forked from hyperopt/hyperopt
-
Notifications
You must be signed in to change notification settings - Fork 102
/
mongoexp.py
1300 lines (1097 loc) · 46.3 KB
/
mongoexp.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
"""
Mongodb-based Trials Object
===========================
Components involved:
- mongo
e.g. mongod ...
- driver
e.g. hyperopt-mongo-search mongo://address bandit_json bandit_algo_json
- worker
e.g. hyperopt-mongo-worker --loop mongo://address
Mongo
=====
Mongo (daemon process mongod) is used for IPC between the driver and worker.
Configure it as you like, so that hyperopt-mongo-search can communicate with it.
I think there is some support in this file for an ssh+mongo connection type.
The experiment uses the following collections for IPC:
* jobs - documents of a standard form used to store suggested trials and their
results. These documents have keys:
* spec : subdocument returned by bandit_algo.suggest
* exp_key: an identifier of which driver suggested this trial
* cmd: a tuple (protocol, ...) identifying bandit.evaluate
* state: 0, 1, 2, 3 for job state (new, running, ok, fail)
* owner: None for new jobs, (hostname, pid) for started jobs
* book_time: time a job was reserved
* refresh_time: last time the process running the job checked in
* result: the subdocument returned by bandit.evaluate
* error: for jobs of state 3, a reason for failure.
* logs: a dict of sequences of strings received by ctrl object
* info: info messages
* warn: warning messages
* error: error messages
* fs - a gridfs storage collection (used for pickling)
* drivers - documents describing drivers. These are used to prevent two drivers
from using the same exp_key simultaneously, and to attach saved states.
* exp_key
* workdir: [optional] path where workers should chdir to
Attachments:
* pkl: [optional] saved state of experiment class
* bandit_args_kwargs: [optional] pickled (clsname, args, kwargs) to
reconstruct bandit in worker processes
The MongoJobs, and CtrlObj classes as well as the main_worker
method form the abstraction barrier around this database layout.
Worker
======
A worker looks up a job in a mongo database, maps that job document to a
runnable python object, calls that object, and writes the return value back to
the database.
A worker *reserves* a job by atomically identifying a document in the jobs
collection whose owner is None and whose state is 0, and setting the state to
1. If it fails to identify such a job, it loops with a random sleep interval
of a few seconds and polls the database.
If hyperopt-mongo-worker is called with a --loop argument then it goes back to
the database after finishing a job to identify and perform another one.
CtrlObj
-------
The worker allocates a CtrlObj and passes it to bandit.evaluate in addition to
the subdocument found at job['spec']. A bandit can use ctrl.info, ctrl.warn,
ctrl.error and so on like logger methods, and those messages will be written
to the mongo database (to job['logs']). They are not written synchronously
though, they are written when the bandit.evaluate function calls
ctrl.checkpoint().
Ctrl.checkpoint does several things:
* flushes logging messages to the database
* updates the refresh_time
* optionally updates the result subdocument
The main_worker routine calls Ctrl.checkpoint(rval) once after the
bandit.evalute function has returned before setting the state to 2 or 3 to
finalize the job in the database.
"""
__authors__ = ["James Bergstra", "Dan Yamins"]
__license__ = "3-clause BSD License"
__contact__ = "github.com/jaberg/hyperopt"
import copy
try:
import dill as cPickle
except ImportError:
import cPickle
import hashlib
import logging
import optparse
import os
import shutil
import signal
import socket
import subprocess
import sys
import time
import urlparse
import warnings
import numpy
import pymongo
import gridfs
from bson import SON
logger = logging.getLogger(__name__)
from .base import JOB_STATES
from .base import (JOB_STATE_NEW, JOB_STATE_RUNNING, JOB_STATE_DONE,
JOB_STATE_ERROR)
from .base import Trials
from .base import trials_from_docs
from .base import InvalidTrial
from .base import Ctrl
from .base import SONify
from .base import spec_from_misc
from .utils import coarse_utcnow
from .utils import fast_isin
from .utils import get_most_recent_inds
from .utils import json_call
import plotting
class OperationFailure(Exception):
"""Proxy that could be factored out if we also want to use CouchDB and
JobmanDB classes with this interface
"""
class Shutdown(Exception):
"""
Exception for telling mongo_worker loop to quit
"""
class WaitQuit(Exception):
"""
Exception for telling mongo_worker loop to quit
"""
class InvalidMongoTrial(InvalidTrial):
pass
class BanditSwapError(Exception):
"""Raised when the search program tries to change the bandit attached to
an experiment.
