-
Notifications
You must be signed in to change notification settings - Fork 3.3k
/
parsing.py
330 lines (260 loc) · 11.8 KB
/
parsing.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
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
# Copyright The Lightning AI team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Utilities used for parameter parsing."""
import copy
import inspect
import pickle
import types
from dataclasses import fields, is_dataclass
from typing import Any, Dict, List, Literal, MutableMapping, Optional, Sequence, Tuple, Type, Union
from torch import nn
import lightning.pytorch as pl
from lightning.fabric.utilities.data import AttributeDict as _AttributeDict
from lightning.pytorch.utilities.rank_zero import rank_zero_warn
def is_picklable(obj: object) -> bool:
"""Tests if an object can be pickled."""
try:
pickle.dumps(obj)
return True
except (pickle.PickleError, AttributeError, RuntimeError, TypeError):
return False
def clean_namespace(hparams: MutableMapping) -> None:
"""Removes all unpicklable entries from hparams."""
del_attrs = [k for k, v in hparams.items() if not is_picklable(v)]
for k in del_attrs:
rank_zero_warn(
f"Attribute '{k}' removed from hparams because it cannot be pickled. You can suppress this warning by"
f" setting `self.save_hyperparameters(ignore=['{k}'])`.",
)
del hparams[k]
def parse_class_init_keys(cls: Type) -> Tuple[str, Optional[str], Optional[str]]:
"""Parse key words for standard ``self``, ``*args`` and ``**kwargs``.
Examples:
>>> class Model:
... def __init__(self, hparams, *my_args, anykw=42, **my_kwargs):
... pass
>>> parse_class_init_keys(Model)
('self', 'my_args', 'my_kwargs')
"""
init_parameters = inspect.signature(cls.__init__).parameters
# docs claims the params are always ordered
# https://docs.python.org/3/library/inspect.html#inspect.Signature.parameters
init_params = list(init_parameters.values())
# self is always first
n_self = init_params[0].name
def _get_first_if_any(
params: List[inspect.Parameter],
param_type: Literal[inspect._ParameterKind.VAR_POSITIONAL, inspect._ParameterKind.VAR_KEYWORD],
) -> Optional[str]:
for p in params:
if p.kind == param_type:
return p.name
return None
n_args = _get_first_if_any(init_params, inspect.Parameter.VAR_POSITIONAL)
n_kwargs = _get_first_if_any(init_params, inspect.Parameter.VAR_KEYWORD)
return n_self, n_args, n_kwargs
def get_init_args(frame: types.FrameType) -> Dict[str, Any]: # pragma: no-cover
"""For backwards compatibility: #16369."""
_, local_args = _get_init_args(frame)
return local_args
def _get_init_args(frame: types.FrameType) -> Tuple[Optional[Any], Dict[str, Any]]:
_, _, _, local_vars = inspect.getargvalues(frame)
if "__class__" not in local_vars:
return None, {}
cls = local_vars["__class__"]
init_parameters = inspect.signature(cls.__init__).parameters
self_var, args_var, kwargs_var = parse_class_init_keys(cls)
filtered_vars = [n for n in (self_var, args_var, kwargs_var) if n]
exclude_argnames = (*filtered_vars, "__class__", "frame", "frame_args")
# only collect variables that appear in the signature
local_args = {k: local_vars[k] for k in init_parameters}
# kwargs_var might be None => raised an error by mypy
if kwargs_var:
local_args.update(local_args.get(kwargs_var, {}))
local_args = {k: v for k, v in local_args.items() if k not in exclude_argnames}
self_arg = local_vars.get(self_var, None)
return self_arg, local_args
def collect_init_args(
frame: types.FrameType,
path_args: List[Dict[str, Any]],
inside: bool = False,
classes: Tuple[Type, ...] = (),
) -> List[Dict[str, Any]]:
"""Recursively collects the arguments passed to the child constructors in the inheritance tree.
Args:
frame: the current stack frame
path_args: a list of dictionaries containing the constructor args in all parent classes
inside: track if we are inside inheritance path, avoid terminating too soon
classes: the classes in which to inspect the frames
Return:
A list of dictionaries where each dictionary contains the arguments passed to the
constructor at that level. The last entry corresponds to the constructor call of the
most specific class in the hierarchy.
"""
_, _, _, local_vars = inspect.getargvalues(frame)
# frame.f_back must be of a type types.FrameType for get_init_args/collect_init_args due to mypy
if not isinstance(frame.f_back, types.FrameType):
return path_args
local_self, local_args = _get_init_args(frame)
if "__class__" in local_vars and (not classes or isinstance(local_self, classes)):
# recursive update
path_args.append(local_args)
return collect_init_args(frame.f_back, path_args, inside=True, classes=classes)
if not inside:
return collect_init_args(frame.f_back, path_args, inside=False, classes=classes)
return path_args
def save_hyperparameters(
obj: Any,
*args: Any,
ignore: Optional[Union[Sequence[str], str]] = None,
frame: Optional[types.FrameType] = None,
given_hparams: Optional[Dict[str, Any]] = None,
) -> None:
"""See :meth:`~lightning.pytorch.LightningModule.save_hyperparameters`"""
if len(args) == 1 and not isinstance(args, str) and not args[0]:
# args[0] is an empty container
return
if not frame:
current_frame = inspect.currentframe()
# inspect.currentframe() return type is Optional[types.FrameType]: current_frame.f_back called only if available
if current_frame:
frame = current_frame.f_back
if not isinstance(frame, types.FrameType):
raise AttributeError("There is no `frame` available while being required.")
