/
u.py
994 lines (819 loc) · 29.9 KB
/
u.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
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
import subprocess, sys, os, re, tempfile, zipfile, gzip, io, shutil, string, random, itertools, pickle, json, yaml, gc, inspect
from itertools import chain, groupby, islice, product, permutations, combinations
from datetime import datetime
from time import time
from fnmatch import fnmatch
from glob import glob
from tqdm import tqdm
from copy import copy, deepcopy
from collections import OrderedDict, defaultdict, Counter
import warnings
warnings.filterwarnings('ignore')
version = sys.version_info
if version[0] < 3:
from StringIO import StringIO
else:
from io import StringIO
def lrange(*args, **kwargs):
return list(range(*args, **kwargs))
def lchain(*args):
return list(chain(*args))
def lmap(fn, *iterables):
return [fn(*xs) for xs in zip(*iterables)]
def lif(keep, *x):
return x if keep else []
def dif(keep, **kwargs):
return kwargs if keep else {}
def flatten(x):
return [z for y in x for z in y]
def groupby_(xs, key=None):
if callable(key):
key = map(key, xs)
elif key is None:
key = xs
groups = defaultdict(list)
for k, v in zip(key, xs):
groups[k].append(v)
return groups
class Dict(dict if version.major == 3 and version.minor >= 6 else OrderedDict):
def __add__(self, d):
return Dict(**self).merge(d)
def merge(self, *dicts, **kwargs):
for d in dicts:
self.update(d)
self.update(kwargs)
return self
def filter(self, keys):
try: # check for iterable
keys = set(keys)
return Dict((k, v) for k, v in self.items() if k in keys)
except TypeError: # function key
f = keys
return Dict((k, v) for k, v in self.items() if f(k, v))
def map(self, mapper):
if callable(mapper): # function mapper
return Dict((k, mapper(v)) for k, v in self.items())
else: # dictionary mapper
return Dict((k, mapper[v]) for k, v in self.items())
def load_json(path):
with open(path, 'r+') as f:
return json.load(f)
def save_json(path, dict_):
with open(path, 'w+') as f:
json.dump(dict_, f, indent=4, sort_keys=True)
def format_json(dict_):
return json.dumps(dict_, indent=4, sort_keys=True)
def format_yaml(dict_):
dict_ = recurse(dict_, lambda x: x._ if isinstance(x, Path) else dict(x) if isinstance(x, Dict) else x)
return yaml.dump(dict_)
def load_text(path, encoding='utf-8'):
with open(path, 'r', encoding=encoding) as f:
return f.read()
def save_text(path, string):
with open(path, 'w') as f:
f.write(string)
def load_pickle(path):
with open(path, 'rb') as f:
return pickle.load(f)
def save_pickle(path, obj):
with open(path, 'wb') as f:
pickle.dump(obj, f)
def wget(link, output_dir):
cmd = 'wget %s -P %s' % (link, output_dir)
shell(cmd)
output_path = Path(output_dir) / os.path.basename(link)
if not output_path.exists(): raise RuntimeError('Failed to run %s' % cmd)
return output_path
def extract(input_path, output_path=None):
if input_path[-3:] == '.gz':
if not output_path:
output_path = input_path[:-3]
with gzip.open(input_path, 'rb') as f_in:
with open(output_path, 'wb') as f_out:
f_out.write(f_in.read())
else:
raise RuntimeError('Don\'t know file extension for ' + input_path)
def rand_string(length):
import string
letters = string.ascii_lowercase
return ''.join(random.choice(letters) for i in range(length))
nexti = nextk = lambda iterable: next(iter(iterable))
nextv = lambda dict: next(iter(dict.values()))
nextkv = lambda dict: next(iter(dict.items()))
