def df(x_train, y_train, n_features):
config = load_json("demo_ca.json")
gc = GCForest(config)
X_train = x_train.values.reshape(-1, 1, len(x_train.columns))
_, _features = gc.fit_transform(X_train, y_train)
_features = _features.sort_values(ascending=False)
return _features.index.values.tolist()[:n_features]
def RUN_2(best_th): #主函数，在获得最优分类阈值的情况下计算模型在测试集上的预测结果
comm_s_TPR = []
comm_s_TNR = []
comm_s_BER = []
comm_s_ACC = []
comm_s_MCC = []
comm_s_F1score = []
comm_s_AUC = []
comm_s_time = []
#将原始数据分为训练集，测试集
tiaocan_train, ceshi_train, tiaocan_train_test, ceshi_true = cross_validation.train_test_split(
comtest.iloc[0:len(comtest), 1:comtest.shape[1] - 1],
comtest.iloc[0:len(comtest), -1],
test_size=0.2,
random_state=0)
x_train = tiaocan_train
y_train = tiaocan_train_test
x_test = ceshi_train
y_true = ceshi_true
x_train = np.array(x_train, dtype=np.float16)
y_train = np.array(y_train, dtype=np.float16)
x_test = np.array(x_test, dtype=np.float16)
y_true = np.array(y_true, dtype=np.float16)
# x_train, y_train = RandomUnderSampler().fit_sample(x_train, y_train) #对训练集使用欠采样的方法达到类平衡
# 设置机器学习模型
##########################################################################################################################
############################## --XGB-- #############################
comm = GCForest(config)
comm.fit_transform(x_train, y_train)
pro_comm_Pre = comm.predict_proba(x_test)
blo_comm_Pre = blo(pro_comm_Pre, best_th) #根据最优分类阈值与预测概率计算画着生死情况
eva_comm = evaluating_indicator(y_true=y_true,
y_test=blo_comm_Pre,
y_test_value=pro_comm_Pre)
comm_s_TPR.append(eva_comm['TPR'])
comm_s_TNR.append(eva_comm['TNR'])
comm_s_BER.append(eva_comm['BER'])
comm_s_ACC.append(eva_comm['ACC'])
comm_s_MCC.append(eva_comm['MCC'])
comm_s_F1score.append(eva_comm['F1_score'])
comm_s_AUC.append(eva_comm['AUC'])
eva_comm = {
"TPR": np.mean(comm_s_TPR),
"TNR": np.mean(comm_s_TNR),
"BER": np.mean(comm_s_BER),
"ACC": np.mean(comm_s_ACC),
"MCC": np.mean(comm_s_MCC),
"F1_score": np.mean(comm_s_F1score),
"AUC": np.mean(comm_s_AUC),
"time": np.mean(comm_s_time)
}
return eva_comm
def get_m_gcForest(mtype="ca"):
"""
@param mtype: "ca" or "gc"
@param n_esti: n_estimators param in get_ca_config
"""
if mtype == "ca":
config = get_ca_config()
gc = GCForest(config)
return gc
if mtype == "gc":
config = get_gc_config()
gc = GCForest(config)
return gc
def run_model(features, adhd_labels, rand_params, verbose=True, test_size=0.2):
"""
Run the gcForest using parameters from the Optimizer. Use random portions of the original dataset
for testing and training (default 20%-80%)
:param kind: (str) The type of functional connectivity we want to use
:param features: (list) A matrix containing phenotypic and functional connectivity c
:param adhd_labels: (list) The correct labels from the dataset
:param rand_params: (dict) The generated random params from the Optimizer
:param verbose: (bool) Whether to print classification report
:param test_size: (float) How much of the dataset to use for testing
:return: (float) accuracy, (float) f1, (float) precision, (float) recall
"""
classifier = GCForest( # Instantiate the gcForest algorithm using the random parameters we generated
config=generate_gcforest_config(rand_params['mlp_layers'], rand_params['mlp_solver'],
rand_params['logistic_regressions'],
rand_params['svc_kernel'], rand_params['xgb_estimators'],
rand_params['rf_estimators'],
rand_params['early_stopping_iterations'], rand_params['positions']),
)
X_train, X_test, y_train, y_test = train_test_split(features, adhd_labels, test_size=test_size)
# Split the data into random subsets (20% test, 80% train by default)
classifier.fit_transform(np.array(X_train), np.array(y_train)) # Train the gcForest model
y_pred = classifier.predict(np.array(X_test)) # Predict off of the test dataset
y_test = np.array(y_test)
if verbose:
print "Classification Report\n", classification_report(y_test, y_pred) # Print out some useful run information
print "Accuracy:", accuracy_score(y_test, y_pred)
print "Confusion Matrix\n", confusion_matrix(y_test, y_pred)
positive_metrics = {
'f1': f1_score(y_test, y_pred), # Calculate the f1 for class "1"
'precision': precision_score(y_test, y_pred), # Calculate the precision for class "1"
'recall': recall_score(y_test, y_pred), # Calculate the recall for class "1"
}
negative_metrics = {
'f1': f1_score(y_test, y_pred, pos_label=0), # Calculate the f1 for class "0"
'precision': precision_score(y_test, y_pred, pos_label=0), # Calculate the precision for class "0"
'recall': recall_score(y_test, y_pred, pos_label=0), # Calculate the recall for class "0"
}
matrix = confusion_matrix(y_test, y_pred)
confusion = { # Return the attributes of the confusion matrix
'true_negative': matrix[0][0], # Predicted false and is false
'false_positive': matrix[0][1], # Predicted false and is true
'false_negative': matrix[1][0], # Predicted true and is false
'true_positive': matrix[1][1] # Predicted true and is true
}
scores = accuracy_score(y_test, y_pred), positive_metrics, negative_metrics, confusion
# Get the accuracy, f1, precision and recall of the model
return scores # Return it
def RUN(): #根据训练集与验证集获取最优分类阈值
tiaocan_train, ceshi_train, tiaocan_train_test, ceshi_true = cross_validation.train_test_split(
comtest.iloc[0:len(comtest), 1:comtest.shape[1] - 1],
comtest.iloc[0:len(comtest), -1],
test_size=0.2,
random_state=0)
position = []
skf = StratifiedKFold(n_splits=10) #设置十折交叉验证
tiaocan_train = np.array(tiaocan_train, dtype=np.float16)
tiaocan_train_test = np.array(tiaocan_train_test, dtype=np.float16)