if given_hparams is not None:
init_args = given_hparams
elif is_dataclass(obj):
init_args = {f.name: getattr(obj, f.name) for f in fields(obj)}
else:
init_args = {}
from lightning.pytorch.core.mixins import HyperparametersMixin
for local_args in collect_init_args(frame, [], classes=(HyperparametersMixin,)):
init_args.update(local_args)
if ignore is None:
ignore = []
elif isinstance(ignore, str):
ignore = [ignore]
elif isinstance(ignore, (list, tuple)):
ignore = [arg for arg in ignore if isinstance(arg, str)]
ignore = list(set(ignore))
init_args = {k: v for k, v in init_args.items() if k not in ignore}
if not args:
# take all arguments
hp = init_args
obj._hparams_name = "kwargs" if hp else None
else:
# take only listed arguments in `save_hparams`
isx_non_str = [i for i, arg in enumerate(args) if not isinstance(arg, str)]
if len(isx_non_str) == 1:
hp = args[isx_non_str[0]]
cand_names = [k for k, v in init_args.items() if v == hp]
obj._hparams_name = cand_names[0] if cand_names else None
else:
hp = {arg: init_args[arg] for arg in args if isinstance(arg, str)}
obj._hparams_name = "kwargs"
# `hparams` are expected here
obj._set_hparams(hp)
for k, v in obj._hparams.items():
if isinstance(v, nn.Module):
rank_zero_warn(
f"Attribute {k!r} is an instance of `nn.Module` and is already saved during checkpointing."
f" It is recommended to ignore them using `self.save_hyperparameters(ignore=[{k!r}])`."
)
# make a deep copy so there are no other runtime changes reflected
obj._hparams_initial = copy.deepcopy(obj._hparams)
class AttributeDict(_AttributeDict):
"""Extended dictionary accessible with dot notation.
>>> ad = AttributeDict({'key1': 1, 'key2': 'abc'})
>>> ad.key1
1
>>> ad.update({'my-key': 3.14})
>>> ad.update(new_key=42)
>>> ad.key1 = 2
>>> ad
"key1": 2
"key2": abc
"my-key": 3.14
"new_key": 42
"""
def _lightning_get_all_attr_holders(model: "pl.LightningModule", attribute: str) -> List[Any]:
"""Special attribute finding for Lightning.
Gets all of the objects or dicts that holds attribute. Checks for attribute in model namespace, the old hparams
namespace/dict, and the datamodule.
"""
holders: List[Any] = []
# Check if attribute in model
if hasattr(model, attribute):
holders.append(model)
# Check if attribute in model.hparams, either namespace or dict
if hasattr(model, "hparams") and attribute in model.hparams:
holders.append(model.hparams)
trainer = model._trainer
# Check if the attribute in datamodule (datamodule gets registered in Trainer)
if trainer is not None and trainer.datamodule is not None:
if hasattr(trainer.datamodule, attribute):
holders.append(trainer.datamodule)
if hasattr(trainer.datamodule, "hparams") and attribute in trainer.datamodule.hparams:
holders.append(trainer.datamodule.hparams)
return holders
def _lightning_get_first_attr_holder(model: "pl.LightningModule", attribute: str) -> Optional[Any]:
"""Special attribute finding for Lightning.
Gets the object or dict that holds attribute, or None. Checks for attribute in model namespace, the old hparams
namespace/dict, and the datamodule, returns the last one that has it.
"""
holders = _lightning_get_all_attr_holders(model, attribute)
if len(holders) == 0:
return None
# using the last holder to preserve backwards compatibility
return holders[-1]
def lightning_hasattr(model: "pl.LightningModule", attribute: str) -> bool:
"""Special hasattr for Lightning.
Checks for attribute in model namespace, the old hparams namespace/dict, and the datamodule.
"""
return _lightning_get_first_attr_holder(model, attribute) is not None
def lightning_getattr(model: "pl.LightningModule", attribute: str) -> Optional[Any]:
"""Special getattr for Lightning. Checks for attribute in model namespace, the old hparams namespace/dict, and the
datamodule.
Raises:
AttributeError:
If ``model`` doesn't have ``attribute`` in any of
model namespace, the hparams namespace/dict, and the datamodule.
"""
holder = _lightning_get_first_attr_holder(model, attribute)
if holder is None:
raise AttributeError(
f"{attribute} is neither stored in the model namespace"
" nor the `hparams` namespace/dict, nor the datamodule."
)
if isinstance(holder, dict):
return holder[attribute]
return getattr(holder, attribute)
def lightning_setattr(model: "pl.LightningModule", attribute: str, value: Any) -> None:
"""Special setattr for Lightning. Checks for attribute in model namespace and the old hparams namespace/dict. Will
also set the attribute on datamodule, if it exists.
Raises:
AttributeError:
If ``model`` doesn't have ``attribute`` in any of
model namespace, the hparams namespace/dict, and the datamodule.
"""
holders = _lightning_get_all_attr_holders(model, attribute)
if len(holders) == 0:
raise AttributeError(
f"{attribute} is neither stored in the model namespace"
" nor the `hparams` namespace/dict, nor the datamodule."
)
for holder in holders:
if isinstance(holder, dict):
holder[attribute] = value
else:
setattr(holder, attribute, value)