps` must be either None or a positive number.') graph = graph or ops.get_default_graph() with graph.as_default(): if global_step is None: global_step = variables.get_or_create_global_step() saver = saver or tf_saver.Saver() with ops.name_scope('init_ops'): if init_op == _USE_DEFAULT: init_op = tf_variables.initialize_all_variables() if ready_op == _USE_DEFAULT: ready_op = tf_variables.report_uninitialized_variables() if local_init_op == _USE_DEFAULT: local_init_op = control_flow_ops.group( tf_variables.initialize_local_variables(), data_flow_ops.initialize_all_tables()) if summary_op == _USE_DEFAULT: summary_op = logging_ops.merge_all_summaries() if summary_writer == _USE_DEFAULT: summary_writer = supervisor.Supervisor.USE_DEFAULT cleanup_op = None if is_chief and sync_optimizer is not None: if not isinstance( sync_optimizer, (sync_replicas_optimizer.SyncReplicasOptimizer, sync_replicas_optimizer.SyncReplicasOptimizerV2)): raise ValueError( '`sync_optimizer` must be a tf.train.SyncReplicasOptimizer or ' 'tf.train.SyncReplicasOptimizerV2.') # Need to create these BEFORE the supervisor finalizes the graph: with ops.control_dependencies([init_op]): init_tokens_op = sync_optimizer.get_init_tokens_op() init_op = init_tokens_op chief_queue_runner = sync_optimizer.get_chief_queue_runner() if isinstance(sync_optimizer, sync_replicas_optimizer.SyncReplicasOptimizer): cleanup_op = sync_optimizer.get_clean_up_op() if train_step_kwargs == _USE_DEFAULT: with ops.name_scope('train_step'): train_step_kwargs = {} if number_of_steps: should_stop_op = math_ops.greater_equal( global_step, number_of_steps) else: should_stop_op = constant_op.constant(False) train_step_kwargs['should_stop'] = should_stop_op train_step_kwargs['should_log'] = math_ops.equal( math_ops.mod(global_step, log_every_n_steps), 0) if is_chief and trace_every_n_steps is not None: train_step_kwargs['should_trace'] = math_ops.equal( math_ops.mod(global_step, trace_every_n_steps), 0) train_step_kwargs['logdir'] = logdir sv = supervisor.Supervisor(graph=graph, is_chief=is_chief, logdir=logdir, init_op=init_op, init_feed_dict=init_feed_dict, local_init_op=local_init_op, ready_op=ready_op, summary_op=summary_op, summary_writer=summary_writer, global_step=global_step, saver=saver, save_summaries_secs=save_summaries_secs, save_model_secs=save_interval_secs, init_fn=init_fn) if summary_writer is not None: train_step_kwargs['summary_writer'] = sv.summary_writer should_retry = True while should_retry: try: should_retry = False with sv.managed_session(master, start_standard_services=False, config=session_config) as sess: logging.info('Starting Session.') if is_chief: if logdir: sv.start_standard_services(sess) elif startup_delay_steps > 0: _wait_for_step( sess, global_step, min(startup_delay_steps, number_of_steps or sys.maxint)) sv.start_queue_runners(sess) logging.info('Starting Queues.') if is_chief and sync_optimizer is not None: sv.start_queue_runners(sess, [chief_queue_runner]) try: while not sv.should_stop(): total_loss, should_stop = train_step_fn( sess, train_op, global_step, train_step_kwargs) if should_stop: logging.info('Stopping Training.') break if logdir and sv.is_chief: logging.info( 'Finished training! Saving model to disk.') sv.saver.save(sess, sv.save_path, global_step=sv.global_step) except: if sv.is_chief and cleanup_op is not None: logging.info('About to execute sync_clean_up_op!') sess.run(cleanup_op) raise except errors.AbortedError: # Always re-run on AbortedError as it indicates a restart of one of the # distributed tensorflow servers. logging.info('Retrying training!') should_retry = True return total_loss
def train(train_op, logdir, train_step_fn=train_step, train_step_kwargs=_USE_DEFAULT, log_every_n_steps=1, graph=None, master='', is_chief=True, global_step=None, number_of_steps=None, init_op=_USE_DEFAULT, init_feed_dict=None, local_init_op=_USE_DEFAULT, init_fn=None, ready_op=_USE_DEFAULT, summary_op=_USE_DEFAULT, save_summaries_secs=600, summary_writer=_USE_DEFAULT, startup_delay_steps=0, saver=None, save_interval_secs=600, sync_optimizer=None, session_config=None): """Runs a training loop using a TensorFlow supervisor. When the sync_optimizer is supplied, gradient updates are applied synchronously. Otherwise, gradient updates are applied asynchronous. Args: train_op: A `Tensor` that, when executed, will apply the gradients and return the loss value. logdir: The directory where training logs are written to. If None, model checkpoints and summaries will not be