-
Notifications
You must be signed in to change notification settings - Fork 100
/
easy_train.py
2629 lines (2274 loc) · 90.5 KB
/
easy_train.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
995
996
997
998
999
1000
#!/usr/bin/env python3
import time
import random
import sys
import psutil
import re
import subprocess
import importlib
import importlib.metadata
import argparse
import math
import logging
import time
EXITCODE_OK = 0
EXITCODE_MISSING_DEPENDENCIES = 2
EXITCODE_TRAINING_LIKELY_NOT_FINISHED = 3
EXITCODE_TRAINING_NOT_FINISHED = 4
logging.basicConfig()
LOGGER = logging.getLogger(__name__)
LOGGER.setLevel(logging.DEBUG)
LOGGER.addHandler(logging.StreamHandler(stream=sys.stdout))
LOGGER.propagate = False
def validate_python_version():
if sys.version_info >= (3, 7):
LOGGER.info(f'Found python version {sys.version}. OK.')
return True
else:
LOGGER.error(f'Found python version {sys.version} but 3.7 is required. Exiting.')
return False
# Functions for checking external dependencies.
def run_for_version(name):
process = subprocess.Popen(
[name, '--version'],
shell=False,
bufsize=-1,
universal_newlines=True,
stdin=subprocess.DEVNULL,
stdout=subprocess.PIPE,
stderr=subprocess.STDOUT
)
return process.stdout.read()
def validate_cmake():
success = True
try:
out = run_for_version('cmake')
parts = out.split('\n')[0].split()
version_str = parts[-1]
major_version = int(version_str.split('.')[0])
minor_version = int(version_str.split('.')[1])
success = (major_version, minor_version) >= (3, 4)
if success:
LOGGER.info(f'Found cmake executable version {version_str}. OK.')
else:
LOGGER.error(f'Found cmake executable version {version_str} but at least 3.4 required. Exiting.')
except:
success = False
LOGGER.error('No cmake executable found. Exiting.')
return success
def validate_make():
success = True
try:
out = run_for_version('make')
parts = out.split('\n')[0].split()
version_str = parts[-1]
major_version = int(version_str.split('.')[0])
success = major_version >= 3
if success:
LOGGER.info(f'Found make executable version {version_str}. OK.')
else:
LOGGER.error(f'Found make executable version {version_str} but at least 3 required. Exiting.')
except:
success = False
LOGGER.error('No make executable found. Exiting.')
return success
def validate_gcc():
success = True
try:
out = run_for_version('gcc')
parts = out.split('\n')[0].split()
for part in parts:
try:
version_str = part # sometimes there are trailing strings in the version number
major_version = int(version_str.split('.')[0])
minor_version = int(version_str.split('.')[1])
success = (major_version, minor_version) >= (9, 2)
except:
continue
if success:
LOGGER.info(f'Found gcc executable version {version_str}. OK.')
else:
LOGGER.error(f'Found gcc executable version {version_str} but at least 9.2 required. Exiting.')
except:
success = False
LOGGER.error('No gcc executable found. Exiting.')
return success
def maybe_int(v):
try:
return int(v)
except:
return v
class PackageInfo:
'''
Represents an [installed] python package.
'''
def __init__(self, name):
self._spec = importlib.util.find_spec(name)
self._version_str = None
self._version_tup = None
try:
if self._spec:
self._version_str = importlib.metadata.version(name)
self._version_tup = tuple(maybe_int(v) for v in self._version_str.split('.'))
except:
pass
@property
def exists(self):
return self._spec is not None
def is_version_at_least(self, desired):
return self._version_tup and self._version_tup >= desired
@property
def version(self):
return self._version_str
# Functions for checking required python packages.
def validate_asciimatics():
pkg = PackageInfo('asciimatics')
if pkg.exists:
LOGGER.info('Found asciimatics package. OK.')
return True
else:
LOGGER.error('No asciimatics package found. Run `pip install asciimatics`. Exiting.')