times = 0
position = []
for train, test in skf.split(tiaocan_train, tiaocan_train_test):
alltime_start = time.time()
times = times + 1
x_train = tiaocan_train[train]
y_train = tiaocan_train_test[train]
x_test = tiaocan_train[test]
y_true = tiaocan_train_test[test]
# x_train, y_train = RandomUnderSampler().fit_sample(x_train, y_train) #使用欠采样的方法进行类平衡
# 设置机器学习模型
#
############################## --XGB-- #############################
comm = GCForest(config)
comm.fit_transform(x_train, y_train) #模型训练
pro_comm_Pre = comm.predict_proba(x_test)
############################### 敏感性特异性相近 ########################################
RightIndex = []
for jj in range(100): #计算模型在不同分类阈值下的各项指标
blo_comm_Pre = blo(pro_comm_Pre, jj)
eva_comm = evaluating_indicator(y_true=y_true,
y_test=blo_comm_Pre,
y_test_value=pro_comm_Pre)
RightIndex.append(abs(eva_comm['TPR'] - eva_comm['TNR']))
RightIndex = np.array(RightIndex, dtype=np.float16)
position = np.argmin(RightIndex) #选择出使得敏感性特异性最小的阈值作为分类阈值输出
alltime_end = time.time()
print('done_0, 第%s次验证 , time: %s s ' %
(times, alltime_end - alltime_start))
######################################################################################
return position.mean() #计算交叉验证输出的多个阈值的平均值作为最优分类阈值
def cross_validation(X, y, k, cpu):
config = get_toy_config(cpu = cpu)
classifier = GCForest(config)
cv = StratifiedKFold(n_splits = k)
res = {}
i=1
for train, test in cv.split(X, y):
tt = classifier.fit_transform(X[train], y[train])
yscore = classifier.predict_proba(X[test])
tmpID = "fold_" + str(i)
curDic = {}
curDic["yscore"] = yscore
curDic["ytest"] = y[test]
res[tmpID] = curDic
i = i + 1
return res
def run_gcforest(train_X,
test_X,
train_y,
test_y,
rounds=3,
layers=100,
seed=0):
config = get_toy_config(rounds, layers, seed)
gc = GCForest(config) # should be a dict
X_train_enc = gc.fit_transform(train_X, train_y)
ypred = np.array([i[1] for i in gc.predict_proba(test_X)])
metrics = gen_eval_metrics(test_y, ypred)
accuracy = metrics[0]
#cor = sum([int(ypred[i] + 0.5) == test_y[i] for i in range(len(ypred))])
#accuracy = cor / len(test_y)
print('Fold accuracy: ' + str(accuracy))
return metrics
def run_classification_configuration(X_train_10_fold, X_test_10_fold, y_train_10_fold, y_test_10_fold,test_idx_10_fold, train_idx_10_fold,trees,max_depth, min_child_weight,layer,cw=0.001):
i = 0
folds_AUC_testing = []
folds_AUPR_testing = []
folds_AUC_training = []
folds_AUPR_training = []
test_true_predict_compare = []
train_true_predict_compare = []
for X_train, X_test, y_train, y_test, test_idx_fold, train_idx_fold in zip(X_train_10_fold, X_test_10_fold, y_train_10_fold, y_test_10_fold, test_idx_10_fold, train_idx_10_fold):
# X_train, X_test = X_train[:,12:], X_test[:,12:]
# X_train, X_test = X_train[:,:12], X_test[:,:12]
config = get_toy_config(trees,max_depth, min_child_weight,cw,layer)
gc = GCForest(config)
#print(config)
X_train_enc = gc.fit_transform(X_train, y_train, X_test, y_test)
y_pred_train = gc.predict(X_train)
y_predprob_train = gc.predict_proba(X_train)
y_pred_test = gc.predict(X_test)
y_predprob_test = gc.predict_proba(X_test)
y_predprob_test_df = pd.DataFrame(y_predprob_test)
y_predprob_train_df = pd.DataFrame(y_predprob_train)
test_true_predict_compare.append([test_idx_fold, y_pred_test, y_test, y_predprob_test[:,0], y_predprob_test[:,1]]) #10-cv
train_true_predict_compare.append([train_idx_fold, y_pred_train, y_train, y_predprob_train[:,0], y_predprob_train[:,1]]) #10-cv
precision_training, recall_training, _ = precision_recall_curve(y_train, y_predprob_train[:,1], pos_label=1)
precision_testing, recall_testing, _ = precision_recall_curve(y_test, y_predprob_test[:,1], pos_label=1)
AUPR_training = auc(recall_training,precision_training)
AUPR_testing = auc(recall_testing, precision_testing)
AUC_training = roc_auc_score(y_train, y_predprob_train[:,1])
AUC_testing = roc_auc_score(y_test, y_predprob_test[:,1])
folds_AUC_testing.append(AUC_testing)
folds_AUPR_testing.append(AUPR_testing)
folds_AUC_training.append(AUC_training)
folds_AUPR_training.append(AUPR_training)
Avg_AUPR_training = np.mean(folds_AUPR_training)
Avg_AUPR_testing = np.mean(folds_AUPR_testing)
Avg_AUC_training = np.mean(folds_AUC_training)
Avg_AUC_testing = np.mean(folds_AUC_testing)
return [Avg_AUPR_training,Avg_AUPR_testing,folds_AUPR_testing, Avg_AUC_training,Avg_AUC_testing,folds_AUC_testing,folds_AUPR_training,folds_AUC_training], [test_true_predict_compare,train_true_predict_compare]
def fit(self, xtrain: pd.DataFrame, ytrain: pd.Series):
"""
Fit model
:param xtrain: training features
:param ytrain: training labels
"""
clf = GCForest(self.config)
if self.scaler is None:
clf.fit_transform(xtrain, ytrain)
else:
xtrain_norm = self.scaler.fit_transform(xtrain)
clf.fit_transform(xtrain_norm, ytrain)
return DeepRandomForestModel(clf, self.scaler)
# -*- coding:utf-8 -*-
import pandas as pd
from utils import avg_importance
from sklearn.model_selection import StratifiedKFold
import gcforest.data_load as load
from gcforest.gcforest import GCForest
import utils
cv = StratifiedKFold(n_splits=5, shuffle=True, random_state=0)
config = utils.load_json("demo_ca.json")
gc = GCForest(config)
datasets = ['cirrhosis', 'obesity', 't2d']
for dataset_idx, name in enumerate(datasets):
thre_features = {}
X = None
Y = None
if name == 'cirrhosis':
X, Y = load.cirrhosis_data()
elif name == 't2d':
X, Y = load.t2d_data()
elif name == 'obesity':
X, Y = load.obesity_data()
else:
raise Exception('the dataset is not defined!!!')