written. train_step_fn: The function to call in order to execute a single gradient step. The function must have take exactly four arguments: the current session, the `train_op` `Tensor`, a global step `Tensor` and a dictionary. train_step_kwargs: A dictionary which is passed to the `train_step_fn`. By default, two `Boolean`, scalar ops called "should_stop" and "should_log" are provided. log_every_n_steps: The frequency, in terms of global steps, that the loss and global step and logged. graph: The graph to pass to the supervisor. If no graph is supplied the default graph is used. master: The BNS name of the tensorflow master. is_chief: Specifies whether or not the training is being run by the primary replica during replica training. global_step: The `Tensor` representing the global step. If left as `None`, then slim.variables.get_or_create_global_step() is used. number_of_steps: The max number of gradient steps to take during training. If the value is left as None, training proceeds indefinitely. init_op: The initialization operation. If left to its default value, then the session is initialized by calling `tf.initialize_all_variables()`. init_feed_dict: A feed dictionary to use when executing the `init_op`. local_init_op: The local initialization operation. If left to its default value, then the session is initialized by calling `tf.initialize_local_variables()` and `tf.initialize_all_tables()`. init_fn: An optional callable to be executed after `init_op` is called. The callable must accept one argument, the session being initialized. ready_op: Operation to check if the model is ready to use. If left to its default value, then the session checks for readiness by calling `tf.report_uninitialized_variables()`. summary_op: The summary operation. save_summaries_secs: How often, in seconds, to save summaries. summary_writer: `SummaryWriter` to use. Can be `None` to indicate that no summaries should be written. If unset, we create a SummaryWriter. startup_delay_steps: The number of steps to wait for before beginning. Note that this must be 0 if a sync_optimizer is supplied. saver: Saver to save checkpoints. If None, a default one will be created and used. save_interval_secs: How often, in seconds, to save the model to `logdir`. sync_optimizer: an instance of tf.train.SyncReplicasOptimizer. If the argument is supplied, gradient updates will be synchronous. If left as `None`, gradient updates will be asynchronous. session_config: An instance of `tf.ConfigProto` that will be used to configure the `Session`. If left as `None`, the default will be used. Returns: the value of the loss function after training. Raises: ValueError: if `train_op` is empty or if `startup_delay_steps` is non-zero when `sync_optimizer` is supplied, or if `number_of_steps` is negative. """ if train_op is None: raise ValueError('train_op cannot be None.') if logdir is None: if summary_op != _USE_DEFAULT: raise ValueError('Cannot provide summary_op because logdir=None') if saver is not None: raise ValueError('Cannot provide saver because logdir=None') if sync_optimizer and startup_delay_steps > 0: raise ValueError( 'startup_delay_steps must be zero when sync_optimizer is supplied.') if number_of_steps is not None and number_of_steps <= 0: raise ValueError( '`number_of_steps` must be either None or a positive number.') graph = graph or ops.get_default_graph() with graph.as_default(): if global_step is None: global_step = variables.get_or_create_global_step() saver = saver or tf_saver.Saver() if init_op == _USE_DEFAULT: init_op = tf_variables.initialize_all_variables() if ready_op == _USE_DEFAULT: ready_op = tf_variables.report_uninitialized_variables() if summary_op == _USE_DEFAULT: summary_op = logging_ops.merge_all_summaries() if summary_writer == _USE_DEFAULT: summary_writer = supervisor.Supervisor.USE_DEFAULT if local_init_op == _USE_DEFAULT: local_init_op = control_flow_ops.group( tf_variables.initialize_local_variables(), data_flow_ops.initialize_all_tables()) cleanup_op = None if is_chief and sync_optimizer: if not isinstance(sync_optimizer, sync_replicas_optimizer.SyncReplicasOptimizer): raise ValueError( '`sync_optimizer` must be a tf.train.SyncReplicasOptimizer') # Need to create these BEFORE the supervisor finalizes the graph: with ops.control_dependencies([init_op]): init_tokens_op = sync_optimizer.get_init_tokens_op() init_op = init_tokens_op chief_queue_runner = sync_optimizer.get_chief_queue_runner() cleanup_op = sync_optimizer.get_clean_up_op() if train_step_kwargs == _USE_DEFAULT: train_step_kwargs = {} if number_of_steps: should_stop_op = math_ops.greater_equal(global_step, number_of_steps) else: should_stop_op = constant_op.constant(False) train_step_kwargs['should_stop'] = should_stop_op train_step_kwargs['should_log'] = math_ops.equal( math_ops.mod(global_step, log_every_n_steps), 0) sv = supervisor.Supervisor( graph=graph, is_chief=is_chief, logdir=logdir, init_op=init_op, init_feed_dict=init_feed_dict, local_init_op=local_init_op, ready_op=ready_op, summary_op=summary_op, summary_writer=summary_writer, global_step=global_step, saver=saver, save_summaries_secs=save_summaries_secs, save_model_secs=save_interval_secs, init_fn=init_fn) should_retry = True while should_retry: try: should_retry = False with sv.managed_session( master, start_standard_services=False, config=session_config) as sess: logging.info('Starting Session.') if is_chief: if logdir: sv.start_standard_services(sess) elif startup_delay_steps > 0: _wait_for_step(sess, global_step, min(startup_delay_steps, number_of_steps or sys.maxint)) sv.start_queue_runners(sess) logging.info('Starting Queues.') if is_chief and sync_optimizer: sv.start_queue_runners(sess, [chief_queue_runner]) try: while not sv.should_stop(): total_loss, should_stop = train_step_fn( sess, train_op, global_step, train_step_kwargs) if should_stop: logging.info('Stopping Training.') break if logdir and sv.is_chief: logging.info('Finished training! Saving model to disk.') sv.saver.save(sess, sv.save_path, global_step=sv.global_step) except: if sv.is_chief and cleanup_op is not None: logging.info('About to execute sync_clean_up_op!') sess.run(cleanup_op) raise except errors.AbortedError: # Always re-run on AbortedError as it indicates a restart of one of the # distributed tensorflow servers. logging.info('Retrying training!') should_retry = True return total_loss
def train(train_op, logdir, train_step_fn=train_step, train_step_kwargs=_USE_DEFAULT, log_every_n_steps=1, graph=None, master='', is_chief=True, global_step=None, number_of_steps=None, init_op=_USE_DEFAULT, init_feed_dict=None, local_init_op=None, init_fn=None, summary_op=_USE_DEFAULT, save_summaries_secs=600, startup_delay_steps=0, saver=None, save_interval_secs=600, sync_optimizer=None): """Runs a training loop using a TensorFlow supervisor. When the sync_optimizer is supplied, gradient updates are applied synchronously. Otherwise, gradient updates are applied asynchronous. Args: train_op: A `Tensor` that, when executed, will apply the gradients and return the loss value. logdir: The directory where training logs are written to. train_step_fn: The function to call in order to execute a single gradient step. The function must have take exactly four arguments: the current session, the `train_op` `Tensor`, a global step `Tensor` and a dictionary. train_step_kwargs: A dictionary which is passed to the `train_step_fn`. By default, two `Boolean`, scalar ops called "should_stop" and "should_log" are provided. log_every_n_steps: The frequency, in terms of global steps, that the loss and global step and logged. graph: The graph to pass to the supervisor. If no graph is supplied the default graph is used. master: The BNS name of the tensorflow master. is_chief: Specifies whether or not the training is being run by the primary replica during replica training. global_step: The `Tensor` representing the global step. If left as `None`, then slim.variables.get_or_create_global_step() is used. number_of_steps: The max number of gradient steps to take during training. If the value is left as None, training proceeds indefinitely. init_op: The initialization operation. If left to its default value, then the session is initialized by calling `tf.initialize_all_variables()`. init_feed_dict: A feed dictionary to use when executing the `init_op`. local_init_op: The local initialization operation. If None, then the session is initialized by calling `tf.initialize_local_variables()` and `tf.initialize_all_tables()`. init_fn: An optional callable to be executed after `init_op` is called. The callable must accept one argument, the session being initialized. summary_op: The summary operation. save_summaries_secs: How often, in seconds, to save summaries. startup_delay_steps: The number of steps to wait for before beginning. Note that this must be 0 if a sync_optimizer is supplied. saver: Saver to save checkpoints. If none, a default one will be created and used. save_interval_secs: How often, in seconds, to save the model