return False
def validate_pytorch():
pkg = PackageInfo('torch')
if pkg.exists:
if not 'cu' in pkg.version:
LOGGER.error(f'Found torch without CUDA but CUDA support required. Exiting')
return False
elif pkg.is_version_at_least((1, 7)):
LOGGER.info(f'Found torch version {pkg.version}. OK.')
return True
else:
LOGGER.error(f'Found torch version {pkg.version} but at least 1.8 required. Exiting.')
return False
else:
LOGGER.error('No torch package found. Install at least torch 1.8 with cuda. See https://pytorch.org/. Exiting.')
return False
def validate_pytorchlightning():
pkg = PackageInfo('pytorch_lightning')
if pkg.exists:
LOGGER.info(f'Found pytorch_lightning version {pkg.version}. OK.')
return True
else:
LOGGER.error('No pytorch_lightning found. Run `pip install pytorch-lightning`. Exiting.')
return False
def validate_cupy():
pkg = PackageInfo('cupy')
if pkg.exists:
LOGGER.info(f'Found cupy version {pkg.version}. OK.')
return True
else:
LOGGER.error('No cupy found. Install cupy matching cuda version used by pytorch. See https://cupy.dev/. Exiting.')
return False
def validate_gputil():
pkg = PackageInfo('GPUtil')
if pkg.exists:
LOGGER.info(f'Found GPUtil version {pkg.version}. OK.')
return True
else:
LOGGER.error('No GPUtil found. Run `pip install GPUtil`. Exiting.')
return False
# Validation of required external and package dependencies.
def validate_imports():
success = True
success &= validate_asciimatics()
success &= validate_pytorch()
success &= validate_pytorchlightning()
success &= validate_cupy()
success &= validate_gputil()
return success
def validate_environment_requirements():
success = True
try:
success &= validate_python_version()
success &= validate_make()
success &= validate_cmake()
success &= validate_gcc()
success &= validate_imports()
except Exception as e:
LOGGER.error(e)
return False
return success
# Exit early if the requires packages have not been found
if not validate_environment_requirements():
sys.exit(EXITCODE_MISSING_DEPENDENCIES)
# Only now import the rest of the required packages
from asciimatics.widgets import Frame, ListBox, Layout, Divider, Text, Button, \
TextBox, Widget, VerticalDivider, MultiColumnListBox, Label, PopUpDialog
from asciimatics.scene import Scene
from asciimatics.screen import Screen
from asciimatics.exceptions import ResizeScreenError, NextScene, StopApplication
from asciimatics.utilities import BoxTool
from asciimatics.constants import SINGLE_LINE, DOUBLE_LINE
from asciimatics.event import KeyboardEvent, MouseEvent
from threading import Thread, Lock, Event
import GPUtil
import io
import os
import requests
import zipfile
import shutil
import urllib.request
import urllib.parse
import signal
from datetime import datetime, timedelta
from tqdm.auto import tqdm
from pathlib import Path
# Specify which versions of ordo and c-chess-cli we want.
# We rely on specific well-tested commits because we know exactly what we need.
# repo/branch, commit id
ORDO_GIT = ('michiguel/Ordo', '17eec774f2e4b9fdd2b1b38739f55ea221fb851a')
C_CHESS_CLI_GIT = ('lucasart/c-chess-cli', '6d08fee2e95b259c486b21a886f6911b61f676af')
TIMEOUT = 600.0 # on some systems starting pytorch can be really slow
def terminate_process_on_exit(process):
'''
Create a watchdog process that awaits the termination of this (calling) process
and automatically terminates a given process (python's subprocess object) after.
On Windows this is achieved by a wmic call that is deprecated in windows 10,
and may not work in windows 11.