loocv = LeaveOneOut()
clf_rf = RandomForestClassifier(n_estimators=50, random_state=0)
clf_svm = SVC(kernel='linear',
C=1,
gamma=0.001,
random_state=0,
probability=True)
# xgb_crf = XGBClassifier(n_estimators=50)
# config = gcforest_config()
config = load_json("gc.json")
clf_gc = GCForest(config)
gc_pred_acc = []
# # ==============================================
f, ax = plt.subplots(1, 1)
params = [(clf_rf, 'green', "Random Forest"), (clf_svm, 'black', "SVM"),
(clf_gc, 'red', "Deep Forest")]
# (xgb_crf,'purple', "XGBoosing")]
# params = [(clf_gc,'red',"Deep Forest")]
for x in params:
mean_fpr = np.linspace(0, 1, 100)
tprs = []
aucs = []
i = 1
for train, test in loocv.split(X, Y):
probas_ = None
# In[12]:
config = get_toy_config()
models = [
LogisticRegression(),
LinearDiscriminantAnalysis(),
SVC(probability=True),
DecisionTreeClassifier(),
ExtraTreeClassifier(),
GaussianNB(),
KNeighborsClassifier(),
RandomForestClassifier(random_state=random_seed),
ExtraTreesClassifier(random_state=random_seed),
GCForest(config)
]
# In[16]:
test_entries = []
train_entries = []
for model in models:
model_name = model.__class__.__name__
if model_name == 'GCForest':
model.fit_transform(X_train, y_train, X_test, y_test)
else:
model.fit(X_train, y_train)
y_train_pred = model.predict(X_train)
y_test_pred = model.predict(X_test)
X_train_oversampled_batch, y_train_oversampled_batch = Batch(
X_train_oversampled, y_train_oversampled, batch_size)
X_train_batch, y_train_batch = Batch(X_train, y_train, batch_size)
X_valid_batch, y_valid_batch = Batch(X_valid, y_valid, batch_size)
# # GcForest
#
# ## train gc
# #### 1.train GcForest on oversampled datasets
# In[21]:
config = get_toy_config()
gc = GCForest(config)
X_train_enc = gc.fit_transform(X_train_oversampled, y_train_oversampled)
# dump
with open("../pkl/2018_test.pkl", "wb") as f:
pickle.dump(gc, f, pickle.HIGHEST_PROTOCOL)
# load
with open("../pkl/2018_test.pkl", "rb") as f:
gc = pickle.load(f)
# #### test GcForest on valid datasets
# In[22]:
y_valid_pred = gc.predict(X_valid)
"type": "LogisticRegression"
})
else:
ca_config["estimators"].append({
"n_folds": 2,
"type": "ExtraTreesClassifier",
"n_estimators": 10,
"max_depth": None,
"n_jobs": -1
})
config["cascade"] = ca_config
return config
config = get_toy_config(all_estimators=all_estimators)
gc = GCForest(config)
# If the model you use cost too much memory for you.
# You can use these methods to force gcforest not keeping model in memory
# gc.set_keep_model_in_mem(False), default is TRUE.
n_test = 500
# (X_train, y_train), (X_test, y_test) = mnist.load_data()
X_train, y_train = train_dataset_x[:-n_test], train_dataset_y[:-n_test]
X_test_cv, y_test_cv = train_dataset_x[-n_test:], train_dataset_y[-n_test:]
X_train = X_train[:, np.newaxis, :, :]
X_test_cv = X_test_cv[:, np.newaxis, :, :]
X_train_enc = gc.fit_transform(X_train, y_train)
y_pred_cv = gc.predict(X_test_cv)
acc = accuracy_score(y_test_cv, y_pred_cv)
feature = np.vstack((positive_feature, negative_sample_feature))
label1 = np.ones((len(positive_feature), 1))
label0 = np.zeros((len(negative_sample_feature), 1))
label = np.vstack((label1, label0))
rs = np.random.randint(0, 1000, 1)[0]
kf = StratifiedKFold(label[:, 0], n_folds=5, shuffle=True, random_state=rs)
test_auc_fold = []
test_aupr_fold = []
for train_index, test_index in kf:
Xtrain, Xtest = feature[train_index], feature[test_index]
Ytrain, Ytest = label[train_index], label[test_index]
config = get_toy_config()
rf = GCForest(config)
Ytrain = Ytrain.flatten()
rf.fit_transform(Xtrain, Ytrain)
# deep forest
predict_y = rf.predict(Xtest)
acc = accuracy_score(Ytest, predict_y)
print("Test Accuracy of GcForest = {:.2f} %".format(acc * 100))
prob_predict_y = rf.predict_proba(
Xtest
) # Give a result with probability values，the probability sum is 1
predictions_validation = prob_predict_y[:, 1]
fpr, tpr, _ = roc_curve(Ytest, predictions_validation)
roc_auc = auc(fpr, tpr)
aupr = average_precision_score(Ytest, predictions_validation)
print(roc_auc)
ca_config["estimators"].append({
"n_folds": 5,
"type": "LogisticRegression"
})
config["cascade"] = ca_config
return config
if __name__ == "__main__":
args = parse_args()
if args.model is None:
config = get_toy_config()
else:
config = load_json(args.model)
gc = GCForest(config)
# If the model you use cost too much memory for you.
# You can use these methods to force gcforest not keeping model in memory
gc.set_keep_model_in_mem(False) # default is TRUE.
