def lfw_view2_random_AsyncB(host, dbname, A):
trials = MongoTrials(
'mongo://%s:44556/%s/jobs' % (host, dbname),
exp_key=exp_keys['random_asyncB'],
refresh=True)
helper = BoostHelper(trials.trials)
# XXX: Should the weights be used? I don't think so, we're basically
# doing LPBoost at this point
members = helper.ensemble_members(MultiBandit())[:int(A)]
for ii, dd in enumerate(members):
ccc = helper.continues(dd)
print ii, dd['_id'], dd['tid'], dd['result']['loss'],
print (ccc['_id'] if ccc else None)
namebases = []
for doc in members:
namebase = '%s_%s' % (dbname, doc['_id'])
namebases.append(namebase)
get_view2_features(
slm_desc=doc['spec']['model']['slm'],
preproc=doc['spec']['model']['preproc'],
comparison=doc['spec']['comparison'],
namebase=namebase,
basedir=os.getcwd(),
)
basedirs = [os.getcwd()] * len(namebases)
train_view2(namebases=namebases, basedirs=basedirs)
def lfw_view2_random_AdaboostMixture(host, dbname, A):
trials = MongoTrials(
'mongo://%s:44556/%s/jobs' % (host, dbname),
exp_key=exp_keys['random'],
refresh=True)
bandit = MultiBandit()
mix = AdaboostMixture(trials, bandit, test_mask=True)
# XXX: Should the weights be used? I don't think so, we're basically
# doing LPBoost at this point
specs, weights, tids = mix.mix_models(int(A), ret_tids=True)
assert len(specs) == len(tids)
namebases = []
for spec, tid in zip(specs, tids):
# -- allow this feature cache to be
# reused by AdaboostMixture and
# SimpleMixtures of different
# sizes
#XXX: Potentially affected by the tid/injected jobs bug,
# but unlikely. Rerun just in case once dual svm solver is in.
namebase = '%s_%s' % (dbname, tid)
namebases.append(namebase)
get_view2_features(
slm_desc=spec['model']['slm'],
preproc=spec['model']['preproc'],
comparison=spec['comparison'],
namebase=namebase,
basedir=os.getcwd(),
)
basedirs = [os.getcwd()] * len(namebases)
train_view2(namebases=namebases, basedirs=basedirs)
def test_baby_view2_libsvm_kernel():
c = config_tiny_rnd0
test_set = range(20) + range(500, 520)
lfw.get_view2_features(c['slm'], c['preproc'], 'mult', 'libsvm', os.getcwd(),
test=test_set)
return lfw.train_view2(['libsvm'],[os.getcwd()],
test=test_set, use_libsvm={'kernel': 'precomputed'})
def lfw_view2_randomL(host, dbname):
trials = MongoTrials('mongo://%s:44556/%s/jobs' % (host, dbname),
refresh=False)
#B = main_lfw_driver(trials)
#E = B.get_experiment(name=('random', 'foo'))
mongo_trials = trials.view(exp_key=exp_keys['randomL'], refresh=True)
docs = [d for d in mongo_trials.trials
if d['result']['status'] == hyperopt.STATUS_OK]
local_trials = hyperopt.trials_from_docs(docs)
losses = local_trials.losses()
best_doc = docs[np.argmin(losses)]
#XXX: Potentially affected by the tid/injected jobs bug,
# but unlikely. Rerun just in case once dual svm solver is in.
print best_doc['spec']
namebase = '%s_randomL_%s' % (dbname, best_doc['tid'])
get_view2_features(
slm_desc=best_doc['spec']['model']['slm'],
preproc=best_doc['spec']['model']['preproc'],
comparison=best_doc['spec']['comparison'],
namebase=namebase,
basedir=os.getcwd(),
)
namebases = [namebase]
basedirs = [os.getcwd()] * len(namebases)
#train_view2(namebases=namebases, basedirs=basedirs)
# running on the try2 database
# finds id 1674
#train err mean 0.0840740740741
#test err mean 0.199666666667
#running with libsvm:
train_view2(namebases=namebases, basedirs=basedirs,
use_libsvm={'kernel':'precomputed'})
def test_baby_view2():
c = config_tiny_rnd0
lfw.get_view2_features(c['slm'], c['preproc'], 'mult', '', os.getcwd(),
test=50)
return lfw.train_view2([''],[os.getcwd()], test=50)