See: https://stackoverflow.com/a/22559493/3763139
https://superuser.com/a/1299350/388191
TODO: powershell version
TODO: linux version
'''
if sys.platform == "win32":
try:
# We cannot execute from string so we write the script to a file.
# Doesn't do anything if the file already exists.
with open('.process_watchdog_helper.bat', 'x') as file:
file.write(""":waitforpid
tasklist /nh /fi "pid eq %1" 2>nul | find "%1" >nul
if %ERRORLEVEL%==0 (
timeout /t 5 /nobreak >nul
goto :waitforpid
) else (
wmic process where processid="%2" call terminate >nul
)""")
except:
pass
subprocess.Popen(
['.process_watchdog_helper.bat', str(os.getpid()), str(process.pid)],
stdout=subprocess.DEVNULL,
stderr=subprocess.DEVNULL
)
elif sys.platform == "linux":
# TODO: this
pass
# Exits the process forcefully after a specified amount of seconds with a given error code
TUI_SCREEN = None
def schedule_exit(timeout_seconds, errcode):
def f():
time.sleep(timeout_seconds)
LOGGER.info(f'Performing a scheduled exit.')
if TUI_SCREEN:
if sys.platform == 'win32':
TUI_SCREEN.close(restore=True)
else:
# We cannot call .close directly because it tries to reset signals...
# But resetting signals won't work from a non-main thread...
import curses
TUI_SCREEN._screen.keypad(0)
curses.echo()
curses.nocbreak()
curses.endwin()
os._exit(errcode)
thread = Thread(target=f)
thread.setDaemon(True)
thread.start()
if sys.platform == "win32":
import ctypes
WINAPI_CreateMutex = ctypes.windll.kernel32.CreateMutexA
WINAPI_CreateMutex.argtypes = [ctypes.wintypes.LPCVOID, ctypes.wintypes.BOOL, ctypes.c_char_p]
WINAPI_CreateMutex.restype = ctypes.wintypes.HANDLE
WINAPI_WaitForSingleObject = ctypes.windll.kernel32.WaitForSingleObject
WINAPI_WaitForSingleObject.argtypes = [ctypes.wintypes.HANDLE, ctypes.wintypes.DWORD]
WINAPI_WaitForSingleObject.restype = ctypes.wintypes.DWORD
WINAPI_ReleaseMutex = ctypes.windll.kernel32.ReleaseMutex
WINAPI_ReleaseMutex.argtypes = [ctypes.wintypes.HANDLE]
WINAPI_ReleaseMutex.restype = ctypes.wintypes.BOOL
WINAPI_CloseHandle = ctypes.windll.kernel32.CloseHandle
WINAPI_CloseHandle.argtypes = [ctypes.wintypes.HANDLE]
WINAPI_CloseHandle.restype = ctypes.wintypes.BOOL
class SystemWideMutex:
def __init__(self, name):
# \ is a reserved character so we have to convert them to / to be recognized as
# directory delimiters
# encode as utf-8 because LPCSTR is bytes not str
self.name = str(os.path.abspath(name)).replace('\\', '/').encode('utf-8')
self.acquired = False
self.file = open(self.name, 'a+')
self.handle = WINAPI_CreateMutex(None, False, self.name)
if not self.handle:
raise ctypes.WinError()
def acquire(self):
ret = WINAPI_WaitForSingleObject(self.handle, 0xFFFFFFFF)
if ret in (0, 0x80):
# 0 - normally acquired
# 0x80 - acquired by other process terminating
self.acquired = True
return True
else:
raise ctypes.WinError()
def release(self):
ret = WINAPI_ReleaseMutex(self.handle)
if not ret:
raise ctypes.WinError()
self.acquired = False
def close(self):
if self.handle is None:
return
self.file.close()
ret = WINAPI_CloseHandle(self.handle)
if not ret:
raise ctypes.WinError()
try:
os.remove(self.name)
except:
pass
self.handle = None
def __enter__(self):
self.acquire()
return self
def __exit__(self, exc_type, exc_val, exc_tb):
self.release()
self.close()
else:
import fcntl
class SystemWideMutex:
def __init__(self, name):
self.name = name
self.acquired = False
self.file = open(self.name, 'a+')
def acquire(self):
fcntl.lockf(self.file, fcntl.LOCK_EX)
self.acquired = True
def release(self):
fcntl.lockf(self.file, fcntl.LOCK_UN)
self.acquired = False
def close(self):
if self.file is None:
return
os.unlink(self.name)
self.file.close()
self.file = None
def __del__(self):
self.close()
def __enter__(self):
self.acquire()
return self
def __exit__(self, exc_type, exc_val, exc_tb):
self.release()
class DecayingRunningAverage:
'''
Represents an average of a list of values with exponential decay of old values.