"""
class ReserveTimeout(Exception):
"""No job was reserved in the alotted time
"""
def read_pw():
username = 'hyperopt'
password = open(os.path.join(os.getenv('HOME'), ".hyperopt")).read()[:-1]
return dict(
username=username,
password=password)
def authenticate_for_db(db):
d = read_pw()
db.authenticate(d['username'], d['password'])
def parse_url(url, pwfile=None):
"""Unpacks a url of the form
protocol://[username[:pw]]@hostname[:port]/db/collection
:rtype: tuple of strings
:returns: protocol, username, password, hostname, port, dbname, collection
:note:
If the password is not given in the url but the username is, then
this function will read the password from file by calling
``open(pwfile).read()[:-1]``
"""
protocol=url[:url.find(':')]
ftp_url='ftp'+url[url.find(':'):]
# -- parse the string as if it were an ftp address
tmp = urlparse.urlparse(ftp_url)
logger.info( 'PROTOCOL %s'% protocol)
logger.info( 'USERNAME %s'% tmp.username)
logger.info( 'HOSTNAME %s'% tmp.hostname)
logger.info( 'PORT %s'% tmp.port)
logger.info( 'PATH %s'% tmp.path)
try:
_, dbname, collection = tmp.path.split('/')
except:
print >> sys.stderr, "Failed to parse '%s'"%(str(tmp.path))
raise
logger.info( 'DB %s'% dbname)
logger.info( 'COLLECTION %s'% collection)
if tmp.password is None:
if (tmp.username is not None) and pwfile:
password = open(pwfile).read()[:-1]
else:
password = None
else:
password = tmp.password
logger.info( 'PASS %s'% password)
return (protocol, tmp.username, password, tmp.hostname, tmp.port, dbname,
collection)
def connection_with_tunnel(host='localhost',
auth_dbname='admin', port=27017,
ssh=False, user='hyperopt', pw=None):
if ssh:
local_port=numpy.random.randint(low=27500, high=28000)
# -- forward from local to remote machine
ssh_tunnel = subprocess.Popen(
['ssh', '-NTf', '-L',
'%i:%s:%i'%(local_port, '127.0.0.1', port),
host],
#stdin=subprocess.PIPE,
#stdout=subprocess.PIPE,
#stderr=subprocess.PIPE,
)
# -- give the subprocess time to set up
time.sleep(.5)
connection = pymongo.Connection('127.0.0.1', local_port,
document_class=SON)
else:
connection = pymongo.Connection(host, port, document_class=SON)
if user:
if user == 'hyperopt':
authenticate_for_db(connection[auth_dbname])
else:
raise NotImplementedError()
ssh_tunnel=None
return connection, ssh_tunnel
def connection_from_string(s):
protocol, user, pw, host, port, db, collection = parse_url(s)
if protocol == 'mongo':
ssh=False
elif protocol in ('mongo+ssh', 'ssh+mongo'):
ssh=True
else:
raise ValueError('unrecognized protocol for MongoJobs', protocol)
connection, tunnel = connection_with_tunnel(
ssh=ssh,
user=user,
pw=pw,
host=host,
port=port,
)
return connection, tunnel, connection[db], connection[db][collection]
class MongoJobs(object):
"""
# Interface to a Jobs database structured like this
#
# Collections:
#
# db.jobs - structured {config_name, 'cmd', 'owner', 'book_time',
# 'refresh_time', 'state', 'exp_key', 'owner', 'result'}
# This is the collection that the worker nodes write to
#
# db.gfs - file storage via gridFS for all collections
#
"""
def __init__(self, db, jobs, gfs, conn, tunnel, config_name):
self.db = db
self.jobs = jobs
self.gfs = gfs
self.conn=conn
self.tunnel=tunnel
self.config_name = config_name
# TODO: rename jobs -> coll throughout
coll = property(lambda s : s.jobs)
@classmethod
def alloc(cls, dbname, host='localhost',
auth_dbname='admin', port=27017,
jobs_coll='jobs', gfs_coll='fs', ssh=False, user=None, pw=None):
connection, tunnel = connection_with_tunnel(
host, auth_dbname, port, ssh, user, pw)
db = connection[dbname]
gfs = gridfs.GridFS(db, collection=gfs_coll)
return cls(db, db[jobs_coll], gfs, connection, tunnel)
@classmethod
