1# Copyright 2019 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"""Training state management.""" 16from __future__ import absolute_import 17from __future__ import division 18from __future__ import print_function 19 20import os 21 22from tensorflow.python.framework import constant_op 23from tensorflow.python.framework import dtypes 24from tensorflow.python.keras import backend as K 25from tensorflow.python.keras.distribute import distributed_file_utils 26from tensorflow.python.keras.utils import mode_keys 27from tensorflow.python.lib.io import file_io 28from tensorflow.python.ops import variables 29from tensorflow.python.training import checkpoint_management 30from tensorflow.python.training.tracking import util as trackable_util 31 32# Constant for `tf.keras.Model` attribute to store the epoch at which the most 33# recently saved checkpoint was saved. 34CKPT_SAVED_EPOCH = '_ckpt_saved_epoch' 35 36CKPT_SAVED_EPOCH_UNUSED_VALUE = -1 37 38 39class WorkerTrainingState(object): 40 """Training state management class. 41 42 This class provides apis for backing up and restoring the training state. 43 This allows model and epoch information to be saved periodically and restore 44 for fault-tolerance, also known as preemption-recovery purpose. 45 """ 46 47 def __init__(self, model, checkpoint_dir): 48 self._model = model 49 50 # The epoch at which the checkpoint is saved. Used for fault-tolerance. 51 # GPU device only has int64 dtype registered VarHandleOp. 52 self._ckpt_saved_epoch = variables.Variable( 53 initial_value=constant_op.constant( 54 CKPT_SAVED_EPOCH_UNUSED_VALUE, dtype=dtypes.int64), 55 name='ckpt_saved_epoch') 56 57 # Variable initialization. 58 K.set_value(self._ckpt_saved_epoch, CKPT_SAVED_EPOCH_UNUSED_VALUE) 59 60 # _ckpt_saved_epoch gets tracked and is included in the checkpoint file 61 # when backing up. 62 checkpoint = trackable_util.Checkpoint( 63 model=self._model, ckpt_saved_epoch=self._ckpt_saved_epoch) 64 65 # If this is single-worker training, checkpoint_dir are the same for 66 # write_checkpoint_manager and read_checkpoint_manager. 67 # 68 # If this is multi-worker training, and this worker should not 69 # save checkpoint, we replace the write_checkpoint_manager's checkpoint_dir 70 # with a temp filepath, so it writes to a file that will be removed at the 71 # end of back_up() call. This is necessary because the SyncOnReadVariable 72 # needs to be synced across all the workers in order to be read, and all 73 # workers need to perform `save()`. 74 # But all workers should restore from the same checkpoint_dir as passed in 75 # read_checkpoint_manager. 76 self.read_checkpoint_manager = checkpoint_management.CheckpointManager( 77 checkpoint, 78 directory=os.path.join(checkpoint_dir, 'chief'), 79 max_to_keep=1) 80 write_checkpoint_dir = distributed_file_utils.write_dirpath( 81 checkpoint_dir, self._model.distribute_strategy) 82 if self._model.distribute_strategy.extended.should_checkpoint: 83 self.write_checkpoint_manager = self.read_checkpoint_manager 84 else: 85 self.write_checkpoint_manager = checkpoint_management.CheckpointManager( 86 checkpoint, directory=write_checkpoint_dir, max_to_keep=1) 87 88 def back_up(self, epoch): 89 """Back up the current state of training into a checkpoint file. 90 91 Args: 92 epoch: The current epoch information to be saved. 93 """ 94 K.set_value(self._ckpt_saved_epoch, epoch) 95 # Save the model plus CKPT_SAVED_EPOCH variable. 96 if self.write_checkpoint_manager.save(): 97 distributed_file_utils.remove_temp_dirpath( 98 self.write_checkpoint_manager.directory, 99 self._model.distribute_strategy) 100 101 def restore(self): 102 """Restore the training state from the backed up checkpoint file. 103 104 Returns: 105 True if the training state is successfully restored. False if the training 106 state doesn't need to be restored, or error occurred so it can't. 107 """ 108 self.read_checkpoint_manager.restore_or_initialize() 109 110 def delete_backup(self): 111 """Delete the backup directories. 112 113 Delete the backup directories which should not exist after `fit()` 114 successfully finishes. 115 """ 116 if self.write_checkpoint_manager is self.read_checkpoint_manager: 117 file_io.delete_recursively_v2(self.write_checkpoint_manager.directory) 118 119 def maybe_load_initial_epoch_from_ckpt(self, initial_epoch, mode): 120 """Maybe load initial epoch from ckpt considering possible worker recovery. 121 122 When `_ckpt_saved_epoch` attribute exists and is not 123 `CKPT_SAVED_EPOCH_UNUSED_VALUE`, this is under multi-worker training setting 124 and indicates the worker is recovering from previous failure. In this case, 125 infer `initial_epoch` from `self._ckpt_saved_epoch` to continue previous 126 unfinished training from certain epoch. 127 128 Args: 129 initial_epoch: The original initial_epoch user passes in in `fit()`. 130 mode: The mode for running `model.fit()`. 131 132 Returns: 133 If the training is recovering from previous failure under multi-worker 134 training setting, return the epoch the training is supposed to continue 135 at. Otherwise, return the `initial_epoch` the user passes in. 136 """ 137 138 epoch = K.eval(self._ckpt_saved_epoch) 139 if mode == mode_keys.ModeKeys.TRAIN and epoch >= 0: 140 # The most recently saved epoch is one epoch prior to the epoch it 141 # failed at, so return the value of 'self._ckpt_saved_epoch' plus one. 142 return epoch + 1 143 return initial_epoch 144