(X_train, y_train), (X_test, y_test) = mnist.load_data()
# X_train, y_train = X_train[:2000], y_train[:2000]
X_train = X_train[:, np.newaxis, :, :]
X_test = X_test[:, np.newaxis, :, :]
X_train_enc = gc.fit_transform(X_train, y_train)
# X_enc is the concatenated predict_proba result of each estimators of the last layer of the GCForest model
# X_enc.shape =
# (n_datas, n_estimators * n_classes): If cascade is provided
# (n_datas, n_estimators * n_classes, dimX, dimY): If only finegrained part is provided
# You can also pass X_test, y_test to fit_transform method, then the accracy on test data will be logged when training.
def GAGCForest_prediction(feature_data, result_data):
n_splits = 5
acc_scores = np.zeros(n_splits)
recall_scores = np.zeros(n_splits)
mcc_scores = np.zeros(n_splits)
f1_scores = np.zeros(n_splits)
skfolds = StratifiedKFold(n_splits=n_splits,
shuffle=True,
random_state=random_state).split(
feature_data, result_data)
new_test_pred = np.zeros(feature_data.shape[0])
new_test_proba = np.zeros(feature_data.shape[0])
for j, (train_idx, test_idx) in enumerate(skfolds):
X_train = feature_data[train_idx]
Y_train = result_data[train_idx]
X_test = feature_data[test_idx]
Y_test = result_data[test_idx]
config = get_toy_config()
gc = GCForest(config) # should be a dict
X_train_enc = gc.fit_transform(X_train, Y_train)
y_pred = gc.predict(X_test)
X_test_enc = gc.transform(X_test)
# 获取函数接口地址
AIM_M = __import__('aimfuc')
AIM_F = 'gcforestCM'
"""============================变量设置============================"""
w1 = [0, 1]
w2 = [0, 1]
w3 = [0, 1]
b1 = [1, 1]
b2 = [1, 1]
b3 = [1, 1]
ranges = np.vstack([w1, w2, w3]).T # 生成自变量的范围矩阵
borders = np.vstack([b1, b2, b3]).T # 生成自变量的边界矩阵
# ranges = np.vstack([np.zeros((1, 3)), np.ones((1, 3))]) # 生成自变量的范围矩阵
# print(shape(ranges))
# borders = np.vstack([.ones((1, 3)), np.ones((1, 3))]) # 生成自变量的边界矩阵
precisions = [6] * 3 # 自变量的编码精度
scales = [0] * 3
codes = [1] * 3
# print(np.ones((1, 300)))
# scales = list(np.zeros((1, 300))) # 采用算术刻度
# codes = np.vstack([np.ones((1, 300)), np.ones((1, 300))]) # 变量的编码方式，2个变量均使用格雷编码
# print(shape(codes))
"""========================遗传算法参数设置========================="""
# NIND = 50 # 种群规模
# MAXGEN = 100 # 最大遗传代数
# GGAP = 0.8 # 代沟：子代与父代个体不相同的概率为0.8
# selectStyle = 'sus'; # 遗传算法的选择方式设为"sus"——随机抽样选择
# recombinStyle = 'xovdp' # 遗传算法的重组方式，设为两点交叉
# recopt = 0.9 # 交叉概率
# pm = 0.1 # 变异概率
# SUBPOP = 1 # 设置种群数为1
# maxormin = 1 #
# 设置最大最小化目标标记为1，表示是最小化目标，-1则表示最大化目标
FieldD = ga.crtfld(ranges, borders, precisions, codes, scales) #
# 调用编程模板
[weightarray, pop_trace, var_trace,
times] = new_code_templet(AIM_M,
AIM_F,
None,
None,
FieldD,
problem='R',
maxormin=-1,
MAXGEN=10,
NIND=50,
SUBPOP=1,
GGAP=0.8,
selectStyle='sus',
recombinStyle='xovsp',
recopt=0.9,
pm=0.7,
distribute=True,
proba=X_train_enc,
result=Y_train,
drawing=0)
print('用时：', times, '秒')
# w3 = 1 - weight[0] - weight[1]
# print(weight)
# weightarray = np.concatenate((weight, [w3]), axis=0)
for element in weightarray:
print(element)
test_probaF = X_test_enc[:, ::2].T
test_probaT = X_test_enc[:, 1::2].T
test_predT = np.dot(weightarray, test_probaT)
test_predF = np.dot(weightarray, test_probaF)
test_pred = np.zeros(len(test_predT))
test_proba = np.zeros(len(test_predT))
for i in range(len(test_predT)):
temper = test_predT[i] + test_predF[i]
test_proba = test_predT / temper
if (test_predT[i] > test_predF[i]):
test_pred[i] = 1
else:
test_pred[i] = 0
confmat = confusion_matrix(Y_test, test_pred)
sn = confmat[1, 1] / (confmat[1, 0] + confmat[1, 1])
sp = confmat[0, 0] / (confmat[0, 0] + confmat[0, 1])
print('1. The acc score of the model {}\n'.format(
accuracy_score(Y_test, test_pred)))
print('2. The sp score of the model {}\n'.format(sp))
print('3. The sn score of the model {}\n'.format(sn))
print('4. The mcc score of the model {}\n'.format(
matthews_corrcoef(Y_test, test_pred)))
print('9. The auc score of the model {}\n'.format(
roc_auc_score(Y_test, test_proba, average='macro')))
print('6. The recall score of the model {}\n'.format(
recall_score(Y_test, test_pred, average='macro')))
print('5. The F-1 score of the model {}\n'.format(
f1_score(Y_test, test_pred, average='macro')))
print('7. Classification report \n {} \n'.format(
classification_report(Y_test, test_pred)))
print('8. Confusion matrix \n {} \n'.format(
confusion_matrix(Y_test, test_pred)))
recall = recall_score(Y_test, test_pred, average='macro')
f1 = f1_score(Y_test, test_pred, average='macro')
acc = accuracy_score(Y_test, test_pred)
mcc = matthews_corrcoef(Y_test, test_pred)
recall_scores[j] = recall
f1_scores[j] = f1
acc_scores[j] = acc
mcc_scores[j] = mcc
new_test_pred[test_idx] = test_pred
new_test_proba[test_idx] = test_proba
print("CV- {} recall: {}, acc_score: {} , mcc_score: {}, f1_score: {}".