Every added value has weight of decay**n, where n is the distance from the
last element. For the last added element n==0.
'''
def __init__(self, decay=0.995):
self._decay = decay
self._total = 0.0
self._count = 0.0
@property
def decay(self):
return self._decay
@property
def value(self):
try:
return self._total / self._count
except:
return float('NaN')
def update(self, value):
'''
Adds a new value at the end of the implicit running average list and
updates the counters to reflect the change in the running average.
'''
self._total = self._total * self._decay + value
self._count = self._count * self._decay + 1.0
class SystemResources:
'''
Holds information about the usage of system resources at a time point of creation.
This includes GPU, CPU, and memory usage.
'''
def __init__(self):
self._gpus = dict()
for gpu in GPUtil.getGPUs():
self._gpus[gpu.id] = gpu
self._cpu_usage = psutil.cpu_percent() / 100.0
mem = psutil.virtual_memory()
self._ram_usage_mb = mem[3] // (1024 * 1024)
self._ram_max_mb = mem[0] // (1024 * 1024)
@property
def gpus(self):
return self._gpus
@property
def cpu_usage(self):
return self._cpu_usage
@property
def ram_usage_mb(self):
return self._ram_usage_mb
@property
def ram_max_mb(self):
return self._ram_max_mb
class SystemResourcesMonitor(Thread):
'''
Periodically queries system resources.
Runs as a daemon so does not need to be cleaned up.
'''
def __init__(self, period_seconds):
super(SystemResourcesMonitor, self).__init__()
self._period_seconds = period_seconds
self._mutex = Lock()
self._stop_event = Event()
self._running = True
self._update()
self.setDaemon(True)
self.start()
def _update(self):
self._resources = SystemResources()
def run(self):
while self._running:
self._mutex.acquire()
try:
self._update()
finally:
self._mutex.release()
self._stop_event.wait(timeout=self._period_seconds)
@property
def resources(self):
'''
Returns the most recent system resources measurement.
'''
self._mutex.acquire()
try:
return self._resources
finally:
self._mutex.release()
def stop(self):
self._running = False
self._stop_event.set()
def find_latest_checkpoint(root_dir):
'''
Recursively searches the specified directory for
the .ckpt file with the latest creation date.
'''
ckpts = [file for file in Path(root_dir).rglob("*.ckpt")]
if not ckpts:
return None
return str(max(ckpts, key=lambda p: p.stat().st_ctime_ns))
class OrdoEntry:
'''
Represents a single entry in an ordo file.
Expects players to be named after network paths, if the form experiment_path/run_{}/nn-epoch{}.nnue
'''
NET_PATTERN = re.compile(r'.*?run_(\d+).*?nn-epoch(\d+)\.nnue')
def __init__(self, line=None, network_path=None, elo=None, elo_error=None, run_id=None, epoch=None):
if line:
fields = line.split()
self._network_path = fields[1]
self._elo = float(fields[3])
self._elo_error = float(fields[4])
net_parts = OrdoEntry.NET_PATTERN.search(self._network_path)
self._run_id = int(net_parts[1])
self._epoch = int(net_parts[2])
else:
self._network_path = network_path
self._elo = elo
self._elo_error = elo_error
self._run_id = run_id
self._epoch = epoch
@property
def network_path(self):
return self._network_path
@property
def run_id(self):
return self._run_id
@property
def epoch(self):
return self._epoch
@property
def elo(self):
return self._elo
@property
def elo_error(self):
return self._elo_error
def find_best_checkpoint(root_dir):
'''
Recursively searches the specified directory the best
.ckpt file as determined by an ordo output file that must be
present under the path os.path.join(root_dir, 'ordo.out').