def shell(cmd, wait=True, stdout=subprocess.PIPE, stderr=subprocess.PIPE):
stdout = stdout or subprocess.DEVNULL
stderr = stderr or subprocess.DEVNULL
if not isinstance(cmd, str):
cmd = ' '.join(cmd)
process = subprocess.Popen(cmd, shell=True, stdout=stdout, stderr=stderr)
if not wait:
return process
out, err = process.communicate()
return out.decode().rstrip('\n') if out else '', err.decode().rstrip('\n') if err else ''
def terminal_height():
return int(shell('tput lines')[0])
def terminal_width():
return int(shell('tput cols')[0])
def git_state(dir=None):
cwd = os.getcwd()
dir = dir or shell('git rev-parse --show-toplevel')[0]
os.chdir(dir)
status = shell('git status')[0]
base_commit = shell('git rev-parse HEAD')[0]
diff = shell('git diff %s' % base_commit)[0]
os.chdir(cwd)
return base_commit, diff, status
def attrs(obj):
for k, v in inspect.getmembers(obj):
if inspect.isfunction(v) or inspect.ismethod(v):
print(f'{v.__name__}{inspect.signature(v)}')
elif not callable(v) and not k.startswith('__'):
print(k, v)
def source(obj):
print(inspect.getsource(obj))
def import_module(module_name, module_path):
import imp
module = imp.load_source(module_name, module_path)
return module
def str2num(s):
try: return int(s)
except:
try: return float(s)
except: return s
def get_time_log_path():
return datetime.now().isoformat().replace(':', '_').rsplit('.')[0] + '.log'
_log_path = None
def logger(directory=None):
global _log_path
if directory and not _log_path:
from datetime import datetime
_log_path = Path(directory) / get_time_log_path()
return log
def log(text):
print(text)
if _log_path:
with open(_log_path, 'a') as f:
f.write(text)
f.write('\n')
def installed(pkg):
out, err = shell('dpkg -l %s' % pkg)
if err and err.startswith('dpkg-query: no packages found matching'):
return False
return True
def install(pkgs, root):
root = Path(root)
old_cwd = os.getcwd()
self_installed = set()
os.chdir(root)
while pkgs:
pkg = pkgs.pop()
print('Processing %s' % pkg)
if installed(pkg) or pkg in self_installed:
continue
out, err = shell('apt-cache depends %s' % pkg)
deps = []
for x in out.split('\n'):
x = x.lstrip()
if x.startswith('Depends:'):
splits = x.split(' ')
assert len(splits) == 2
dep = splits[1]
if not (dep in self_installed or installed(dep)):
deps.append(dep)
print('Found needed dependencies %s for %s' % (deps, pkg))
pkgs.extend(deps)
tmp = Path('tmp')
shell('mkdir tmp && cd tmp && apt download %s' % pkg)
for deb in tmp.glob('*.deb'):
shell('dpkg -x %s .' % deb)
print('Installing %s with %s' % (pkg, deb))
self_installed.add(pkg)
tmp.rm()
lib = Path('usr/lib')
real_root = Path('/')
for x in lib, lib / 'x86_64-linux-gnu':
brokens = x.lslinks(exist=False)
for broken in brokens:
real = real_root / broken._up / os.readlink(broken)
if real.exists():
broken.link(real, force=True)
print('Fixing broken link to be %s -> %s' % (broken, real))
else:
print('Could not fix broken link %s' % broken)
os.chdir(old_cwd)
class Path(str):
""""""
@classmethod
def env(cls, var):
return Path(os.environ[var])
def __init__(self, path):
pass
def __add__(self, subpath):
return Path(str(self) + str(subpath))
def __truediv__(self, subpath):
return Path(os.path.join(str(self), str(subpath)))
def __floordiv__(self, subpath):
return (self / subpath)._
def ls(self, show_hidden=True, dir_only=False, file_only=False):
subpaths = [Path(self / subpath) for subpath in os.listdir(self) if show_hidden or not subpath.startswith('.')]