to `logdir`. sync_optimizer: an instance of tf.train.SyncReplicasOptimizer. If the argument is supplied, gradient updates will be synchronous. If left as `None`, gradient updates will be asynchronous. Returns: the value of the loss function after training. Raises: ValueError: if `train_op` is empty or if `startup_delay_steps` is non-zero when `sync_optimizer` is supplied, or if `number_of_steps` is negative. """ if train_op is None: raise ValueError('train_op cannot be None.') if sync_optimizer and startup_delay_steps > 0: raise ValueError( 'startup_delay_steps must be zero when sync_optimizer is supplied.' ) if number_of_steps is not None and number_of_steps <= 0: raise ValueError( '`number_of_steps` must be either None or a positive number.') graph = graph or ops.get_default_graph() with graph.as_default(): if global_step is None: global_step = variables.get_or_create_global_step() saver = saver or tf_saver.Saver() if init_op == _USE_DEFAULT: init_op = tf_variables.initialize_all_variables() if summary_op == _USE_DEFAULT: summary_op = logging_ops.merge_all_summaries() cleanup_op = None if is_chief and sync_optimizer: if not isinstance(sync_optimizer, sync_replicas_optimizer.SyncReplicasOptimizer): raise ValueError( '`sync_optimizer` must be a tf.train.SyncReplicasOptimizer') # Need to create these BEFORE the supervisor finalizes the graph: with ops.control_dependencies([init_op]): init_tokens_op = sync_optimizer.get_init_tokens_op() init_op = init_tokens_op chief_queue_runner = sync_optimizer.get_chief_queue_runner() cleanup_op = sync_optimizer.get_clean_up_op() if train_step_kwargs == _USE_DEFAULT: train_step_kwargs = {} if number_of_steps: should_stop_op = math_ops.greater_equal(global_step, number_of_steps) else: should_stop_op = constant_op.constant(False) train_step_kwargs['should_stop'] = should_stop_op train_step_kwargs['should_log'] = math_ops.equal( math_ops.mod(global_step, log_every_n_steps), 0) sv = supervisor.Supervisor(graph=graph, is_chief=is_chief, logdir=logdir, init_op=init_op, init_feed_dict=init_feed_dict, local_init_op=local_init_op, summary_op=summary_op, global_step=global_step, saver=saver, save_summaries_secs=save_summaries_secs, save_model_secs=save_interval_secs, init_fn=init_fn) with sv.managed_session(master, start_standard_services=False) as sess: if is_chief: sv.start_standard_services(sess) elif not is_chief and startup_delay_steps > 0: _wait_for_step( sess, global_step, min(startup_delay_steps, number_of_steps or sys.maxint)) sv.start_queue_runners(sess) if is_chief and sync_optimizer: sv.start_queue_runners(sess, [chief_queue_runner]) try: while not sv.should_stop(): total_loss, should_stop = train_step_fn( sess, train_op, global_step, train_step_kwargs) if should_stop: break finally: if sv.is_chief and cleanup_op is not None: sess.run(cleanup_op) # This waits for service threads to finish. sv.Stop() if sv.is_chief: logging.info('Finished training! Saving model to disk.') sv.saver.save(sess, sv.save_path, global_step=sv.global_step) return total_loss
def evaluation_loop(master, checkpoint_dir, logdir, num_evals=1, initial_op=None, initial_op_feed_dict=None, eval_op=None, eval_op_feed_dict=None, final_op=None, final_op_feed_dict=None, summary_op=_USE_DEFAULT, summary_op_feed_dict=None, variables_to_restore=None, eval_interval_secs=60, max_number_of_evaluations=None, session_config=None): """Runs TF-Slim's Evaluation Loop. Args: master: The BNS address of the TensorFlow master. checkpoint_dir: The directory where checkpoints are stored. logdir: The directory where the TensorFlow summaries are written to. num_evals: The number of times to run `eval_op`. initial_op: An operation run at the beginning of evaluation. initial_op_feed_dict: A feed dictionary to use when executing `initial_op`. eval_op: A operation run `num_evals` times. eval_op_feed_dict: The feed dictionary to use when executing the `eval_op`. final_op: An operation to execute after all of the `eval_op` executions. The value of `final_op` is returned. final_op_feed_dict: A feed dictionary to use when executing `final_op`. summary_op: The summary_op to evaluate after running TF-Slims metric ops. By default the summary_op is set to tf.merge_all_summaries(). summary_op_feed_dict: An optional feed dictionary to use when running the `summary_op`. variables_to_restore: A list of TensorFlow variables to restore during evaluation. If the argument is left as `None` then slim.variables.GetVariablesToRestore() is used. eval_interval_secs: The minimum