format(j, recall, acc, mcc, f1))
new_confmat = confusion_matrix(result_data, new_test_pred)
sn = new_confmat[1, 1] / (new_confmat[1, 0] + new_confmat[1, 1])
sp = new_confmat[0, 0] / (new_confmat[0, 0] + new_confmat[0, 1])
print(
"---------------------------------遗传算法-----------------------------------------"
)
print('1. The acc score of the model {}\n'.format(
accuracy_score(result_data, new_test_pred)))
print('2. The sp score of the model {}\n'.format(sp))
print('3. The sn score of the model {}\n'.format(sn))
print('4. The mcc score of the model {}\n'.format(
matthews_corrcoef(result_data, new_test_pred)))
print('9. The auc score of the model {}\n'.format(
roc_auc_score(result_data, new_test_proba, average='macro')))
print('6. The recall score of the model {}\n'.format(
recall_score(result_data, new_test_pred, average='macro')))
print('5. The F-1 score of the model {}\n'.format(
f1_score(result_data, new_test_pred, average='macro')))
print('7. Classification report \n {} \n'.format(
classification_report(result_data, new_test_pred)))
print('8. Confusion matrix \n {} \n'.format(
confusion_matrix(result_data, new_test_pred)))
def load_json(path):
import json
"""
支持以//开头的注释
"""
lines = []
with open(path) as f:
for row in f.readlines():
if row.strip().startswith("//"):
continue
lines.append(row)
return json.loads("\n".join(lines))
X, Y = load.obesity_data()
x_tr,x_te,y_tr,y_te = train_test_split(X,Y,random_state=42,stratify=Y)
clf_rf = RandomForestClassifier(n_estimators=200, random_state=0)
clf_rf.fit(x_tr,y_tr)
y_pred = clf_rf.predict(x_te)
print(accuracy_score(y_te,y_pred))
config = load_json("/home/qiang/repo/python/cascade_clf/examples/demo_ca.json")
clf_gc = GCForest(config)
clf_gc.fit_transform(x_tr.values, y_tr)
y_pred = clf_gc.predict(x_te.values)
print(accuracy_score(y_te, y_pred))
# load gc config
# In[12]:
y_train2 = y_train2.values
y_train = y_train.values
y_valid = y_valid.values
y_test = y_test.values
# In[90]:
config = get_toy_config()
model = GCForest(config)
model.fit_transform(X_train2, y_train2, X_test, y_test)
gc_valid_proba = model.predict_proba(X_valid)
gc_pred = model.predict(X_valid)
# In[14]:
models = [
LogisticRegression(),
LinearDiscriminantAnalysis(),
SVC(probability=True),
DecisionTreeClassifier(),
ExtraTreeClassifier(),
from sklearn.ensemble import RandomForestClassifier
from sklearn.metrics import accuracy_score
from sklearn.model_selection import StratifiedKFold
from sklearn.preprocessing import scale, StandardScaler
from sklearn.metrics import roc_curve, auc
from dimension_reduction import KPCA, LLE, pca
import utils.tools as utils
from gcforest.gcforest import GCForest
from gcforest.utils.config_utils import load_json
start = time.time()
path1 = 'gcforest4.json'
config = load_json(path1)
gc = GCForest(config)
extraction = sio.loadmat('yeast_feature_end.mat')
proteinA = extraction.get('feature_A')
protein_A = np.array(proteinA)
proteinB = extraction.get('feature_B')
protein_B = np.array(proteinB)
X_ = np.concatenate((protein_A, protein_B), axis=1)
X_ = np.array(X_)
[row, column] = np.shape(X_)
label_P = np.ones(int(row / 2))
label_N = np.zeros(int(row / 2))
label_ = np.hstack((label_P, label_N))
y_raw = np.mat(label_)
y_raw = np.transpose(y_raw)
y_ = np.array(y_raw)
# ca_config["estimators"].append(
# {"n_folds": 3, "type": "RandomForestClassifier", "n_estimators": 10, "n_jobs": -1,"random_state":0})
# ca_config["estimators"].append(
# {"n_folds": 3, "type": "XGBClassifier", "n_estimators": 10,
# "silent": True, "nthread": -1, "learning_rate": 0.1} )
# ca_config["estimators"].append({"n_folds": 3, "type": "ExtraTreesClassifier","max_depth": None, "n_jobs": -1})
# ca_config["estimators"].append({"n_folds": 3, "type": "LogisticRegression"})
config["cascade"] = ca_config
return config
if __name__ == "__main__":
X, Y = load2.cirrhosis_data()
config = gcforest_config()
gc = GCForest(config)
AUCs = []
for i in range(10):
cv = StratifiedKFold(n_splits=10, shuffle=True)
# # ==============================================
mean_fpr = np.linspace(0, 1, 100)
tprs = []
aucs = []
for train, test in cv.split(X, Y):
x_train = X.iloc[train]
y_train = Y[train]
x_test = X.iloc[test]
y_test = Y[test]
# test = pd.read_csv("../data/water/csv/test2018.csv")
# X_test = test.values[:, 0:-1]
# y_test = test.values[:, -1]
# X_test = clean_pipeline.fit_transform(X_test)
# In[13]:
# X_train, X_valid, y_train, y_valid = train_test_split(X, y, test_size=test_size, stratify = y, random_state = random_seed)
# X_train_oversampled, y_train_oversampled = Smoter(X_train, y_train, is_random=True)
config = get_toy_config()
gc = GCForest(config)
gc.fit_transform(X_train_oversampled, y_train_oversampled, X_valid, y_valid)
# y_valid_pred = gc.predict(X_valid)
# In[13]:
# dump
with open("../pkl/2019_gc.pkl", "wb") as f:
pickle.dump(gc, f, pickle.HIGHEST_PROTOCOL)
# # load
# with open("../pkl/2018_gc.pkl", "rb") as f:
# gc = pickle.load(f)
x_u_s = np.concatenate((x_p_test, x_u), axis=0)
y_u_s = np.concatenate((y_p_test, y_u), axis=0)
y_u_s = np.zeros(y_u_s.shape[0])
x = np.concatenate((x_p_s, x_u_s), axis=0)
y = np.concatenate((y_p_s, y_u_s), axis=0)
x_train, x_test, y_train, y_test = train_test_split(x,
y,
test_size=0.2,
random_state=1)
# scaler = StandardScaler().fit(X_train)
# X_train_transformed = scaler.transform(X_train)
# X_test_transformed = scaler.transform(X_test)
config = get_toy_config()
gc = GCForest(config)
gc.fit_transform(x_train, y_train)
scores = gc.predict_proba(x_u_test)[:, 0]
orderScores = np.argsort(-scores)
orderList = [str(item) for item in orderScores]
orderStr = ','.join(orderList)
top = int(y_u_test.shape[0] * 0.25)
topNIndex = orderScores[:top]
t = 0
while t < top:
index = topNIndex[t]
x_n = x_u[index]
X_n = np.vstack((X_n, x_n))
t += 1
X_n = X_n[1:, :]
ca_config["estimators"].append({
"n_folds": 5,
"type": "LogisticRegression"
})
config["cascade"] = ca_config
return config
if __name__ == "__main__":
args = parse_args()
if args.model is None:
config = get_toy_config()
else:
config = load_json(args.model)
gc = GCForest(config)