isdirs = [os.path.isdir(subpath) for subpath in subpaths]
subdirs = [subpath for subpath, isdir in zip(subpaths, isdirs) if isdir]
files = [subpath for subpath, isdir in zip(subpaths, isdirs) if not isdir]
if dir_only:
return subdirs
if file_only:
return files
return subdirs, files
def lsdirs(self, show_hidden=True):
return self.ls(show_hidden=show_hidden, dir_only=True)
def lsfiles(self, show_hidden=True):
return self.ls(show_hidden=show_hidden, file_only=True)
def lslinks(self, show_hidden=True, exist=None):
dirs, files = self.ls(show_hidden=show_hidden)
return [x for x in dirs + files if x.islink() and (
exist is None or not (exist ^ x.exists()))]
def glob(self, glob_str):
return [Path(p) for p in glob(self / glob_str, recursive=True)]
def re(self, re_pattern):
""" Similar to .glob but uses regex pattern """
subpatterns = lmap(re.compile, re_pattern.split('/'))
matches = []
dirs, files = self.ls()
for pattern in subpatterns[:-1]:
new_dirs, new_files = [], []
for d in filter(lambda x: pattern.fullmatch(x._name), dirs):
d_dirs, d_files = d.ls()
new_dirs.extend(d_dirs)
new_files.extend(d_files)
dirs, files = new_dirs, new_files
return sorted(filter(lambda x: subpatterns[-1].fullmatch(x._name), dirs + files))
def recurse(self, dir_fn=None, file_fn=None):
""" Recursively apply dir_fn and file_fn to all subdirs and files in directory """
if dir_fn is not None:
dir_fn(self)
dirs, files = self.ls()
if file_fn is not None:
list(map(file_fn, files))
for dir in dirs:
dir.recurse(dir_fn=dir_fn, file_fn=file_fn)
def mk(self):
os.makedirs(self, exist_ok=True)
return self
def dir_mk(self):
self._up.mk()
return self
def rm(self):
if self.isfile() or self.islink():
os.remove(self)
elif self.isdir():
shutil.rmtree(self)
return self
def unlink(self):
os.unlink(self)
return self
def mv(self, dest):
shutil.move(self, dest)
def mv_from(self, src):
shutil.move(src, self)
def cp(self, dest):
shutil.copy(self, dest)
def cp_from(self, src):
shutil.copy(src, self)
def link(self, target, force=False):
if self.lexists():
if not force:
return
else:
self.rm()
os.symlink(target, self)
def exists(self):
return os.path.exists(self)
def lexists(self):
return os.path.lexists(self)
def isfile(self):
return os.path.isfile(self)
def isdir(self):
return os.path.isdir(self)
def islink(self):
return os.path.islink(self)
def chdir(self):
os.chdir(self)
def rel(self, start=None):
return Path(os.path.relpath(self, start=start))
def clone(self):
name = self._name
match = re.search('__([0-9]+)$', name)
if match is None:
base = self + '__'
i = 1
else:
initial = match.group(1)
base = self[:-len(initial)]
i = int(initial) + 1
while True:
path = Path(base + str(i))
if not path.exists():
return path
i += 1
@property
def _(self):
return str(self)
@property
def _real(self):
return Path(os.path.realpath(os.path.expanduser(self)))
@property
def _up(self):
path = os.path.dirname(self.rstrip('/'))
if path == '':
path = os.path.dirname(self._real.rstrip('/'))
return Path(path)
@property
def _name(self):
return Path(os.path.basename(self))
@property
def _stem(self):
return Path(os.path.splitext(self)[0])
@property
def _basestem(self):
new = self._stem
while new != self:
new, self = new._stem, new
return new
@property
def _ext(self):
return Path(os.path.splitext(self)[1])
extract = extract
load_json = load_json
save_json = save_json
load_txt = load_sh = load_text
save_txt = save_sh = save_text
load_p = load_pickle
save_p = save_pickle
def save_bytes(self, bytes):
with open(self, 'wb') as f:
f.write(bytes)
def load_csv(self, index_col=0, **kwargs):
return pd.read_csv(self, index_col=index_col, **kwargs)
def save_csv(self, df, float_format='%.5g', **kwargs):
df.to_csv(self, float_format=float_format, **kwargs)
def load_npy(self):
return np.load(self, allow_pickle=True)
def save_npy(self, obj):
np.save(self, obj)
def load_yaml(self):
with open(self, 'r') as f:
return yaml.safe_load(f)
def save_yaml(self, obj):
obj = recurse(obj, lambda x: x._ if isinstance(x, Path) else dict(x) if isinstance(x, Dict) else x)
with open(self, 'w') as f:
yaml.dump(obj, f, default_flow_style=False, allow_unicode=True)
def load_pth(self):
return torch.load(self)
def save_pth(self, obj):
torch.save(obj, self)
def load_pdf(self):
"""
return: PdfReader object.
