def wrapper(db,feature_type, sample, data):
# サンプルのグループを取ってくるだけ
# Sample … (ft,_id,type,cls,group,likelihood,weight)
#print "function: predict"
clf_id = generate_clf_id(algorithm,feature_type,data)
# 予測部
collection = db["classifiers"]
try:
record = collection.find_one({'_id':clf_id})
if record == None:
return error_json("No classifier was found.")
print "load %s..."%record['clf']
clf = load_model(record['clf'])
print "done"
except:
print sys.exc_info()[0]
return error_json(sys.exc_info()[1])
if record.has_key('pca'):
print "load %s..."%record['pca']
pca = load_model(record['pca'])
print "done"
feature = sample.ft
sample.ft = pca.transform(sample.ft)
# sparseな学習をしても通常のベクトルを入力として良い
# (逆にsparseなベクトルを入れると出力のlikelihoodもsparseになる)
#if record.has_key('sparse'):
# sample.ft = lil_matrix(sample.ft).tocsr()
# algorithmに応じた処理(func)を行う
likelihood_list = func(clf,sample)
if record.has_key('pca'):
sample.ft = feature
likelihood_dict = {}
for i,l in enumerate(likelihood_list):
if not record['class_id2name'].has_key(str(i)):
# 本来，ここは通らないはず．稀にclass_id2nameに登録されていないクラスID (カテゴリ数と一致を返す．scikit-learnの仕様??要調査
continue
key = record['class_id2name'][str(i)]
likelihood_dict[key] = l
# 予測結果をデータベースへ追加
sample.likelihood[clf_id] = likelihood_dict
collections = db[feature_type]
if collections.find_one(sample._id):
collections.update_one({"_id":sample._id},{"$set":{'likelihood':sample.likelihood}})
sub_result = {'update::%s'%sample._id}
else:
sub_result = mongointerface.add(db,feature_type,sample)
result = success_json()
result['event'] = {'_id':"predict::"+clf_id+"::"+str(sample._id), 'sub_event':sub_result}
result['result'] = {'id':sample._id, 'likelihood':likelihood_dict}
if sample.ground_truth:
result['result']['ground_truth'] = sample.ground_truth
return result
def cross_validation(db, json_data_s, feature_type, algorithm, fold_num):
print("function: cross_validation")
data = json.loads(json_data_s)
init_data(data)
cv_group_head = "__cross_validation"
# disband all previously taged cross_validation_groups
for i in range(0,fold_num):
group_name = generate_group_name(cv_group_head, i)
mongointerface.disband(db, feature_type, {'group': group_name})
mongointerface.disband(db, feature_type, {'group': cv_group_head})
collections = db[feature_type]
selector = data['selector']
data['selector']['ground_truth'] = {"$exists": True}
samples = collections.find(selector)
# group samples into N groups randomly
samples_count = samples.count()
if samples_count == 0:
return error_json("ERROR: no samples are hit.")
group_assignment = []
remainder = samples_count % fold_num
quotient = int(samples_count / fold_num)
for i in range(0,fold_num):
n = quotient
if i < remainder:
n = n+1
print("group_count[%02d] = %d" % (i,n))
group_assignment += [generate_group_name(cv_group_head, i)] * n
random.shuffle(group_assignment)
# grouping samples into N group
for i in range(samples_count):
s = samples[i]
group_name = group_assignment[i]
#print group_name
groups = s['group']
if not group_name in groups:
groups = mongointerface.ensure_list(groups)
groups.append(group_name)
groups.append(cv_group_head)
_id = s['_id']
collections.update_one({"_id":_id},{"$set":{'group':groups}})
mod = __import__(algorithm+'.classifier', fromlist=['.'])
#print 'train and evaluation'
# evaluate each group by trained classifiers
confusion_matrices = []
# train, predict, and evaluate N classifiers
for i in range(0,fold_num):
## train ##
exclude_group = generate_group_name(cv_group_head, i)
#print exclude_group
_data = copy.deepcopy(data)
_data['selector'] = {'group':{'$not':{'$all':[exclude_group]},'$all':[cv_group_head]},'ground_truth':{"$exists": True}}
_data['overwrite'] = True
_data['name'] = exclude_group
#print _data
result = mod.train(db,feature_type,_data)
if result['status'] != 'success':
return result
## predict ##
selector = {'group':{'$all':[exclude_group]}}
group_samples = mongointerface.get_training_samples(db,feature_type,False,selector)
for s in group_samples:
result = mod.predict(db,feature_type, Sample(s), _data)
if result['status'] != 'success':
return result
_data['selector'] = selector
## evaluate ##
result = mongointerface.evaluate(db, feature_type, _data, algorithm)
if result['status'] != 'success':
return result
confusion_matrices.append(result['confusion_matrix'])
cmat = None
for m in confusion_matrices:
if bool(cmat):
cmat = merge_confusion_matrix(cmat,json.loads(m))
else:
cmat = json.loads(m)
result = success_json()
result['confusion_matrix'] = cmat
clf_id = generate_clf_id(algorithm,feature_type,data)
result['event'] = {'_id':"cross_validation::" + clf_id}
return result