# If the model you use cost too much memory for you.
# You can use these methods to force gcforest not keeping model in memory
# gc.set_keep_model_in_mem(False), default is TRUE.
(X_train, y_train), (X_test, y_test) = mnist.load_data()
# X_train, y_train = X_train[:2000], y_train[:2000]
X_train = X_train[:, np.newaxis, :, :]
X_test = X_test[:, np.newaxis, :, :]
X_train_enc = gc.fit_transform(X_train, y_train)
# X_enc is the concatenated predict_proba result of each estimators of the last layer of the GCForest model
# X_enc.shape =
# (n_datas, n_estimators * n_classes): If cascade is provided
# (n_datas, n_estimators * n_classes, dimX, dimY): If only finegrained part is provided
# You can also pass X_test, y_test to fit_transform method, then the accracy on test data will be logged when training.
def GCForest_prediction(feature_data, result_data):
random_state = 2019
n_splits = 5
folds = StratifiedKFold(n_splits=n_splits,
shuffle=True,
random_state=random_state).split(
feature_data, result_data)
test_pred = np.zeros(feature_data.shape[0])
test_proba = np.zeros(feature_data.shape[0])
acc_scores = np.zeros(n_splits)
recall_scores = np.zeros(n_splits)
mcc_scores = np.zeros(n_splits)
f1_scores = np.zeros(n_splits)
for j, (train_idx, test_idx) in enumerate(folds):
X_train = feature_data[train_idx]
Y_train = result_data[train_idx]
X_test = feature_data[test_idx]
Y_test = result_data[test_idx]
config = get_toy_config()
gc = GCForest(config) # should be a dict
X_train_enc = gc.fit_transform(X_train, Y_train)
part_X_train_enc = X_train_enc[:, ::2]
y_pred = gc.predict(X_test)
X_test_enc = gc.transform(X_test)
part_X_test_enc = X_test_enc[:, ::2]
y_proba = gc.predict_proba(X_test)[:, 1]
acc = accuracy_score(Y_test, y_pred)
print("Test Accuracy of GcForest (save and load) = {:.2f} %".format(
acc * 100))
confmat = confusion_matrix(Y_test, y_pred)
sn = confmat[1, 1] / (confmat[1, 0] + confmat[1, 1])
sp = confmat[0, 0] / (confmat[0, 0] + confmat[0, 1])
print('1. The acc score of the model {}\n'.format(
accuracy_score(Y_test, y_pred)))
print('2. The sp score of the model {}\n'.format(sp))
print('3. The sn score of the model {}\n'.format(sn))
print('4. The mcc score of the model {}\n'.format(
matthews_corrcoef(Y_test, y_pred)))
print('9. The auc score of the model {}\n'.format(
roc_auc_score(Y_test, y_proba, average='macro')))
print('6. The recall score of the model {}\n'.format(
recall_score(Y_test, y_pred, average='macro')))
print('5. The F-1 score of the model {}\n'.format(
f1_score(Y_test, y_pred, average='macro')))
print('7. Classification report \n {} \n'.format(
classification_report(Y_test, y_pred)))
print('8. Confusion matrix \n {} \n'.format(
confusion_matrix(Y_test, y_pred)))
recall = recall_score(Y_test, y_pred, average='macro')
f1 = f1_score(Y_test, y_pred, average='macro')
acc = accuracy_score(Y_test, y_pred)
mcc = matthews_corrcoef(Y_test, y_pred)
recall_scores[j] = recall
f1_scores[j] = f1
acc_scores[j] = acc
mcc_scores[j] = mcc
test_pred[test_idx] = y_pred
test_proba[test_idx] = y_proba
print("CV- {} recall: {}, acc_score: {} , mcc_score: {}, f1_score: {}".
