-
Notifications
You must be signed in to change notification settings - Fork 3.3k
/
base.py
381 lines (309 loc) · 13 KB
/
base.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
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
import argparse
import functools
import operator
from abc import ABC, abstractmethod
from argparse import Namespace
from typing import Union, Optional, Dict, Iterable, Any, Callable, List, Sequence, Mapping, Tuple
import numpy as np
import torch
from pytorch_lightning.utilities import rank_zero_only
class LightningLoggerBase(ABC):
"""
Base class for experiment loggers.
Args:
agg_key_funcs:
Dictionary which maps a metric name to a function, which will
aggregate the metric values for the same steps.
agg_default_func:
Default function to aggregate metric values. If some metric name
is not presented in the `agg_key_funcs` dictionary, then the
`agg_default_func` will be used for aggregation.
Note:
The `agg_key_funcs` and `agg_default_func` arguments are used only when
one logs metrics with the :meth:`~LightningLoggerBase.agg_and_log_metrics` method.
"""
def __init__(
self,
agg_key_funcs: Optional[Mapping[str, Callable[[Sequence[float]], float]]] = None,
agg_default_func: Callable[[Sequence[float]], float] = np.mean
):
self._rank = 0
self._prev_step: int = -1
self._metrics_to_agg: List[Dict[str, float]] = []
self._agg_key_funcs = agg_key_funcs if agg_key_funcs else {}
self._agg_default_func = agg_default_func
def update_agg_funcs(
self,
agg_key_funcs: Optional[Mapping[str, Callable[[Sequence[float]], float]]] = None,
agg_default_func: Callable[[Sequence[float]], float] = np.mean
):
"""
Update aggregation methods.
Args:
agg_key_funcs:
Dictionary which maps a metric name to a function, which will
aggregate the metric values for the same steps.
agg_default_func:
Default function to aggregate metric values. If some metric name
is not presented in the `agg_key_funcs` dictionary, then the
`agg_default_func` will be used for aggregation.
"""
if agg_key_funcs:
self._agg_key_funcs.update(agg_key_funcs)
if agg_default_func:
self._agg_default_func = agg_default_func
@property
@abstractmethod
def experiment(self) -> Any:
"""Return the experiment object associated with this logger."""
def _aggregate_metrics(
self, metrics: Dict[str, float], step: Optional[int] = None
) -> Tuple[int, Optional[Dict[str, float]]]:
"""
Aggregates metrics.
Args:
metrics: Dictionary with metric names as keys and measured quantities as values
step: Step number at which the metrics should be recorded
Returns:
Step and aggregated metrics. The return value could be ``None``. In such case, metrics
are added to the aggregation list, but not aggregated yet.
"""
# if you still receiving metric from the same step, just accumulate it
if step == self._prev_step:
self._metrics_to_agg.append(metrics)
return step, None
# compute the metrics
agg_step, agg_mets = self._reduce_agg_metrics()
# as new step received reset accumulator
self._metrics_to_agg = [metrics]
self._prev_step = step
return agg_step, agg_mets
def _reduce_agg_metrics(self):
"""Aggregate accumulated metrics."""
# compute the metrics
if not self._metrics_to_agg:
agg_mets = None
elif len(self._metrics_to_agg) == 1:
agg_mets = self._metrics_to_agg[0]
else:
agg_mets = merge_dicts(self._metrics_to_agg, self._agg_key_funcs, self._agg_default_func)
return self._prev_step, agg_mets
def _finalize_agg_metrics(self):
"""This shall be called before save/close."""
agg_step, metrics_to_log = self._reduce_agg_metrics()
self._metrics_to_agg = []
if metrics_to_log is not None:
self.log_metrics(metrics=metrics_to_log, step=agg_step)
def agg_and_log_metrics(self, metrics: Dict[str, float], step: Optional[int] = None):
"""
Aggregates and records metrics.
This method doesn't log the passed metrics instantaneously, but instead
it aggregates them and logs only if metrics are ready to be logged.