Can use index and slice obj.pages for the pages, then call Path.save_pdf to save
"""
from pdfrw import PdfReader
return PdfReader(self)
def save_pdf(self, pages):
from pdfrw import PdfWriter
writer = PdfWriter()
writer.addpages(pages)
writer.write(self)
def load(self):
return eval('self.load_%s' % self._ext[1:])()
def save(self, obj):
return eval('self.save_%s' % self._ext[1:])(obj)
def replace_txt(self, replacements, dst=None):
content = self.load_txt()
for k, v in replacements.items():
content = content.replace(k, v)
(dst or self).save_txt(content)
def update_dict(self, updates={}, vars=[], unvars=[], dst=None):
d = self.load()
for k in vars:
d[k] = True
for k in unvars:
d.pop(k, None)
d.update(updates)
(dst or self).save(d)
def torch_strip(self, dst):
self.update_dict(unvars=['opt', 'step'], dst=dst)
def wget(self, link):
if self.isdir():
return Path(wget(link, self))
raise ValueError('Path %s needs to be a directory' % self)
def replace(self, old, new=''):
return Path(super().replace(old, new))
def search(self, pattern):
return re.search(pattern, self)
def search_pattern(self, pattern):
return self.search(pattern).group()
def search_groups(self, pattern):
return self.search(pattern).groups()
def search_group(self, pattern):
return self.search_groups(pattern)[0]
def findall(self, pattern):
return re.findall(pattern, self)
class Namespace(Dict):
def __init__(self, *args, **kwargs):
self.var(*args, **kwargs)
def var(self, *args, **kwargs):
kvs = Dict()
for a in args:
if isinstance(a, str):
kvs[a] = True
else: # a is a dictionary
kvs.update(a)
kvs.update(kwargs)
self.update(kvs)
return self
def unvar(self, *args):
for a in args:
self.pop(a)
return self
def setdefaults(self, *args, **kwargs):
args = [a for a in args if a not in self]
kwargs = {k: v for k, v in kwargs.items() if k not in self}
return self.var(*args, **kwargs)
def __getattr__(self, key):
try:
return self[key]
except KeyError as e:
self.__getattribute__(key)
def __setattr__(self, key, value):
self[key] = value
##### Functions for compute
using_ipython = True
try:
_ = get_ipython().__class__.__name__
except NameError:
using_ipython = False
import numpy as np
import pandas as pd
def _sel(self, col, value):
if isinstance(value, list):
return self[self[col].isin(value)]
return self[self[col] == value]
pd.DataFrame.sel = _sel
import scipy.stats
import scipy as sp
from scipy.stats import pearsonr as pearson, spearmanr as spearman, kendalltau
if not using_ipython:
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
plt_colors = plt.rcParams['axes.prop_cycle'].by_key()['color']
arrayf = lambda *args, **kwargs: np.array(*args, **kwargs, dtype=np.float32)
arrayl = lambda *args, **kwargs: np.array(*args, **kwargs, dtype=np.long)
arrayb = lambda *args, **kwargs: np.array(*args, **kwargs, dtype=np.bool)
arrayo = lambda *args, **kwargs: np.array(*args, **kwargs, dtype=object)
def split(x, sizes):
return np.split(x, np.cumsum(sizes[:-1]))
def recurse(x, fn):
if isinstance(x, dict):
return type(x)((k, recurse(v, fn)) for k, v in x.items())
elif isinstance(x, (list, tuple)):
return type(x)(recurse(v, fn) for v in x)
return fn(x)
def from_numpy(x):
def helper(x):
if type(x).__module__ == np.__name__:
if isinstance(x, np.ndarray):