def new_from_connection_str(cls, conn_str, gfs_coll='fs', config_name='spec'):
connection, tunnel, db, coll = connection_from_string(conn_str)
gfs = gridfs.GridFS(db, collection=gfs_coll)
return cls(db, coll, gfs, connection, tunnel, config_name)
def __iter__(self):
return self.jobs.find()
def __len__(self):
try:
return self.jobs.count()
except:
return 0
def create_jobs_indexes(self):
jobs = self.db.jobs
for k in ['exp_key', 'result.loss', 'book_time']:
jobs.create_index(k)
def create_drivers_indexes(self):
drivers = self.db.drivers
drivers.create_index('exp_key', unique=True)
def create_indexes(self):
self.create_jobs_indexes()
self.create_drivers_indexes()
def jobs_complete(self, cursor=False):
c = self.jobs.find(spec=dict(state=JOB_STATE_DONE))
return c if cursor else list(c)
def jobs_error(self, cursor=False):
c = self.jobs.find(spec=dict(state=JOB_STATE_ERROR))
return c if cursor else list(c)
def jobs_running(self, cursor=False):
if cursor:
raise NotImplementedError()
rval = list(self.jobs.find(spec=dict(state=JOB_STATE_RUNNING)))
#TODO: mark some as MIA
rval = [r for r in rval if not r.get('MIA', False)]
return rval
def jobs_dead(self, cursor=False):
if cursor:
raise NotImplementedError()
rval = list(self.jobs.find(spec=dict(state=JOB_STATE_RUNNING)))
#TODO: mark some as MIA
rval = [r for r in rval if r.get('MIA', False)]
return rval
def jobs_queued(self, cursor=False):
c = self.jobs.find(spec=dict(state=JOB_STATE_NEW))
return c if cursor else list(c)
def insert(self, job, safe=True):
"""Return a job dictionary by inserting the job dict into the database"""
try:
cpy = copy.deepcopy(job)
# this call adds an _id field to cpy
_id = self.jobs.insert(cpy, safe=safe, check_keys=True)
# so now we return the dict with the _id field
assert _id == cpy['_id']
return cpy
except pymongo.errors.OperationFailure, e:
raise OperationFailure(e)
def delete(self, job, safe=True):
"""Delete job[s]"""
try:
self.jobs.remove(job, safe=safe)
except pymongo.errors.OperationFailure, e:
raise OperationFailure(e)
def delete_all(self, cond={}, safe=True):
"""Delete all jobs and attachments"""
try:
for d in self.jobs.find(spec=cond, fields=['_id', '_attachments']):
logger.info('deleting job %s' % d['_id'])
for name, file_id in d.get('_attachments', []):
try:
self.gfs.delete(file_id)
except gridfs.errors.NoFile:
logger.error('failed to remove attachment %s:%s' % (
name, file_id))
self.jobs.remove(d, safe=safe)
except pymongo.errors.OperationFailure, e:
raise OperationFailure(e)
def delete_all_error_jobs(self, safe=True):
return self.delete_all(cond={'state': JOB_STATE_ERROR}, safe=safe)
def reserve(self, host_id, cond=None, exp_key=None):
now = coarse_utcnow()
if cond is None:
cond = {}
else:
cond = copy.copy(cond) #copy is important, will be modified, but only the top-level
if exp_key is not None:
cond['exp_key'] = exp_key
#having an owner of None implies state==JOB_STATE_NEW, so this effectively
#acts as a filter to make sure that only new jobs get reserved.
if cond.get('owner') is not None:
raise ValueError('refusing to reserve owned job')
else:
cond['owner'] = None
cond['state'] = JOB_STATE_NEW #theoretically this is redundant, theoretically
try:
rval = self.jobs.find_and_modify(
cond,
{'$set':
{'owner': host_id,
'book_time': now,
'state': JOB_STATE_RUNNING,
'refresh_time': now,
}
},
new=True,
safe=True,
upsert=False)
except pymongo.errors.OperationFailure, e:
logger.error('Error during reserve_job: %s'%str(e))
rval = None
return rval
def refresh(self, doc, safe=False):
self.update(doc, dict(refresh_time=coarse_utcnow()), safe=False)
def update(self, doc, dct, safe=True, collection=None):
"""Return union of doc and dct, after making sure that dct has been
added to doc in `collection`.