Args:
metrics: Dictionary with metric names as keys and measured quantities as values
step: Step number at which the metrics should be recorded
"""
agg_step, metrics_to_log = self._aggregate_metrics(metrics=metrics, step=step)
if metrics_to_log:
self.log_metrics(metrics=metrics_to_log, step=agg_step)
@abstractmethod
def log_metrics(self, metrics: Dict[str, float], step: Optional[int] = None):
"""
Records metrics.
This method logs metrics as as soon as it received them. If you want to aggregate
metrics for one specific `step`, use the
:meth:`~pytorch_lightning.loggers.base.LightningLoggerBase.agg_and_log_metrics` method.
Args:
metrics: Dictionary with metric names as keys and measured quantities as values
step: Step number at which the metrics should be recorded
"""
pass
@staticmethod
def _convert_params(params: Union[Dict[str, Any], Namespace]) -> Dict[str, Any]:
# in case converting from namespace
if isinstance(params, Namespace):
params = vars(params)
if params is None:
params = {}
return params
@staticmethod
def _flatten_dict(params: Dict[str, Any], delimiter: str = '/') -> Dict[str, Any]:
"""
Flatten hierarchical dict, e.g. ``{'a': {'b': 'c'}} -> {'a/b': 'c'}``.
Args:
params: Dictionary containing the hyperparameters
delimiter: Delimiter to express the hierarchy. Defaults to ``'/'``.
Returns:
Flattened dict.
Examples:
>>> LightningLoggerBase._flatten_dict({'a': {'b': 'c'}})
{'a/b': 'c'}
>>> LightningLoggerBase._flatten_dict({'a': {'b': 123}})
{'a/b': 123}
"""
def _dict_generator(input_dict, prefixes=None):
prefixes = prefixes[:] if prefixes else []
if isinstance(input_dict, dict):
for key, value in input_dict.items():
if isinstance(value, (dict, Namespace)):
value = vars(value) if isinstance(value, Namespace) else value
for d in _dict_generator(value, prefixes + [key]):
yield d
else:
yield prefixes + [key, value if value is not None else str(None)]
else:
yield prefixes + [input_dict if input_dict is None else str(input_dict)]
return {delimiter.join(keys): val for *keys, val in _dict_generator(params)}
@staticmethod
def _sanitize_params(params: Dict[str, Any]) -> Dict[str, Any]:
"""
Returns params with non-primitvies converted to strings for logging.
>>> params = {"float": 0.3,
... "int": 1,
... "string": "abc",
... "bool": True,
... "list": [1, 2, 3],
... "namespace": Namespace(foo=3),
... "layer": torch.nn.BatchNorm1d}
>>> import pprint
>>> pprint.pprint(LightningLoggerBase._sanitize_params(params)) # doctest: +NORMALIZE_WHITESPACE
{'bool': True,
'float': 0.3,
'int': 1,
'layer': "<class 'torch.nn.modules.batchnorm.BatchNorm1d'>",
'list': '[1, 2, 3]',
'namespace': 'Namespace(foo=3)',
'string': 'abc'}
"""
return {k: v if type(v) in [bool, int, float, str, torch.Tensor] else str(v) for k, v in params.items()}
@abstractmethod
def log_hyperparams(self, params: argparse.Namespace):
"""
Record hyperparameters.
Args:
params: :class:`~argparse.Namespace` containing the hyperparameters
"""
def save(self) -> None:
"""Save log data."""
self._finalize_agg_metrics()
def finalize(self, status: str) -> None:
"""
Do any processing that is necessary to finalize an experiment.
Args:
status: Status that the experiment finished with (e.g. success, failed, aborted)
"""
self.save()
def close(self) -> None:
"""Do any cleanup that is necessary to close an experiment."""
self.save()
@property
@abstractmethod
def name(self) -> str:
"""Return the experiment name."""
@property
@abstractmethod
def version(self) -> Union[int, str]:
"""Return the experiment version."""
class LoggerCollection(LightningLoggerBase):
"""
The :class:`LoggerCollection` class is used to iterate all logging actions over
the given `logger_iterable`.
