def test(clfs = None):
if clfs == None:
return
print('begin test')
#test_loader
load_test = rst.Load(testdata_path)
#测试数据，原始数据
test_docs = load_test.datas
#测试数据label,原始label。like:<Opinion target="food" category="FOOD#QUALITY" polarity="negative" from="4" to="8"/>
test_docs_labels = load_test.labels
#同上
load_train = rst.Load(traindata_path)
#加载特征
feature_loader = fp.LoadFeature(load_train.datas,load_train.labels)
testX = feature_loader.get_all_feature(test_docs)
#测试数据label转化为矩阵形式
testY = raw_label_process(test_docs_labels)
#预测概率初始化
predict_prob = np.zeros((len(test_docs), category_num))
#预测
for col, clf in enumerate(clfs):
prob = clf.predict_proba(testX)[:, 1]
predict_prob[:, col] = prob
# 根据threshold决定每一句预测的label(字符串形式)
labels = []
for docid, test_doc in enumerate(test_docs):
# up_threshold_ids : a tuple
up_threshold_ids = np.where(predict_prob[docid, :] >= paras['threshold'])
ids = up_threshold_ids[0]
# for id in up_threshold_ids:
if len(ids) == 0:
labels.append(None)
continue
label = []
for id in ids:
label.append(categorys[id])
labels.append(label)
#生成预测结果xml,利用评估工具A.jar来评测
rst.generate_xml(labels, input_path=testdata_path, output_path=output_path)
#利用自己的评测函数来评测，二者评测结果差不多
threshod_list = [paras['threshold'] for j in range(category_num)]
threshod_vec = np.array(threshod_list)
for row in range(len(test_docs)):
predict_prob[row, :] = np.less_equal(threshod_vec, predict_prob[row, :])
predict_prob =predict_prob.astype(int)
# get f score predict_prob and testY
fscore = get_fscore(testY, predict_prob)
print('fscore:'+str(fscore))
def train():
print('load train datas and labels')
load = rst.Load(traindata_path)
train_docs = load.datas
train_docs_labels = load.labels
feature_loader = fp.LoadFeature(train_docs,train_docs_labels)
print('begin train')
clfs = []
trainX = feature_loader.get_all_feature(train_docs)
#tfidf_matrix 是稀疏矩阵,转化为正常矩阵
trainY = raw_label_process(train_docs_labels)
#trainX, trainY = data_feature_precess(train_docs, train_all_docs_raw_str_labels,train_docs)
for category,trainX,y in getdata(trainX,trainY):
print('train the classfier:'+category)
clf = MLPClassifier(solver=paras['solver'],
hidden_layer_sizes=paras['hidden_layer_sizes'],
alpha=paras['alpha'],
learning_rate_init=paras['learning_rate_init'],
random_state=paras['random_state'],
verbose = False
)
#clf = SVC(kernel='linear', probability=True)
clf.fit(trainX,y)
clfs.append(clf)
return clfs
def model_selection2():
load = rst.Load(traindata_path)
docs = load.datas
docs_labels = load.labels
#key paras value:fscore
paras_fscore_map = {}
# 5 - cross validate
for i in range(5):
#random_state = 0 : 每次随机种子不一样
train_docs,eval_docs,train_docs_labels,eval_docs_labels = train_test_split(docs,docs_labels,test_size=0.2,random_state=0)
feature_loader = fp.LoadFeature(train_docs)
trainX = feature_loader.get_all_feature(train_docs)
trainY = raw_label_process(train_docs_labels)
evalX = feature_loader.get_all_feature(eval_docs)
evalY = raw_label_process(eval_docs_labels)
paras_group = []
for i1,i2,i3,i4,i5 in paras_generate():
paras_str = str(i1)+','+str(i2)+','+str(i3)+','+str(i4)+","+str(i5)
print('begin'+paras_str)
f = 0.0
clfs = []
for category, trainX, y in getdata(trainX,trainY):
#print('train the classfier in model selection MODEL:' + category)
clf = MLPClassifier(solver=i3,
hidden_layer_sizes=i2,
alpha=i4,
learning_rate_init=i1,
random_state=paras['random_state']
)
clf.fit(trainX, y)
clfs.append(clf)
predict_prob = np.zeros((len(eval_docs), category_num))
for col, clf in enumerate(clfs):
prob = clf.predict_proba(evalX)[:, 1]
predict_prob[:, col] = prob