return recurse(list(x), helper)
return np.asscalar(x)
return x
return recurse(x, helper)
def smooth(y, box_pts):
box = np.ones(box_pts) / box_pts
y_smooth = np.convolve(y, box, mode='same')
return y_smooth
def gsmooth(y, sigma):
from scipy.ndimage.filters import gaussian_filter1d
return gaussian_filter1d(y, sigma=sigma)
def normalize(x, eps=1e-8):
return (x - x.mean()) / x.std()
def inverse_map(arr):
inv_map = np.zeros(len(arr))
inv_map[arr] = np.arange(len(arr))
return inv_map
def pad_arrays(arrs, value):
max_len = max(len(x) for x in arrs)
return np.array([np.concatenate([x, np.full(max_len - len(x), value)]) for x in arrs])
def sorted_segment_maps(segments):
r = Namespace()
r.segment_idxs = np.argsort(segments)
starts = np.cumsum(segments) - segments
starts = starts[r.segment_idxs]
r.segments = segments[r.segment_idxs]
r.unit_idxs = np.array([i for s, length in zip(starts, r.segments) for i in range(s, s + length)])
r.unit_idxs_r = inverse_map(r.unit_idxs)
r.segment_uniques, r.segment_blocks, r.segment_counts = zip(*((seg, sum(segs), len(list(segs))) for seg, segs in groupby(r.segments)))
return r
def get_gpu_info(ssh_fn=lambda x: x):
nvidia_str, _ = shell(ssh_fn('nvidia-smi --query-gpu=index,name,memory.used,memory.total,utilization.gpu --format=csv,nounits'))
nvidia_str = nvidia_str.replace('[Not Supported]', '100').replace(', ', ',')
nvidia_str_io = StringIO(nvidia_str)
gpu_df = pd.read_csv(nvidia_str_io, index_col=0)
devices_str = os.environ.get('CUDA_VISIBLE_DEVICES')
if devices_str:
devices = list(map(int, devices_str.split(',')))
gpu_df = gpu_df.loc[devices]
gpu_df.index = gpu_df.index.map({k: i for i, k in enumerate(devices)})
out_df = pd.DataFrame(index=gpu_df.index)
out_df['memory_total'] = gpu_df['memory.total [MiB]']
out_df['memory_used'] = gpu_df['memory.used [MiB]']
out_df['memory_free'] = out_df['memory_total'] - out_df['memory_used']
out_df['utilization'] = gpu_df['utilization.gpu [%]'] / 100
out_df['utilization_free'] = 1 - out_df['utilization']
return out_df
def get_process_gpu_info(pid=None, ssh_fn=lambda x: x):
nvidia_str, _ = shell(ssh_fn('nvidia-smi --query-compute-apps=pid,gpu_name,used_gpu_memory --format=csv,nounits'))
nvidia_str_io = StringIO(nvidia_str.replace(', ', ','))
gpu_df = pd.read_csv(nvidia_str_io, index_col=0)
if pid is None:
return gpu_df
if pid == -1:
pid = os.getpid()
return gpu_df.loc[pid]
##### torch functions
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
def to_torch(x, device='cuda' if torch.cuda.is_available() else 'cpu', **kwargs):
def helper(x):
if x is None:
return None
elif isinstance(x, torch.Tensor):
return x.to(device=device, **kwargs)
elif np.isscalar(x):
return x
return torch.from_numpy(x).to(device=device, **kwargs)
return recurse(x, helper)
def from_torch(t, force_scalar=False):
def helper(t):
if not isinstance(t, torch.Tensor):
return t
x = t.detach().cpu().numpy()
if force_scalar and (x.size == 1 or np.isscalar(x)):
return np.asscalar(x)
return x
return recurse(t, helper)
def count_params(network, requires_grad=False):
return sum(p.numel() for p in network.parameters() if not requires_grad or p.requires_grad)
def report_memory(device=None, max=False):
if device:
device = torch.device(device)
if max:
alloc = torch.cuda.max_memory_allocated(device=device)