The path to the checkpoint must have 'nn-epoch' in it,
other checkpoints are not considered.
Returns None if the ordo file does not exist or
no suitable checkpoint has been found.
'''
ckpts = [str(file) for file in Path(root_dir).rglob("*.ckpt")]
nnues = [str(file) for file in Path(root_dir).rglob("*.nnue")]
ordo_file_path = os.path.join(root_dir, 'ordo.out')
with open(ordo_file_path, 'r') as ordo_file:
entries = []
lines = ordo_file.readlines()
for line in lines:
if 'nn-epoch' in line:
try:
entries.append(OrdoEntry(line=line))
except:
pass
entries.sort(key=lambda x:-x.elo+x.elo_error)
run_id = entries[0].run_id
epoch = entries[0].epoch
for ckpt in ckpts:
if f'run_{run_id}' in ckpt and f'epoch={epoch}' in ckpt:
return ckpt
# fallback to .nnue if no checkpoint file
for nnue in nnues:
if f'run_{run_id}' in nnue and f'nn-epoch{epoch}' in nnue:
return nnue
return None
# A global instance of the resource monitor.
# There is no need to have more than one.
RESOURCE_MONITOR = SystemResourcesMonitor(2)
# A regex pattern for a float number.
NUMERIC_CONST_PATTERN = '[-+]?(?:(?:\d*\.\d+)|(?:\d+\.?))(?:[Ee][+-]?\d+)?'
class TrainingRun(Thread):
'''
Manages a single pytorch training run.
Starts it as a subprocess.
Provides information about the current state of training.
Runs as a separate thread and must be stopped before exiting.
'''
# The regex pattern for extracting information from the pytorch lightning's tqdm process bar output
ITERATION_PATTERN = re.compile(f'Epoch (\\d+).*?(\\d+)/(\\d+).*?({NUMERIC_CONST_PATTERN})it/s, loss=({NUMERIC_CONST_PATTERN})')
def __init__(
self,
gpu_id,
run_id,
nnue_pytorch_directory,
training_dataset,
validation_dataset,
num_data_loader_threads,
num_pytorch_threads,
num_epochs,
batch_size,
random_fen_skipping,
smart_fen_skipping,
wld_fen_skipping,
features,
lr,
gamma,
lambda_,
network_save_period,
save_last_network,
seed,
root_dir,
epoch_size,
validation_size,
start_from_model=None,
resume_training=False,
start_lambda=None,
end_lambda=None,
additional_args=[]
):
super(TrainingRun, self).__init__()
self._gpu_id = gpu_id
self._run_id = run_id
# Use abspaths because we will be running the script with a different cwd
self._nnue_pytorch_directory = os.path.abspath(nnue_pytorch_directory)
self._training_dataset = os.path.abspath(training_dataset)
self._validation_dataset = os.path.abspath(validation_dataset)
self._num_data_loader_threads = num_data_loader_threads
self._num_pytorch_threads = num_pytorch_threads
self._num_epochs = num_epochs
self._batch_size = batch_size
self._random_fen_skipping = random_fen_skipping
self._smart_fen_skipping = smart_fen_skipping
self._wld_fen_skipping = wld_fen_skipping
self._features = features
self._lr = lr
self._gamma = gamma
self._lambda = lambda_
self._start_lambda = start_lambda
self._end_lambda = end_lambda
self._network_save_period = network_save_period
self._save_last_network = save_last_network
self._seed = seed
self._root_dir = os.path.abspath(root_dir)
self._epoch_size = epoch_size
self._validation_size = validation_size
self._start_from_model = start_from_model
self._resume_training = resume_training
self._additional_args = additional_args
# State for the status updates
self._current_step_in_epoch = None
self._num_steps_in_epoch = None
self._current_epoch = None
self._current_loss = None
self._momentary_iterations_per_second = None
self._smooth_iterations_per_second = DecayingRunningAverage()
self._has_finished = False
self._has_started = False
self._networks = []
self._process = None
self._running = False
self._error = None
# For speed calculation
self._last_time = None
self._last_step = None
def _get_stringified_args(self):
args = [
self._training_dataset,
self._validation_dataset,
f'--num-workers={self._num_data_loader_threads}',