This function does not modify either `doc` or `dct`.
safe=True means error-checking is done. safe=False means this function will succeed
regardless of what happens with the db.
"""
if collection is None:
collection = self.coll
dct = copy.deepcopy(dct)
if '_id' not in doc:
raise ValueError('doc must have an "_id" key to be updated')
if '_id' in dct:
if dct['_id'] != doc['_id']:
raise ValueError('cannot update the _id field')
del dct['_id']
if 'version' in dct:
if dct['version'] != doc['version']:
warnings.warn('Ignoring "version" field in update dictionary')
if 'version' in doc:
doc_query = dict(_id=doc['_id'], version=doc['version'])
dct['version'] = doc['version']+1
else:
doc_query = dict(_id=doc['_id'])
dct['version'] = 1
try:
# warning - if doc matches nothing then this function succeeds
# N.B. this matches *at most* one entry, and possibly zero
collection.update(
doc_query,
{'$set': dct},
safe=True,
upsert=False,
multi=False,)
except pymongo.errors.OperationFailure, e:
# translate pymongo failure into generic failure
raise OperationFailure(e)
# update doc in-place to match what happened on the server side
doc.update(dct)
if safe:
server_doc = collection.find_one(
dict(_id=doc['_id'], version=doc['version']))
if server_doc is None:
raise OperationFailure('updated doc not found : %s'
% str(doc))
elif server_doc != doc:
if 0:# This is all commented out because it is tripping on the fact that
# str('a') != unicode('a').
# TODO: eliminate false alarms and catch real ones
mismatching_keys = []
for k, v in server_doc.items():
if k in doc:
if doc[k] != v:
mismatching_keys.append((k, v, doc[k]))
else:
mismatching_keys.append((k, v, '<missing>'))
for k,v in doc.items():
if k not in server_doc:
mismatching_keys.append((k, '<missing>', v))
raise OperationFailure('local and server doc documents are out of sync: %s'%
repr((doc, server_doc, mismatching_keys)))
return doc
def attachment_names(self, doc):
def as_str(name_id):
assert isinstance(name_id[0], basestring), name_id
return str(name_id[0])
return map(as_str, doc.get('_attachments', []))
def set_attachment(self, doc, blob, name, collection=None):
"""Attach potentially large data string `blob` to `doc` by name `name`
blob must be a string
doc must have been saved in some collection (must have an _id), but not
necessarily the jobs collection.
name must be a string
Returns None
"""
# If there is already a file with the given name for this doc, then we will delete it
# after writing the new file
attachments = doc.get('_attachments', [])
name_matches = [a for a in attachments if a[0] == name]
# the filename is set to something so that fs.list() will display the file
new_file_id = self.gfs.put(blob, filename='%s_%s' % (doc['_id'], name))
logger.info('stored blob of %i bytes with id=%s and filename %s_%s' % (
len(blob), str(new_file_id), doc['_id'], name))
new_attachments = ([a for a in attachments if a[0] != name]
+ [(name, new_file_id)])
try:
ii = 0
doc = self.update(doc, {'_attachments': new_attachments},
collection=collection)
# there is a database leak until we actually delete the files that
# are no longer pointed to by new_attachments
while ii < len(name_matches):
self.gfs.delete(name_matches[ii][1])
ii += 1
except:
while ii < len(name_matches):
logger.warning("Leak during set_attachment: old_file_id=%s" % (
name_matches[ii][1]))
ii += 1
raise
assert len([n for n in self.attachment_names(doc) if n == name]) == 1
#return new_file_id
def get_attachment(self, doc, name):
"""Retrieve data attached to `doc` by `attach_blob`.
Raises OperationFailure if `name` does not correspond to an attached blob.
Returns the blob as a string.
