def hyperopt_rval(self, save_grams):
rval = copy.deepcopy(self._results)
rval['attachments'] = {}
rval['grams'] = {}
rval['weights'] = {}
rval['trace'] = copy.deepcopy(foobar._trace)
saved = set()
def jsonify_train_results(rkey):
for norm_key in rval[rkey]:
for task_name in rval[rkey][norm_key]:
svm_dct = rval[rkey][norm_key][task_name]
ens = svm_dct.pop('ens')
rval['weights'].setdefault(norm_key, {})
rval['weights'][norm_key][task_name] = ens._weights
# -- stash these as attachments because they fill up the db.
xmean = svm_dct.pop('xmean')
xstd = svm_dct.pop('xstd')
if save_grams:
xmean_key = 'xmean_%s_%s_%s' % (rkey, norm_key, task_name)
xstd_key = 'xstd_%s_%s_%s' % (rkey, norm_key, task_name)
rval['attachments'][xmean_key] = cPickle.dumps(xmean, -1)
rval['attachments'][xstd_key] = cPickle.dumps(xstd, -1)
rval['grams'].setdefault(norm_key, [])
for (inorm_key, sample1, sample2) in ens._grams:
if inorm_key != norm_key:
# -- we're only interested in saving the grams
# calculated by this run.
continue
if (norm_key, sample1, sample2) in saved:
# -- already saved this one
continue
att_key = 'gram_%s_%s_%s.pkl' % (
norm_key, sample1, sample2)
info('saving %s' % att_key)
gram = ens._grams[(norm_key, sample1, sample2)]
rval['attachments'][att_key] = dumps_gram(
gram.astype('float32'))
rval['grams'][norm_key].append((sample1, sample2))
saved.add((norm_key, sample1, sample2))
saved.add((norm_key, sample2, sample1))
jsonify_train_results('train_image_match_indexed')
jsonify_train_results('retrain_classifier_image_match_indexed')
return SONify(rval)
def SONify(self, foo):
rval = SONify(foo)
assert bson.BSON.encode(dict(a=rval))
return rval
def test_nested_w_bool(self):
thing = dict(a=1, b='2', c=True, d=False, e=np.int(3), f=[1l])
assert thing == SONify(thing)
def test_nested_w_bool(self):
thing = dict(a=1, b="2", c=True, d=False, e=int(3), f=[1])
assert thing == SONify(thing)