Args:
logger_iterable: An iterable collection of loggers
"""
def __init__(self, logger_iterable: Iterable[LightningLoggerBase]):
super().__init__()
self._logger_iterable = logger_iterable
def __getitem__(self, index: int) -> LightningLoggerBase:
return [logger for logger in self._logger_iterable][index]
@property
def experiment(self) -> List[Any]:
return [logger.experiment for logger in self._logger_iterable]
def log_metrics(self, metrics: Dict[str, float], step: Optional[int] = None) -> None:
[logger.log_metrics(metrics, step) for logger in self._logger_iterable]
def log_hyperparams(self, params: Union[Dict[str, Any], Namespace]) -> None:
[logger.log_hyperparams(params) for logger in self._logger_iterable]
def save(self) -> None:
[logger.save() for logger in self._logger_iterable]
def finalize(self, status: str) -> None:
[logger.finalize(status) for logger in self._logger_iterable]
def close(self) -> None:
[logger.close() for logger in self._logger_iterable]
@property
def name(self) -> str:
return '_'.join([str(logger.name) for logger in self._logger_iterable])
@property
def version(self) -> str:
return '_'.join([str(logger.version) for logger in self._logger_iterable])
class DummyExperiment(object):
""" Dummy experiment """
def nop(*args, **kw):
pass
def __getattr__(self, _):
return self.nop
class DummyLogger(LightningLoggerBase):
""" Dummy logger for internal use. Is usefull if we want to disable users
logger for a feature, but still secure that users code can run """
def __init__(self):
super().__init__()
self._experiment = DummyExperiment()
@property
def experiment(self):
return self._experiment
def log_metrics(self, metrics, step):
pass
def log_hyperparams(self, params):
pass
@property
def name(self):
pass
@property
def version(self):
pass
def merge_dicts(
dicts: Sequence[Mapping],
agg_key_funcs: Optional[Mapping[str, Callable[[Sequence[float]], float]]] = None,
default_func: Callable[[Sequence[float]], float] = np.mean
) -> Dict:
"""
Merge a sequence with dictionaries into one dictionary by aggregating the
same keys with some given function.
Args:
dicts:
Sequence of dictionaries to be merged.
agg_key_funcs:
Mapping from key name to function. This function will aggregate a
list of values, obtained from the same key of all dictionaries.
If some key has no specified aggregation function, the default one
will be used. Default is: ``None`` (all keys will be aggregated by the
default function).
default_func:
Default function to aggregate keys, which are not presented in the
`agg_key_funcs` map.
Returns:
Dictionary with merged values.
Examples:
>>> import pprint
>>> d1 = {'a': 1.7, 'b': 2.0, 'c': 1, 'd': {'d1': 1, 'd3': 3}}
>>> d2 = {'a': 1.1, 'b': 2.2, 'v': 1, 'd': {'d1': 2, 'd2': 3}}
>>> d3 = {'a': 1.1, 'v': 2.3, 'd': {'d3': 3, 'd4': {'d5': 1}}}
>>> dflt_func = min
>>> agg_funcs = {'a': np.mean, 'v': max, 'd': {'d1': sum}}
>>> pprint.pprint(merge_dicts([d1, d2, d3], agg_funcs, dflt_func))
{'a': 1.3,
'b': 2.0,
'c': 1,
'd': {'d1': 3, 'd2': 3, 'd3': 3, 'd4': {'d5': 1}},
'v': 2.3}
"""
agg_key_funcs = agg_key_funcs or dict()
keys = list(functools.reduce(operator.or_, [set(d.keys()) for d in dicts]))
d_out = {}
for k in keys:
fn = agg_key_funcs.get(k)
values_to_agg = [v for v in [d_in.get(k) for d_in in dicts] if v is not None]
if isinstance(values_to_agg[0], dict):
d_out[k] = merge_dicts(values_to_agg, fn, default_func)
else:
d_out[k] = (fn or default_func)(values_to_agg)
return d_out