format(j, recall, acc, mcc, f1))
confmat = confusion_matrix(result_data, test_pred)
sn = confmat[1, 1] / (confmat[1, 0] + confmat[1, 1])
sp = confmat[0, 0] / (confmat[0, 0] + confmat[0, 1])
print(
"--------------------------------------深度森林------------------------------------"
)
print('1. The acc score of the model {}\n'.format(
accuracy_score(result_data, test_pred)))
print('2. The sp score of the model {}\n'.format(sp))
print('3. The sn score of the model {}\n'.format(sn))
print('4. The mcc score of the model {}\n'.format(
matthews_corrcoef(result_data, test_pred)))
print('9. The auc score of the model {}\n'.format(
roc_auc_score(result_data, test_proba, average='macro')))
print('6. The recall score of the model {}\n'.format(
recall_score(result_data, test_pred, average='macro')))
print('5. The F-1 score of the model {}\n'.format(
f1_score(result_data, test_pred, average='macro')))
print('7. Classification report \n {} \n'.format(
classification_report(result_data, test_pred)))
print('8. Confusion matrix \n {} \n'.format(
confusion_matrix(result_data, test_pred)))
for i in range(0, len(val_labels)):
if val_labels[i] == 0:
val_fea_0.append(val_fea[i])
else:
val_fea_1.append(val_fea[i])
test_fea = val_fea_1[:int(len(val_fea_1) /
2)] + val_fea_0[:int(len(val_fea_0) / 2)]
test_labels = [1] * int(len(val_fea_1) / 2) + [0] * int(len(val_fea_0) / 2)
train_fea = val_fea_1[int(len(val_fea_1) / 2
):] * 1 + val_fea_0[int(len(val_fea_0) / 2):] * 1
train_labels = [1] * (len(val_fea_1) - int(len(val_fea_1) / 2)) * 1 + [
0
] * (len(val_fea_0) - int(len(val_fea_0) / 2)) * 1
train_data = [[t, l] for t, l in zip(train_fea, train_labels)]
test_data = [[d, l] for d, l in zip(test_fea, test_labels)]
random.shuffle(train_data)
random.shuffle(test_data)
test_fea = [d[0] for d in test_data]
test_labels = [d[1] for d in test_data]
train_fea = [d[0] for d in train_data]
train_labels = [d[1] for d in train_data]
gc = GCForest(get_toy_config()) # should be a dict
X_train_enc = gc.fit_transform(np.array(train_fea), np.array(train_labels))
i = 0
while os.path.exists('./gcForest_model/' + str(i)):
i += 1
os.makedirs('./gcForest_model/' + str(i))
#pickle.dump(gc,open('./gcForest_model/'+ str(i)+'/model.pkl','wb+'),protocol=True)
y_pred = gc.predict(np.array(test_fea))
print(classification_report(test_labels, y_pred))
def run_classification_configuration(
X_train_10_fold, X_test_10_fold, y_train_10_fold, y_test_10_fold,
test_idx_10_fold, train_idx_10_fold, rf_tree, rf_max_depth, rf_tree_2,
rf_max_depth_2, xgb_tree, xgb_max_depth, min_child_weight, lr,
xgb_tree_2, xgb_max_depth_2, min_child_weight_2, lr_2, layer):
folds_AUC_testing, folds_AUPR_testing = [], []
folds_AUC_training, folds_AUPR_training = [], []
folds_metrics3_training, folds_metrics3_testing = [], []
test_true_predict_compare, train_true_predict_compare = [], []
folds_recall_50, folds_recall_100 = [], []
for X_train, X_test, y_train, y_test, test_idx_fold, train_idx_fold in zip(
X_train_10_fold, X_test_10_fold, y_train_10_fold, y_test_10_fold,
test_idx_10_fold, train_idx_10_fold):
config = get_toy_config(rf_tree, rf_max_depth, rf_tree_2,
rf_max_depth_2, xgb_tree, xgb_max_depth,
min_child_weight, lr, xgb_tree_2,
xgb_max_depth_2, min_child_weight_2, lr_2,
layer)
gc = GCForest(config)
print(config)
X_train_enc = gc.fit_transform(X_train, y_train)
y_pred_train = gc.predict(X_train)
y_predprob_train = gc.predict_proba(X_train)
y_pred_test = gc.predict(X_test)
y_predprob_test = gc.predict_proba(X_test)
temp = pd.DataFrame([y_test, y_predprob_test[:, 1],
y_pred_test]).T.sort_values(by=1, ascending=False)
recall_50 = precision_recall_fscore_support(temp.iloc[:50, :][0],
temp.iloc[:50, :][2],
pos_label=1,
average='binary')[1]
recall_100 = precision_recall_fscore_support(temp.iloc[:25, :][0],
temp.iloc[:25, :][2],
pos_label=1,
average='binary')[1]
test_true_predict_compare.append([
test_idx_fold, y_pred_test, y_test, y_predprob_test[:, 0],
y_predprob_test[:, 1]
]) #10-cv
train_true_predict_compare.append([
train_idx_fold, y_pred_train, y_train, y_predprob_train[:, 0],
y_predprob_train[:, 1]
]) #10-cv
precision_training, recall_training, _ = precision_recall_curve(
y_train, y_predprob_train[:, 1], pos_label=1)
precision_testing, recall_testing, _ = precision_recall_curve(
y_test, y_predprob_test[:, 1], pos_label=1)
AUPR_training, AUPR_testing = auc(recall_training,
precision_training), auc(
recall_testing,
precision_testing)
AUC_training, AUC_testing = roc_auc_score(
y_train,
y_predprob_train[:, 1]), roc_auc_score(y_test, y_predprob_test[:,
1])
metrics3_testing = precision_recall_fscore_support(
y_test, y_pred_test, pos_label=1, average='binary')[:3]
metrics3_training = precision_recall_fscore_support(
y_train, y_pred_train, pos_label=1, average='binary')[:3]
folds_AUC_testing.append(AUC_testing)
folds_AUPR_testing.append(AUPR_testing)
folds_metrics3_testing.append(metrics3_testing)
folds_AUC_training.append(AUC_training)
folds_AUPR_training.append(AUPR_training)
folds_metrics3_training.append(metrics3_training)
folds_recall_50.append(recall_50)
folds_recall_100.append(recall_100)
Avg_AUPR_training = np.mean(folds_AUPR_training)
Avg_AUPR_testing = np.mean(folds_AUPR_testing)
Avg_AUC_training = np.mean(folds_AUC_training)
Avg_AUC_testing = np.mean(folds_AUC_testing)
Avg_metrics3_training = np.mean(folds_metrics3_training, axis=0)
Avg_metrics3_testing = np.mean(folds_metrics3_testing, axis=0)
return [
Avg_AUPR_training, Avg_AUPR_testing, folds_AUPR_testing,
Avg_AUC_training, Avg_AUC_testing, folds_AUC_testing,
folds_AUPR_training, folds_AUC_training, Avg_metrics3_testing,
Avg_metrics3_training, folds_recall_50, folds_recall_100