"""
attachments = doc.get('_attachments', [])
file_ids = [a[1] for a in attachments if a[0] == name]
if not file_ids:
raise OperationFailure('Attachment not found: %s' % name)
if len(file_ids) > 1:
raise OperationFailure('multiple name matches', (name, file_ids))
return self.gfs.get(file_ids[0]).read()
def delete_attachment(self, doc, name, collection=None):
attachments = doc.get('_attachments', [])
file_id = None
for i,a in enumerate(attachments):
if a[0] == name:
file_id = a[1]
break
if file_id is None:
raise OperationFailure('Attachment not found: %s' % name)
#print "Deleting", file_id
del attachments[i]
self.update(doc, {'_attachments':attachments}, collection=collection)
self.gfs.delete(file_id)
class MongoTrials(Trials):
"""Trials maps on to an entire mongo collection. It's basically a wrapper
around MongoJobs for now.
As a concession to performance, this object permits trial filtering based
on the exp_key, but I feel that's a hack. The case of `cmd` is similar--
the exp_key and cmd are semantically coupled.
WRITING TO THE DATABASE
-----------------------
The trials object is meant for *reading* a trials database. Writing
to a database is different enough from writing to an in-memory
collection that no attempt has been made to abstract away that
difference. If you want to update the documents within
a MongoTrials collection, then retrieve the `.handle` attribute (a
MongoJobs instance) and use lower-level methods, or pymongo's
interface directly. When you are done writing, call refresh() or
refresh_tids() to bring the MongoTrials up to date.
"""
async = True
def __init__(self, arg, exp_key=None, cmd=None, workdir=None,
refresh=True):
if isinstance(arg, MongoJobs):
self.handle = arg
else:
connection_string = arg
self.handle = MongoJobs.new_from_connection_str(connection_string)
self.handle.create_indexes()
self._exp_key = exp_key
self.cmd = cmd
self.workdir = workdir
if refresh:
self.refresh()
def view(self, exp_key=None, cmd=None, workdir=None, refresh=True):
rval = self.__class__(self.handle,
exp_key=self._exp_key if exp_key is None else exp_key,
cmd=self.cmd if cmd is None else cmd,
workdir=self.workdir if workdir is None else workdir,
refresh=refresh)
return rval
def refresh_tids(self, tids):
""" Sync documents with `['tid']` in the list of `tids` from the
database (not *to* the database).
Local trial documents whose tid is not in `tids` are not
affected by this call. Local trial documents whose tid is in `tids` may
be:
* *deleted* (if db no longer has corresponding document), or
* *updated* (if db has an updated document) or,
* *left alone* (if db document matches local one).
Additionally, if the db has a matching document, but there is no
local trial with a matching tid, then the db document will be
*inserted* into the local collection.
"""
exp_key = self._exp_key
if exp_key != None:
query = {'exp_key' : exp_key}
else:
query = {}
t0 = time.time()
query['state'] = {'$ne': JOB_STATE_ERROR}
if tids is not None:
query['tid'] = {'$in': list(tids)}
orig_trials = getattr(self, '_trials', [])
_trials = orig_trials[:] #copy to make sure it doesn't get screwed up
if _trials:
db_data = list(self.handle.jobs.find(query,
fields=['_id', 'version']))
# -- pull down a fresh list of ids from mongo
if db_data:
#make numpy data arrays
db_data = numpy.rec.array([(x['_id'], int(x['version']))
for x in db_data],
names=['_id', 'version'])
db_data.sort(order=['_id', 'version'])
db_data = db_data[get_most_recent_inds(db_data)]
existing_data = numpy.rec.array([(x['_id'],
int(x['version'])) for x in _trials],
names=['_id', 'version'])
existing_data.sort(order=['_id', 'version'])
#which records are in db but not in existing, and vice versa
db_in_existing = fast_isin(db_data['_id'], existing_data['_id'])
existing_in_db = fast_isin(existing_data['_id'], db_data['_id'])
#filtering out out-of-date records
_trials = [_trials[_ind] for _ind in existing_in_db.nonzero()[0]]
#new data is what's in db that's not in existing
new_data = db_data[numpy.invert(db_in_existing)]
#having removed the new and out of data data,
#concentrating on data in db and existing for state changes
db_data = db_data[db_in_existing]
existing_data = existing_data[existing_in_db]
try:
assert len(db_data) == len(existing_data)
assert (existing_data['_id'] == db_data['_id']).all()
assert (existing_data['version'] <= db_data['version']).all()
except:
reportpath = os.path.join(os.getcwd(),
'hyperopt_refresh_crash_report_' + \
str(numpy.random.randint(1e8)) + '.pkl')
logger.error('HYPEROPT REFRESH ERROR: writing error file to %s' % reportpath)
_file = open(reportpath, 'w')
cPickle.dump({'db_data': db_data,
'existing_data': existing_data},
_file)
_file.close()
raise
same_version = existing_data['version'] == db_data['version']
_trials = [_trials[_ind] for _ind in same_version.nonzero()[0]]
version_changes = existing_data[numpy.invert(same_version)]
#actually get the updated records
update_ids = new_data['_id'].tolist() + version_changes['_id'].tolist()
num_new = len(update_ids)
update_query = copy.deepcopy(query)
update_query['_id'] = {'$in': update_ids}
updated_trials = list(self.handle.jobs.find(update_query))
_trials.extend(updated_trials)
else:
num_new = 0
_trials = []
else:
#this case is for performance, though should be able to be removed
#without breaking correctness.