clfs.clear()
#根据predict_label和threshold来决定预测label
threshod_list = [i5 for j in range(category_num)]
threshod_vec = np.array(threshod_list)
for row in range(len(eval_docs)):
predict_prob[row,:] = np.less_equal(threshod_vec,predict_prob[row,:])
predict_prob.astype(int)
#get f score predict_prob and testY
fscore = get_fscore(evalY,predict_prob)
if i == 0:
paras_fscore_map[paras_str] = fscore
else:
paras_fscore_map[paras_str] = (paras_fscore_map[paras_str] * (i-1)+fscore)/i
# print((fscore,[i1,i2,i3,i4,i5]))
# paras_group.append((fscore,[i1,i2,i3,i4,i5]))
sorted(paras_fscore_map.items(),lambda x,y : x[1] > y[1])
for item in paras_fscore_map.items():
print(item)
def qselect(A, k):
if len(A) < k: return A
pivot = A[-1]
right = [pivot] + [x for x in A[:-1] if x[1] >= pivot[1]]
rlen = len(right)
if rlen == k:
return right
if rlen > k:
return qselect(right, k)
else:
left = [x for x in A[:-1] if x[1] < pivot[1]]
return qselect(left, k - rlen) + right
if __name__ == '__main__':
# docs = [
# 'Judging from previous posts this used to be a good place, but not any longer.',
# 'The food was lousy - too sweet or too salty and the portions tiny.',
#
# ]
# #<Opinion target="place" category="RESTAURANT#GENERAL" polarity="negative" from="51" to="56"/>
# #<Opinion target="food" category="FOOD#QUALITY" polarity="negative" from="4" to="8"/>
# raw_sent_labels = [
# [{'Opinion target':'good place'}],
# [{'Opinion target':'good food'}]
# ]
traindata_path = './restaurant2015/ABSA-15_Restaurants_Train_Final.xml'
load = rst.Load(traindata_path)
get_namelist(load.datas, load.labels)
def cross_validating():
load = rst.Load(traindata_path)
docs = load.datas
docs_labels = load.labels
# key paras value:fscore
paras_fscore_map = {}
for i in range(5):
#random_state = 0 : 每次随机种子不一样
train_docs, eval_docs, train_docs_labels, eval_docs_labels = train_test_split(
docs, docs_labels, test_size=0.2, random_state=0)
trainY = tn.raw_label_process(train_docs_labels)
evalY = tn.raw_label_process(eval_docs_labels)
paras_group = []
for i1, i2 in gridsearch_threshold():
feature_loader = fp.LoadFeature(train_docs, train_docs_labels, i1,
i2)
trainX = feature_loader.get_all_feature(train_docs)
print(trainX.shape)
evalX = feature_loader.get_all_feature(eval_docs)
paras_str = str(i1) + ',' + str(i2)
print('begin' + paras_str)
f = 0.0
clfs = []
for category, trainX, y in tn.getdata(trainX, trainY):
#print('train the classfier in model selection MODEL:' + category)
clf = MLPClassifier(
solver=paras['solver'],
hidden_layer_sizes=paras['hidden_layer_sizes'],
alpha=paras['alpha'],
learning_rate_init=paras['learning_rate_init'],
random_state=paras['random_state'],
verbose=False)
clf.fit(trainX, y)
clfs.append(clf)
predict_prob = np.zeros((len(eval_docs), tn.category_num))
for col, clf in enumerate(clfs):
prob = clf.predict_proba(evalX)[:, 1]
predict_prob[:, col] = prob
clfs.clear()
#根据predict_label和threshold来决定预测label
threshod_list = [
paras['threshold'] for j in range(tn.category_num)
]
threshod_vec = np.array(threshod_list)
for row in range(len(eval_docs)):
predict_prob[row, :] = np.less_equal(threshod_vec,
predict_prob[row, :])
predict_prob.astype(int)
#get f score predict_prob and testY
fscore = tn.get_fscore(evalY, predict_prob)
if i == 0:
paras_fscore_map[paras_str] = fscore
else:
paras_fscore_map[paras_str] = (paras_fscore_map[paras_str] *
(i - 1) + fscore) / i
# print((fscore,[i1,i2,i3,i4,i5]))
# paras_group.append((fscore,[i1,i2,i3,i4,i5]))
paras_fscore_map = sorted(paras_fscore_map.items(), key=lambda x: x[1])
for item in list(paras_fscore_map.items()):
print(item)