Newer
Older
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
# SPDX-FileCopyrightText: Copyright © 2023 Idiap Research Institute <contact@idiap.ch>
#
# SPDX-License-Identifier: GPL-3.0-or-later
import logging
import os
import torch
logger = logging.getLogger(__name__)
class Checkpointer:
"""A simple pytorch checkpointer.
Parameters
----------
model : torch.nn.Module
Network model, eventually loaded from a checkpointed file
optimizer : :py:mod:`torch.optim`, Optional
Optimizer
scheduler : :py:mod:`torch.optim`, Optional
Learning rate scheduler
path : :py:class:`str`, Optional
Directory where to save checkpoints.
"""
def __init__(self, model, optimizer=None, scheduler=None, path="."):
self.model = model
self.optimizer = optimizer
self.scheduler = scheduler
self.path = os.path.realpath(path)
def save(self, name, **kwargs):
data = {}
data["model"] = self.model.state_dict()
if self.optimizer is not None:
data["optimizer"] = self.optimizer.state_dict()
if self.scheduler is not None:
data["scheduler"] = self.scheduler.state_dict()
data.update(kwargs)
name = f"{name}.pth"
outf = os.path.join(self.path, name)
logger.info(f"Saving checkpoint to {outf}")
torch.save(data, outf)
with open(self._last_checkpoint_filename, "w") as f:
f.write(name)
def load(self, f=None, strict=True):
"""Loads model, optimizer and scheduler from file.
Parameters
==========
f : :py:class:`str`, Optional
Name of a file (absolute or relative to ``self.path``), that
contains the checkpoint data to load into the model, and optionally
into the optimizer and the scheduler. If not specified, loads data
from current path.
partial : :py:class:`bool`, Optional
If True, loading is not strict and only the model is loaded
"""
if f is None:
f = self.last_checkpoint()
if f is None:
# no checkpoint could be found
logger.warning("No checkpoint found (and none passed)")
return {}
# loads file data into memory
logger.info(f"Loading checkpoint from {f}...")
checkpoint = torch.load(f, map_location=torch.device("cpu"))
# converts model entry to model parameters
self.model.load_state_dict(checkpoint.pop("model"), strict=strict)
if strict:
if self.optimizer is not None:
self.optimizer.load_state_dict(checkpoint.pop("optimizer"))
if self.scheduler is not None:
self.scheduler.load_state_dict(checkpoint.pop("scheduler"))
return checkpoint
@property
def _last_checkpoint_filename(self):
return os.path.join(self.path, "last_checkpoint")
def has_checkpoint(self):
return os.path.exists(self._last_checkpoint_filename)
def last_checkpoint(self):
if self.has_checkpoint():
with open(self._last_checkpoint_filename) as fobj:
return os.path.join(self.path, fobj.read().strip())
return None