_trials = list(self.handle.jobs.find(query))
if _trials:
_trials = [_trials[_i] for _i in get_most_recent_inds(_trials)]
num_new = len(_trials)
logger.debug('Refresh data download took %f seconds for %d ids' %
(time.time() - t0, num_new))
if tids is not None:
# -- If tids were given, then _trials only contains
# documents with matching tids. Here we augment these
# fresh matching documents, with our current ones whose
# tids don't match.
new_trials = _trials
tids_set = set(tids)
assert all(t['tid'] in tids_set for t in new_trials)
old_trials = [t for t in orig_trials if t['tid'] not in tids_set]
_trials = new_trials + old_trials
# -- reassign new trials to self, in order of increasing tid
jarray = numpy.array([j['_id'] for j in _trials])
jobsort = jarray.argsort()
self._trials = [_trials[_idx] for _idx in jobsort]
self._specs = [_trials[_idx]['spec'] for _idx in jobsort]
self._results = [_trials[_idx]['result'] for _idx in jobsort]
self._miscs = [_trials[_idx]['misc'] for _idx in jobsort]
def refresh(self):
self.refresh_tids(None)
def _insert_trial_docs(self, docs):
rval = []
for doc in docs:
rval.append(self.handle.jobs.insert(doc, safe=True))
return rval
def count_by_state_unsynced(self, arg):
exp_key = self._exp_key
# TODO: consider searching by SON rather than dict
if isinstance(arg, int):
if arg not in JOB_STATES:
raise ValueError('invalid state', arg)
query = dict(state=arg)
else:
assert hasattr(arg, '__iter__')
states = list(arg)
assert all([x in JOB_STATES for x in states])
query = dict(state={'$in': states})
if exp_key != None:
query['exp_key'] = exp_key
rval = self.handle.jobs.find(query).count()
return rval
def delete_all(self, cond=None):
if cond is None:
cond = {}
else:
cond = dict(cond)
if self._exp_key:
cond['exp_key'] = self._exp_key
# -- remove all documents matching condition
self.handle.delete_all(cond)
gfs = self.handle.gfs
for filename in gfs.list():
try:
fdoc = gfs.get_last_version(filename=filename, **cond)
except gridfs.errors.NoFile:
continue
gfs.delete(fdoc._id)
self.refresh()
def new_trial_ids(self, N):
db = self.handle.db
# N.B. that the exp key is *not* used here. It was once, but it caused
# a nasty bug: tids were generated by a global experiment
# with exp_key=None, running a BanditAlgo that introduced sub-experiments
# with exp_keys, which ran jobs that did result injection. The tids of
# injected jobs were sometimes unique within an experiment, and
# sometimes not. Hilarious!
#
# Solution: tids are generated to be unique across the db, not just
# within an exp_key.
#
# -- mongo docs say you can't upsert an empty document
query = {'a': 0}
doc = None
while doc is None:
doc = db.job_ids.find_and_modify(
query,
{'$inc' : {'last_id': N}},
upsert=True,
safe=True)
if doc is None:
logger.warning('no last_id found, re-trying')
time.sleep(1.0)
lid = doc.get('last_id', 0)
return range(lid, lid + N)
def trial_attachments(self, trial):
"""
Attachments to a single trial (e.g. learned weights)
Returns a dictionary interface to the attachments.