number of seconds between evaluations. max_number_of_evaluations: the max number of iterations of the evaluation. If the value is left as 'None', the evaluation continues indefinitely. session_config: An instance of `tf.ConfigProto` that will be used to configure the `Session`. If left as `None`, the default will be used. Returns: The value of `final_op` or `None` if `final_op` is `None`. """ if summary_op == _USE_DEFAULT: summary_op = logging_ops.merge_all_summaries() global_step = variables.get_or_create_global_step() saver = tf_saver.Saver(variables_to_restore or variables.get_variables_to_restore()) summary_writer = summary_io.SummaryWriter(logdir) sv = supervisor.Supervisor(graph=ops.get_default_graph(), logdir=logdir, summary_op=None, summary_writer=None, global_step=None, saver=saver) number_of_evaluations = 0 for checkpoint_path in checkpoints_iterator(checkpoint_dir, eval_interval_secs): logging.info('Starting evaluation at ' + time.strftime('%Y-%m-%d-%H:%M:%S', time.gmtime())) with sv.managed_session(master, start_standard_services=False, config=session_config) as sess: sv.saver.restore(sess, checkpoint_path) sv.start_queue_runners(sess) final_op_value = evaluation( sess, num_evals=num_evals, initial_op=initial_op, initial_op_feed_dict=initial_op_feed_dict, eval_op=eval_op, eval_op_feed_dict=eval_op_feed_dict, final_op=final_op, final_op_feed_dict=final_op_feed_dict, summary_op=summary_op, summary_op_feed_dict=summary_op_feed_dict, summary_writer=summary_writer, global_step=global_step) logging.info('Finished evaluation at ' + time.strftime('%Y-%m-%d-%H:%M:%S', time.gmtime())) number_of_evaluations += 1 if (max_number_of_evaluations and number_of_evaluations >= max_number_of_evaluations): logging.info('Reached max_number_of_evaluations=%s. Exit', max_number_of_evaluations) break return final_op_value
def evaluate_once(master, checkpoint_path, logdir, num_evals=1, eval_op=None, eval_op_feed_dict=None, final_op=None, final_op_feed_dict=None, summary_op=_USE_DEFAULT, summary_op_feed_dict=None, variables_to_restore=None, session_config=None): """Evaluates the model at the given checkpoint path. Args: master: The BNS address of the TensorFlow master. checkpoint_path: The path to a checkpoint to use for evaluation. logdir: The directory where the TensorFlow summaries are written to. num_evals: The number of times to run `eval_op`. eval_op: A operation run `num_evals` times. eval_op_feed_dict: The feed dictionary to use when executing the `eval_op`. final_op: An operation to execute after all of the `eval_op` executions. The value of `final_op` is returned. final_op_feed_dict: A feed dictionary to use when executing `final_op`. summary_op: The summary_op to evaluate after running TF-Slims metric ops. By default the summary_op is set to tf.merge_all_summaries(). summary_op_feed_dict: An optional feed dictionary to use when running the `summary_op`. variables_to_restore: A list of TensorFlow variables to restore during evaluation. If the argument is left as `None` then slim.variables.GetVariablesToRestore() is used. session_config: An instance of `tf.ConfigProto` that will be used to configure the `Session`. If left as `None`, the default will be used. Returns: The value of `final_op` or `None` if `final_op` is `None`. """ if summary_op == _USE_DEFAULT: summary_op = logging_ops.merge_all_summaries() global_step = variables.get_or_create_global_step() init_op = control_flow_ops.group(tf_variables.initialize_all_variables(), tf_variables.initialize_local_variables(), data_flow_ops.initialize_all_tables()) saver = tf_saver.Saver(variables_to_restore or variables.get_variables_to_restore()) summary_writer = summary_io.SummaryWriter(logdir) sv = supervisor.Supervisor(graph=ops.get_default_graph(), logdir=logdir, init_op=init_op, summary_op=None, summary_writer=None, global_step=None, saver=None) logging.info('Starting evaluation at ' + time.strftime('%Y-%m-%d-%H:%M:%S', time.gmtime())) with sv.managed_session( master, start_standard_services=False, config=session_config) as sess: saver.restore(sess, checkpoint_path) sv.start_queue_runners(sess) final_op_value = evaluation(sess, num_evals=num_evals, eval_op=eval_op, eval_op_feed_dict=eval_op_feed_dict, final_op=final_op, final_op_feed_dict=final_op_feed_dict, summary_op=summary_op, summary_op_feed_dict=summary_op_feed_dict, summary_writer=summary_writer, global_step=global_step) logging.info('Finished evaluation at ' + time.strftime('%Y-%m-%d-%H:%M:%S', time.gmtime())) return final_op_value