else:
alloc = torch.cuda.memory_allocated(device=device)
alloc /= 1024 ** 2
print('%.3f MBs' % alloc)
return alloc
numels = Counter()
for obj in gc.get_objects():
if torch.is_tensor(obj):
print(type(obj), obj.size())
numels[obj.device] += obj.numel()
print()
for device, numel in sorted(numels.items()):
print('%s: %s elements, %.3f MBs' % (str(device), numel, numel * 4 / 1024 ** 2))
def clear_gpu_memory():
gc.collect()
torch.cuda.empty_cache()
for obj in gc.get_objects():
if torch.is_tensor(obj):
obj.cpu()
gc.collect()
torch.cuda.empty_cache()
def main_only(method):
def wrapper(self, *args, **kwargs):
if self.main:
return method(self, *args, **kwargs)
return wrapper
class Config(Namespace):
def __init__(self, res, *args, **kwargs):
self.res = Path(res)._real
super(Config, self).__init__()
self.load()
self.var(*args, **kwargs)
self.setdefaults(
name=self.res._real._name,
main=True,
logger=True,
device='cuda' if torch.cuda.is_available() else 'cpu',
debug=False,
opt_level='O0',
disable_amp=False
)
def __repr__(self):
return format_yaml(dict(self))
def __hash__(self):
return hash(repr(self))
@property
def path(self):
return self.res / 'config.yaml'
def load(self):
if self.path.exists():
for k, v in self.path.load().items():
self[k] = v
return self
never_save = {'res', 'name', 'main', 'logger', 'distributed', 'parallel', 'device', 'steps', 'debug'}
@property
def attrs_save(self):
return {k: v for k, v in self.items() if k not in self.never_save}
def save(self, force=False):
if force or not self.path.exists():
self.res.mk()
self.path.save(from_numpy(self.attrs_save))
return self
@classmethod
def from_args(cls, *globals_locals):
import argparse
parser = argparse.ArgumentParser(description='Model arguments')
parser.add_argument('res', type=Path, help='Result directory')
parser.add_argument('kwargs', nargs='*', help='Extra arguments that goes into the config')
args = parser.parse_args()
kwargs = {}
for kv in args.kwargs:
splits = kv.split('=')
if len(splits) == 1:
v = True
else:
v = splits[1]
try:
v = eval(v, *globals_locals)
except (SyntaxError, NameError):
pass
kwargs[splits[0]] = v
return cls(args.res, **kwargs).save()
def try_save_commit(self, base_dir=None):
base_commit, diff, status = git_state(base_dir)
save_dir = (self.res / 'commit').mk()
(save_dir / 'hash.txt').save(base_commit)
(save_dir / 'diff.txt').save(diff)
(save_dir / 'status.txt').save(status)
return self
@main_only
def log(self, text):
logger(self.res if self.logger else None)(text)
### Train result saving ###
@property
def train_results(self):
return self.res / 'train_results.csv'
def load_train_results(self):
if self.train_results.exists():
return pd.read_csv(self.train_results, index_col=0)
return None
@main_only
def save_train_results(self, results):
results.to_csv(self.train_results, float_format='%.6g')
### Set stopped early ###
@property
def stopped_early(self):
return self.res / 'stopped_early'
@main_only
def set_stopped_early(self):
self.stopped_early.save_txt('')
### Set training state ###
@property
def training(self):
return self.res / 'is_training'
@main_only
def set_training(self, is_training):
if is_training:
if self.main and self.training.exists():
self.log('Another training is found, continue (yes/n)?')
