-
Notifications
You must be signed in to change notification settings - Fork 5.6k
/
checkpoint_manager.py
225 lines (185 loc) · 7.76 KB
/
checkpoint_manager.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
# coding: utf-8
import heapq
import gc
import logging
from typing import Any, Callable, Optional
from ray.tune.result import NODE_IP
from ray.tune.utils.util import flatten_dict, is_nan
logger = logging.getLogger(__name__)
class _TuneCheckpoint:
"""Describes a checkpoint of trial state.
Checkpoint may be saved in different storage.
Attributes:
storage: Storage type.
value: If storage==MEMORY, it is a Python object.
If storage==PERSISTENT, it is a path to persistent storage,
or a future that will be resolved to such a path.
"""
MEMORY = "memory"
PERSISTENT = "persistent"
def __init__(
self,
storage: str,
value: Any,
result: Optional[dict] = None,
node_ip: Optional[str] = None,
):
self.storage = storage
self.value = value
self.result = result or {}
self.node_ip = node_ip or self.result.get(NODE_IP, None)
# The logical order of checkpoints (both in memory and persistent)
# The more recent checkpoints have larger order.
# The most recent checkpoint is used to restore the trial.
self.order = 0
@staticmethod
def from_object(value=None):
"""Creates a checkpoint from a Python object."""
return _TuneCheckpoint(_TuneCheckpoint.MEMORY, value)
@property
def is_ready(self):
"""Returns whether the checkpoint is ready to be used for restoration.
A PERSISTENT checkpoint is considered ready once its value is resolved
to an actual path. MEMORY checkpoints are always considered ready since
they are transient.
"""
if self.storage == _TuneCheckpoint.PERSISTENT:
return isinstance(self.value, str)
return self.storage == _TuneCheckpoint.MEMORY
def __repr__(self):
return f"Checkpoint({self.storage}, {self.value})"
class QueueItem:
def __init__(self, priority, value):
self.priority = priority
self.value = value
def __lt__(self, other):
return self.priority < other.priority
def __repr__(self):
return f"QueueItem({repr(self.value)})"
class CheckpointManager:
"""Manages checkpoints on the driver for a trial."""
def __init__(
self,
keep_checkpoints_num: int,
checkpoint_score_attr: str,
delete_fn: Callable[[str], None],
):
"""Initializes a new CheckpointManager.
`newest_persistent_checkpoint` and `newest_memory_checkpoint` are
initialized to Checkpoint objects with values of None.
Args:
keep_checkpoints_num: Keep at least this many checkpoints.
checkpoint_score_attr: Attribute to use to determine which
checkpoints to keep.
delete_fn: Function that deletes checkpoints. Must be
idempotent.
"""
self.keep_checkpoints_num = keep_checkpoints_num or float("inf")
assert (
self.keep_checkpoints_num > 0
), "keep_checkpoints_num must be greater than 0."
self._checkpoint_score_desc = checkpoint_score_attr.startswith("min-")
if self._checkpoint_score_desc:
self._checkpoint_score_attr = checkpoint_score_attr[4:]
else:
self._checkpoint_score_attr = checkpoint_score_attr
self.delete = delete_fn
self.newest_persistent_checkpoint = _TuneCheckpoint(
_TuneCheckpoint.PERSISTENT, None
)
self._newest_memory_checkpoint = _TuneCheckpoint(_TuneCheckpoint.MEMORY, None)
self._best_checkpoints = []
self._membership = set()
self._cur_order = 0
@property
def newest_checkpoint(self):
"""Returns the newest checkpoint (based on training iteration)."""
newest_checkpoint = max(
[self.newest_persistent_checkpoint, self.newest_memory_checkpoint],
key=lambda c: c.order,
)
return newest_checkpoint
@property
def newest_memory_checkpoint(self):
return self._newest_memory_checkpoint
def replace_newest_memory_checkpoint(self, new_checkpoint):
# Forcibly remove the memory checkpoint
del self._newest_memory_checkpoint
# Apparently avoids memory leaks on k8s/k3s/pods
gc.collect()
self._newest_memory_checkpoint = new_checkpoint
def on_checkpoint(self, checkpoint: _TuneCheckpoint):
"""Starts tracking checkpoint metadata on checkpoint.
Checkpoints get assigned with an `order` as they come in.
The order is monotonically increasing.
Sets the newest checkpoint. For PERSISTENT checkpoints: Deletes
previous checkpoint as long as it isn't one of the best ones. Also
deletes the worst checkpoint if at capacity.
Args:
checkpoint: Trial state checkpoint.
"""
self._cur_order += 1
checkpoint.order = self._cur_order
if checkpoint.storage == _TuneCheckpoint.MEMORY:
self.replace_newest_memory_checkpoint(checkpoint)
return
old_checkpoint = self.newest_persistent_checkpoint
if old_checkpoint.value == checkpoint.value:
# Overwrite the order of the checkpoint.
old_checkpoint.order = checkpoint.order
return
self.newest_persistent_checkpoint = checkpoint
# Remove the old checkpoint if it isn't one of the best ones.
if old_checkpoint.value and old_checkpoint not in self._membership:
self.delete(old_checkpoint)
try:
# NaN metrics are treated as worst checkpoint
# The tuple structure is (not is_nan(), metric), which makes
# the nan values to be always considered as the worst
# metrics by the heap
queue_item = QueueItem(self._priority(checkpoint), checkpoint)
except KeyError:
logger.error(
"Result dict has no key: {}. "
"checkpoint_score_attr must be set to a key in the "
"result dict.".format(self._checkpoint_score_attr)
)
return
if len(self._best_checkpoints) < self.keep_checkpoints_num:
heapq.heappush(self._best_checkpoints, queue_item)
self._membership.add(checkpoint)
elif queue_item.priority >= self._best_checkpoints[0].priority:
worst = heapq.heappushpop(self._best_checkpoints, queue_item).value
self._membership.add(checkpoint)
if worst in self._membership:
self._membership.remove(worst)
# Don't delete the newest checkpoint. It will be deleted on the
# next on_checkpoint() call since it isn't in self._membership.
if worst.value != checkpoint.value:
self.delete(worst)
def best_checkpoints(self):
"""Returns best PERSISTENT checkpoints, sorted by score."""
checkpoints = sorted(self._best_checkpoints, key=lambda c: c.priority)
return [queue_item.value for queue_item in checkpoints]
def _priority(self, checkpoint):
result = flatten_dict(checkpoint.result)
priority = result[self._checkpoint_score_attr]
if self._checkpoint_score_desc:
priority = -priority
return (
not is_nan(priority),
priority if not is_nan(priority) else 0,
checkpoint.order,
)
def __getstate__(self):
state = self.__dict__.copy()
# Avoid serializing the memory checkpoint.
state["_newest_memory_checkpoint"] = _TuneCheckpoint(
_TuneCheckpoint.MEMORY, None
)
# Avoid serializing lambda since it may capture cyclical dependencies.
state.pop("delete")
return state
def __setstate__(self, state):
self.__dict__.update(state)
self.delete = None