"""
# don't offer more here than in MongoCtrl
class Attachments(object):
def __contains__(_self, name):
return name in self.handle.attachment_names(doc=trial)
def __len__(_self):
return len(self.handle.attachment_names(doc=trial))
def __iter__(_self):
return iter(self.handle.attachment_names(doc=trial))
def __getitem__(_self, name):
try:
return self.handle.get_attachment(
doc=trial,
name=name)
except OperationFailure:
raise KeyError(name)
def __setitem__(_self, name, value):
self.handle.set_attachment(
doc=trial,
blob=value,
name=name,
collection=self.handle.db.jobs)
def __delitem__(_self, name):
raise NotImplementedError('delete trial_attachment')
def keys(self):
return [k for k in self]
def values(self):
return [self[k] for k in self]
def items(self):
return [(k, self[k]) for k in self]
return Attachments()
@property
def attachments(self):
"""
Attachments to a Trials set (such as bandit args).
Support syntax for load: self.attachments[name]
Support syntax for store: self.attachments[name] = value
"""
gfs = self.handle.gfs
query = {}
if self._exp_key:
query['exp_key'] = self._exp_key
class Attachments(object):
def __iter__(_self):
if query:
# -- gfs.list does not accept query kwargs
# (at least, as of pymongo 2.4)
filenames = [fname
for fname in gfs.list()
if fname in _self]
else:
filenames = gfs.list()
return iter(filenames)
def __contains__(_self, name):
return gfs.exists(filename=name, **query)
def __getitem__(_self, name):
try:
rval = gfs.get_version(filename=name, **query).read()
return rval
except gridfs.NoFile:
raise KeyError(name)
def __setitem__(_self, name, value):
if gfs.exists(filename=name, **query):
gout = gfs.get_last_version(filename=name, **query)
gfs.delete(gout._id)
gfs.put(value, filename=name, **query)
def __delitem__(_self, name):
gout = gfs.get_last_version(filename=name, **query)
gfs.delete(gout._id)
return Attachments()
class MongoWorker(object):
poll_interval = 3.0 # -- seconds
workdir = None
def __init__(self, mj,
poll_interval=poll_interval,
workdir=workdir,
exp_key=None,
logfilename='logfile.txt',
):
"""
mj - MongoJobs interface to jobs collection
poll_interval - seconds
workdir - string
exp_key - restrict reservations to this key
"""
self.mj = mj
self.poll_interval = poll_interval
self.workdir = workdir
self.exp_key = exp_key
self.logfilename = logfilename
def make_log_handler(self):
self.log_handler = logging.FileHandler(self.logfilename)
self.log_handler.setFormatter(
logging.Formatter(
fmt='%(levelname)s (%(name)s): %(message)s'))
self.log_handler.setLevel(logging.INFO)
def run_one(self,
host_id=None,
reserve_timeout=None,
erase_created_workdir=False,
):
if host_id == None:
host_id = '%s:%i'%(socket.gethostname(), os.getpid()),
job = None
start_time = time.time()
mj = self.mj
while job is None:
if (reserve_timeout
and (time.time() - start_time) > reserve_timeout):
raise ReserveTimeout()
job = mj.reserve(host_id, exp_key=self.exp_key)
if not job:
interval = (1 +
numpy.random.rand()
* (float(self.poll_interval) - 1.0))
logger.info('no job found, sleeping for %.1fs' % interval)
time.sleep(interval)
logger.debug('job found: %s' % str(job))
# -- don't let the cmd mess up our trial object
spec = spec_from_misc(job['misc'])
ctrl = MongoCtrl(
trials=MongoTrials(mj, exp_key=job['exp_key'], refresh=False),
read_only=False,
current_trial=job)
if self.workdir is None:
workdir = job['misc'].get('workdir', os.getcwd())
if workdir is None:
workdir = ''
workdir = os.path.join(workdir, str(job['_id']))
else:
workdir = self.workdir
workdir = os.path.abspath(os.path.expanduser(workdir))
cwd = os.getcwd()
sentinal = None
if not os.path.isdir(workdir):
# -- figure out the closest point to the workdir in the filesystem