|
1
|
|
|
from typing import Tuple |
|
2
|
|
|
|
|
3
|
|
|
import tensorflow as tf |
|
4
|
|
|
|
|
5
|
|
|
|
|
6
|
|
|
class CheckpointManagerCallback(tf.keras.callbacks.Callback): |
|
7
|
|
|
def __init__( |
|
8
|
|
|
self, model, directory, period: int = 1, save_on_train_end: bool = True |
|
9
|
|
|
): |
|
10
|
|
|
""" |
|
11
|
|
|
Callback wrapping `tf.train.CheckpointManager`. |
|
12
|
|
|
|
|
13
|
|
|
:param model: model |
|
14
|
|
|
:param directory: directory to store the checkpoints |
|
15
|
|
|
:param period: save the checkpoint every X epochs |
|
16
|
|
|
:param save_on_train_end: save the checkpoint as the training ends |
|
17
|
|
|
""" |
|
18
|
|
|
super().__init__() |
|
19
|
|
|
self._directory = directory |
|
20
|
|
|
|
|
21
|
|
|
self._checkpoint = tf.train.Checkpoint(model=model, optimizer=model.optimizer) |
|
22
|
|
|
self._manager = tf.train.CheckpointManager( |
|
23
|
|
|
checkpoint=self._checkpoint, directory=self._directory, max_to_keep=None |
|
24
|
|
|
) |
|
25
|
|
|
self._period = period |
|
26
|
|
|
self._save_on_train_end = save_on_train_end |
|
27
|
|
|
self._restored = False |
|
28
|
|
|
self._epoch_count = None |
|
29
|
|
|
self._last_save = None |
|
30
|
|
|
|
|
31
|
|
|
def _on_begin(self): |
|
32
|
|
|
if not self._restored: |
|
33
|
|
|
self.restore() |
|
34
|
|
|
|
|
35
|
|
|
def restore(self, save_path=None): |
|
36
|
|
|
if save_path is None: |
|
37
|
|
|
save_path = self._manager.latest_checkpoint |
|
38
|
|
|
self._checkpoint.restore(save_path) |
|
39
|
|
|
self._restored = True |
|
40
|
|
|
|
|
41
|
|
|
def on_train_begin(self, logs=None): |
|
42
|
|
|
self._on_begin() |
|
43
|
|
|
|
|
44
|
|
|
def on_epoch_end(self, epoch, logs=None): |
|
45
|
|
|
epochs_finished = epoch + 1 |
|
46
|
|
|
self._epoch_count = epochs_finished |
|
47
|
|
|
if epochs_finished % self._period == 0: |
|
48
|
|
|
self._save() |
|
49
|
|
|
|
|
50
|
|
|
def on_train_end(self, logs=None): |
|
51
|
|
|
if self._save_on_train_end: |
|
52
|
|
|
self._save() |
|
53
|
|
|
|
|
54
|
|
|
def _save(self): |
|
55
|
|
|
""" |
|
56
|
|
|
checkpoint saved as f"{self._directory}/ckpt-{self._epoch_count}" |
|
57
|
|
|
""" |
|
58
|
|
|
if self._last_save != self._epoch_count: |
|
59
|
|
|
self._manager.save(checkpoint_number=self._epoch_count) |
|
60
|
|
|
self._last_save = self._epoch_count |
|
61
|
|
|
|
|
62
|
|
|
|
|
63
|
|
|
def build_checkpoint_callback( |
|
64
|
|
|
model: tf.keras.Model, |
|
65
|
|
|
dataset: tf.data.Dataset, |
|
66
|
|
|
log_dir: str, |
|
67
|
|
|
save_period: int, |
|
68
|
|
|
ckpt_path: str, |
|
69
|
|
|
) -> Tuple[CheckpointManagerCallback, int]: |
|
70
|
|
|
""" |
|
71
|
|
|
Function to prepare callbacks for training. |
|
72
|
|
|
|
|
73
|
|
|
:param model: model to train |
|
74
|
|
|
:param dataset: dataset for training |
|
75
|
|
|
:param log_dir: directory of logs |
|
76
|
|
|
:param save_period: save the checkpoint every X epochs |
|
77
|
|
|
:param ckpt_path: path to restore ckpt |
|
78
|
|
|
:return: a list of callbacks |
|
79
|
|
|
""" |
|
80
|
|
|
# fit the model for 1 step to initialise optimiser arguments as trackable Variables |
|
81
|
|
|
model.fit( |
|
82
|
|
|
x=dataset, |
|
83
|
|
|
steps_per_epoch=1, |
|
84
|
|
|
epochs=1, |
|
85
|
|
|
verbose=0, |
|
86
|
|
|
) |
|
87
|
|
|
checkpoint_manager_callback = CheckpointManagerCallback( |
|
88
|
|
|
model, log_dir + "/save", period=save_period |
|
89
|
|
|
) |
|
90
|
|
|
if ckpt_path: |
|
91
|
|
|
initial_epoch_str = ckpt_path.split("-")[-1] |
|
92
|
|
|
assert initial_epoch_str.isdigit(), ( |
|
93
|
|
|
f"Checkpoint path for checkpoint manager " |
|
94
|
|
|
f"must be of form ckpt-epoch_count, got {ckpt_path}" |
|
95
|
|
|
) |
|
96
|
|
|
initial_epoch = int(initial_epoch_str) |
|
97
|
|
|
checkpoint_manager_callback.restore(ckpt_path) |
|
98
|
|
|
else: |
|
99
|
|
|
initial_epoch = 0 |
|
100
|
|
|
return checkpoint_manager_callback, initial_epoch |
|
101
|
|
|
|
DeepRegNet /
DeepReg