f'--threads={self._num_pytorch_threads}',
f'--max_epoch={self._num_epochs}',
f'--batch-size={self._batch_size}',
f'--random-fen-skipping={self._random_fen_skipping}',
f'--gpus={self._gpu_id},',
f'--features={self._features}',
f'--lr={self._lr}',
f'--gamma={self._gamma}',
f'--lambda={self._lambda}',
f'--network-save-period={self._network_save_period}',
f'--save-last-network={self._save_last_network}',
f'--seed={self._seed}',
f'--epoch-size={self._epoch_size}',
f'--validation-size={self._validation_size}',
f'--default_root_dir={self._root_dir}',
]
if self._smart_fen_skipping:
args.append('--smart-fen-skipping')
else:
args.append('--no-smart-fen-skipping')
if not self._wld_fen_skipping:
args.append('--no-wld-fen-skipping')
if self._start_lambda:
args.append(f'--start-lambda={self._start_lambda}')
if self._end_lambda:
args.append(f'--end-lambda={self._end_lambda}')
resumed = False
if self._resume_training:
ckpt_path = find_latest_checkpoint(self._root_dir)
if ckpt_path:
args.append(f'--resume_from_checkpoint={ckpt_path}')
resumed = True
if self._start_from_model and not resumed:
args.append(f'--resume-from-model={self._start_from_model}')
for arg in self._additional_args:
args.append(arg)
return args
def run(self):
if self._resume_training and os.path.exists(os.path.join(self._root_dir, 'training_finished')):
self._has_started = True
self._has_finished = True
self._running = False
return
self._running = True
cmd = [sys.executable, 'train.py'] + self._get_stringified_args()
LOGGER.info(f'Running training with command: {cmd}')
LOGGER.info(f'Also known as: {" ".join(cmd)}')
LOGGER.info(f'Running in working directory: {self._nnue_pytorch_directory}')
self._process = subprocess.Popen(
cmd,
cwd=self._nnue_pytorch_directory,
shell=False,
bufsize=-1,
stdin=subprocess.DEVNULL,
stdout=subprocess.PIPE,
stderr=subprocess.STDOUT,
)
terminate_process_on_exit(self._process)
reader = io.TextIOWrapper(self._process.stdout)
while self._process.poll() is None and self._running:
if not self._running:
break
# \r is properly recognized as a newline delimiter so we can just read by lines
line = reader.readline().strip()
if not self._has_finished:
try:
matches = TrainingRun.ITERATION_PATTERN.search(line)
if matches:
self._current_epoch = int(matches.group(1))
self._current_step_in_epoch = int(matches.group(2))
self._num_steps_in_epoch = int(matches.group(3))
# There appears to be a pytorch lightning bug where it displays
# negative speed when running from checkpoint. So we work around this
# by computing our own speed.
# Only update every 10 steps to avoid the it/s to blow up.
# (With a higher update frequence might be affected by IO caching)
curr_step = self._current_step_in_epoch
if curr_step == self._last_step:
continue
curr_time = time.perf_counter_ns()
if self._last_time is None or curr_step < self._last_step:
self._last_time = curr_time
self._last_step = curr_step
continue
#self._momentary_iterations_per_second = float(matches.group(4))
if curr_step % 10 == 0:
self._momentary_iterations_per_second = (curr_step-self._last_step)/((curr_time-self._last_time)/1e9)
self._smooth_iterations_per_second.update(self._momentary_iterations_per_second)
self._last_time = curr_time
self._last_step = curr_step
self._current_loss = float(matches.group(5))
self._has_started = True
# Provide some output for the cli interface.
if self._current_step_in_epoch % 100 == 0:
LOGGER.info(line)
else:
# Actually this is where most of the errors from pytorch must be handled.
LOGGER.info(line)
except:
# Usually errors. Aside from that all output should be catched above. We want these logged.
LOGGER.info(line)
pass
if 'CUDA_ERROR_OUT_OF_MEMORY' in line or 'CUDA out of memory' in line:
self._process.terminate()
self._error = 'Cuda out of memory error.'