ans = input('> ')
if ans != 'yes':
exit()
self.training.save_txt('')
else:
self.training.rm()
### Model loading ###
def init_model(self, net, opt=None, step='max', train=True):
if train:
assert not self.training.exists(), 'Training already exists'
# configure parallel training
devices = os.environ.get('CUDA_VISIBLE_DEVICES')
self.n_gpus = 0 if self.device == 'cpu' else 1 if self.device.startswith('cuda:') else len(get_gpu_info()) if devices is None else len(devices.split(','))
can_parallel = self.n_gpus > 1
self.setdefaults(distributed=can_parallel) # use distributeddataparallel
self.setdefaults(parallel=can_parallel and not self.distributed) # use dataparallel
self.local_rank = 0
self.world_size = 1 # number of processes
if self.distributed:
self.local_rank = int(os.environ['LOCAL_RANK']) # rank of the current process
self.world_size = int(os.environ['WORLD_SIZE'])
assert self.world_size == self.n_gpus
torch.cuda.set_device(self.local_rank)
torch.distributed.init_process_group(backend='nccl', init_method='env://')
self.main = self.local_rank == 0
net.to(self.device)
if train and not self.disable_amp:
# configure mixed precision
net, opt = amp.initialize(net, opt, opt_level=self.opt_level, loss_scale=self.get('loss_scale'), verbosity=0 if self.opt_level == 'O0' else 1)
step = self.set_state(net, opt=opt, step=step)
if self.distributed:
import apex
net = apex.parallel.DistributedDataParallel(net)
elif self.parallel:
net = nn.DataParallel(net)
if train:
net.train()
return net, opt, step
else:
net.eval()
return net, step
def load_model(self, step='best', train=False):
'''
step can be 'best', 'max', an integer, or None
'''
model = import_module('model', str(self.model))
net = model.get_net(self)
opt = model.get_opt(self, net) if train else None
return self.init_model(net, opt=opt, step=step, train=train)
@property
def models(self):
return (self.res / 'models').mk()
def model_save(self, step):
return self.models / ('model-%s.pth' % step)
def model_step(self, path):
m = re.match('.+/model-(\d+)\.pth', path)
if m:
return int(m.groups()[0])
@property
def model_best(self):
return self.models / 'best_model.pth'
@main_only
def link_model_best(self, model_save):
self.model_best.rm().link(Path(model_save).rel(self.models))
def get_saved_model_steps(self):
_, save_paths = self.models.ls()
if len(save_paths) == 0:
return []
return sorted([x for x in map(self.model_step, save_paths) if x is not None])
def set_state(self, net, opt=None, step='max', path=None):
state = self.load_state(step=step, path=path)
if state is None:
return 0
if self.get('append_module_before_load'):
state['net'] = OrderedDict(('module.' + k, v) for k, v in state['net'].items())
net.load_state_dict(state['net'])
if opt:
if 'opt' in state:
opt.load_state_dict(state['opt'])
else:
self.log('No state for optimizer to load')
if 'amp' in state and self.opt_level != 'O0':
amp.load_state_dict(state['amp'])
return state.get('step', 0)
@main_only
def get_state(self, net, opt, step):
try:
net_dict = net.module.state_dict()
except AttributeError:
net_dict = net.state_dict()
state = dict(step=step, net=net_dict, opt=opt.state_dict())
try:
state['amp'] = amp.state_dict()
except:
pass
return to_torch(state, device='cpu')
def load_state(self, step='max', path=None):
'''
step: best, max, integer, None if path is specified
path: None if step is specified
'''
if path is None:
if step == 'best':
path = self.model_best
else:
if step == 'max':
steps = self.get_saved_model_steps()
if len(steps) == 0:
return None
step = max(steps)
path = self.model_save(step)
save_path = Path(path)
if save_path.exists():
return to_torch(torch.load(save_path), device=self.device)
return None
@main_only
def save_state(self, step, state, clean=True, link_best=False):
save_path = self.model_save(step)
if save_path.exists():
return save_path
torch.save(state, save_path)
self.log('Saved model %s at step %s' % (save_path, step))
if clean and self.get('max_save'):
self.clean_models(keep=self.max_save)
if link_best:
self.link_model_best(save_path)
self.log('Linked %s to new saved model %s' % (self.model_best, save_path))
return save_path