], [test_true_predict_compare, train_true_predict_compare]
def run_classification_configuration(
X_train_10_fold, X_test_10_fold, y_train_10_fold, y_test_10_fold,
test_idx_10_fold, train_idx_10_fold, rf_tree, rf_max_depth, rf_tree_2,
rf_max_depth_2, xgb_tree, xgb_max_depth, min_child_weight, lr,
xgb_tree_2, xgb_max_depth_2, min_child_weight_2, lr_2, layer, mode,
seed):
folds_AUC_testing, folds_AUPR_testing = [], []
folds_AUC_training, folds_AUPR_training = [], []
folds_metrics3_training, folds_metrics3_testing = [], []
test_true_predict_compare, train_true_predict_compare = [], []
folds_recall_25, folds_recall_50, folds_recall_100, folds_recall_200, folds_recall_400 = [], [], [], [], []
folds_G_mean = []
i = 0
for X_train, X_test, y_train, y_test, test_idx_fold, train_idx_fold in zip(
X_train_10_fold, X_test_10_fold, y_train_10_fold, y_test_10_fold,
test_idx_10_fold, train_idx_10_fold):
config = get_toy_config(rf_tree, rf_max_depth, rf_tree_2,
rf_max_depth_2, xgb_tree, xgb_max_depth,
min_child_weight, lr, xgb_tree_2,
xgb_max_depth_2, min_child_weight_2, lr_2,
layer)
gc = GCForest(config)
X_train_enc = gc.fit_transform(X_train, y_train)
y_pred_train = gc.predict(X_train)
y_predprob_train = gc.predict_proba(X_train)
y_pred_test = gc.predict(X_test)
y_predprob_test = gc.predict_proba(X_test)
temp = pd.DataFrame([y_test, y_predprob_test[:, 1],
y_pred_test]).T.sort_values(by=1, ascending=False)
recall_25 = precision_recall_fscore_support(temp.iloc[:25, :][0],
temp.iloc[:25, :][2],
pos_label=1,
average='binary')[1]
recall_50 = precision_recall_fscore_support(temp.iloc[:50, :][0],
temp.iloc[:50, :][2],
pos_label=1,
average='binary')[1]
test_true_predict_compare.append([
test_idx_fold, y_pred_test, y_test, y_predprob_test[:, 0],
y_predprob_test[:, 1]
]) #10-cv
train_true_predict_compare.append([
train_idx_fold, y_pred_train, y_train, y_predprob_train[:, 0],
y_predprob_train[:, 1]
]) #10-cv
precision_training, recall_training, _ = precision_recall_curve(
y_train, y_predprob_train[:, 1], pos_label=1)
precision_testing, recall_testing, _ = precision_recall_curve(
y_test, y_predprob_test[:, 1], pos_label=1)
AUPR_training, AUPR_testing = auc(recall_training,
precision_training), auc(
recall_testing,
precision_testing)
AUC_training, AUC_testing = roc_auc_score(
y_train,
y_predprob_train[:, 1]), roc_auc_score(y_test, y_predprob_test[:,
1])
metrics3_testing = precision_recall_fscore_support(
y_test, y_pred_test, pos_label=1, average='binary')[:3]
metrics3_training = precision_recall_fscore_support(
y_train, y_pred_train, pos_label=1, average='binary')[:3]
tn, fp, fn, tp = confusion_matrix(y_test, y_pred_test,
labels=[0, 1]).ravel()
specificity = float(tn) / float(tn + fp)
recall = metrics3_testing[1]
G_mean = np.sqrt(recall * specificity)
folds_AUC_testing.append(AUC_testing)
folds_AUPR_testing.append(AUPR_testing)
folds_metrics3_testing.append(metrics3_testing)
folds_AUC_training.append(AUC_training)
folds_AUPR_training.append(AUPR_training)
folds_metrics3_training.append(metrics3_training)
folds_G_mean.append(G_mean)
folds_recall_25.append(recall_25)
folds_recall_50.append(recall_50)
i += 1
Avg_AUPR_training = np.mean(folds_AUPR_training)
Avg_AUPR_testing = np.mean(folds_AUPR_testing)
Avg_AUC_training = np.mean(folds_AUC_training)
Avg_AUC_testing = np.mean(folds_AUC_testing)
Avg_metrics3_training = np.mean(folds_metrics3_training, axis=0)
Avg_metrics3_testing = np.mean(folds_metrics3_testing, axis=0)
Avg_G_mean = np.mean(folds_G_mean)
return [
Avg_AUPR_training,
Avg_AUPR_testing,
folds_AUPR_testing, #012
Avg_AUC_training,
Avg_AUC_testing,
folds_AUC_testing, #345
folds_AUPR_training,
folds_AUC_training, #67
Avg_metrics3_testing,
Avg_metrics3_training, #89
folds_recall_25,
folds_recall_50,
folds_G_mean
], [test_true_predict_compare, train_true_predict_compare
] #folds_recall_100, folds_recall_200, folds_recall_400,
from sklearn import preprocessing
from sklearn.model_selection import train_test_split
from sklearn.ensemble import RandomForestClassifier
from sklearn.metrics import accuracy_score
from sklearn.model_selection import StratifiedKFold
from sklearn.metrics import roc_curve, auc
from dimension_reduction import elasticNet
import utils.tools as utils
from gcforest.gcforest import GCForest
from gcforest.utils.config_utils import load_json
start = time.time()
path1 = 'gcforest4.json'
config = load_json(path1)
gc = GCForest(config)
mask_data = sio.loadmat('yeast_elastic_mask_scale_0.03_0.1.mat')
mask = mask_data.get('yeast_elastic_mask')
extraction = sio.loadmat('yeast_feature_end.mat')
proteinA = extraction.get('feature_A')
protein_A = np.array(proteinA)
proteinB = extraction.get('feature_B')
protein_B = np.array(proteinB)
X_ = np.concatenate((protein_A, protein_B), axis=1)
X_ = np.array(X_)
[row, column] = np.shape(X_)
label_P = np.ones(int(row / 2))
label_N = np.zeros(int(row / 2))
label_ = np.hstack((label_P, label_N))
y_raw = np.mat(label_)
args = parser.parse_args()
return args
if __name__ == "__main__":
# config
args = parse_args()
if args.model == 'ca':
config = load_json('./mnist-ca.json')
elif args.model == 'gc':
config = load_json('./mnist-gc.json')
else:
config = load_json('./mnist-gc.json')
gc = GCForest(config)
# gc.set_keep_model_in_mem(False)
gc.set_keep_model_in_mem(True)
# data
data_num_train = 60000 # The number of figures
data_num_test = 10000 # test num
fig_w = 45 # width of each figure
X_train = np.fromfile("./data/mnist_train/mnist_train_data",
dtype=np.uint8)
y_train = np.fromfile("./data/mnist_train/mnist_train_label",
dtype=np.uint8)
X_test = np.fromfile("./data/mnist_test/mnist_test_data", dtype=np.uint8)
y_test = np.fromfile("./data/mnist_test/mnist_test_label", dtype=np.uint8)