break
# Since _num_steps_in_epochs includes validation steps, that we cannot actually catch
# and we don't know how to account for validation steps, and the trainer exits silently,
# we can just estimate whether it finished with a success by using some margin...
# NOTE: We still cannot catch when the trainer exits with no work, which for example
# happens when resuming from a checkpoint at the end of training.
if self._has_started and self._current_epoch == self._num_epochs - 1 and self._current_step_in_epoch >= self._num_steps_in_epoch * 0.9:
self._has_finished = True
if self._running and not self._has_finished:
if not self._error:
self._error = 'Unknown error occured.'
LOGGER.warning(f'Training run {self._run_id} exited unexpectedly.')
LOGGER.error(f'Error: {self._error}')
else:
LOGGER.info(f'Training run {self._run_id} finished.')
self._has_started = True
self._running = False
def stop(self):
self._running = False
self.join()
if self._process:
self._process.terminate()
self._process.wait()
@property
def gpu_id(self):
return self._gpu_id
@property
def run_id(self):
return self._run_id
@property
def current_step_in_epoch(self):
return self._current_step_in_epoch
@property
def current_epoch(self):
return self._current_epoch
@property
def num_steps_in_epoch(self):
return self._num_steps_in_epoch
@property
def num_epochs(self):
return self._num_epochs
@property
def current_loss(self):
return self._current_loss
@property
def momentary_iterations_per_second(self):
return self._momentary_iterations_per_second
@property
def smooth_iterations_per_second(self):
return self._smooth_iterations_per_second.value
@property
def has_finished(self):
return self._has_finished
@property
def has_started(self):
return self._has_started
@property
def networks(self):
return self._networks
@property
def is_running(self):
return self._running
@property
def error(self):
return self._error
@property
def batch_size(self):
return self._batch_size
def requests_get_content(url, *args, **kwargs):
try:
result = requests.get(url, *args, **kwargs)
result.raise_for_status()
return result.content
except Exception as e:
raise Exception(f'GET request to {url} failed')
def get_zipfile_members_strip_common_prefix(zipfile):
'''
Removes a common previx from zipfile entries.
So for example will remove the top-level directory.
'''
parts = []
for name in zipfile.namelist():
if not name.endswith('/'):
parts.append(name.split('/')[:-1])
offset = len('/'.join(os.path.commonprefix(parts)) + '/')
for zipinfo in zipfile.infolist():
name = zipinfo.filename
if len(name) > offset:
zipinfo.filename = name[offset:]
yield zipinfo
def git_download_branch_or_commit(directory, repo, branch_or_commit):
'''
Github proves an API to download zips of specific commits, so
we don't need to use git clone.
'''
url = f'http://github.com/{repo}/zipball/{branch_or_commit}'
zipped_content = requests_get_content(url, timeout=TIMEOUT)
zipped_input = zipfile.ZipFile(io.BytesIO(zipped_content), mode='r')
zipped_input.extractall(directory, get_zipfile_members_strip_common_prefix(zipped_input))
# Utility functions for dependency setup and executable location.
def make_ordo_executable_path(directory):
path = os.path.join(directory, 'ordo')
if sys.platform == "win32":
path += '.exe'
return path
def is_ordo_setup(directory):
try:
ordo_path = make_ordo_executable_path(directory)
with subprocess.Popen([ordo_path, '--help'], stdout=subprocess.DEVNULL) as process:
if process.wait(timeout=TIMEOUT):
return False
return True
except:
return False
def setup_ordo(directory):
if is_ordo_setup(directory):
LOGGER.info(f'Ordo already setup in {directory}')
return
LOGGER.info(f'Setting up ordo in {directory}.')
git_download_branch_or_commit(directory, *ORDO_GIT)
if sys.platform == "win32":
# need to append -DMINGW
# ugly hack for a dumb makefile