def bl_IsRec_best(a_dataset):
model_name = 'IsRec_best' # 'IsRec' 'IsRec_best_modified'
epoch_num = 20
neighbor_size = 15
topTopicNum = 3
cluster_mode = 'LDA'
cluster_mode_topic_nums = [50] # 10,25,75,,100,125,150
train_data, test_data = get_train_test_data(a_dataset.train_data,
a_dataset.test_data)
for cluster_mode_topic_num in cluster_mode_topic_nums:
HINRec_model = HINRec_new(
model_name=model_name,
semantic_mode='TF_IDF',
epoch_num=epoch_num,
neighbor_size=neighbor_size,
topTopicNum=topTopicNum,
cluster_mode=cluster_mode,
cluster_mode_topic_num=cluster_mode_topic_num)
if os.path.exists(HINRec_model.weight_path):
print('have trained,return!')
else:
HINRec_model.train(test_data)
HINRec_model.save_model()
evalute_by_epoch(HINRec_model,
HINRec_model,
HINRec_model.model_name,
test_data,
evaluate_by_slt_apiNum=True) # )
def bl_PasRec(a_dataset):
model_name = 'PasRec_2path' # 'PasRec_2path'
epoch_num = 20 # 之前是40 40比20差点
neighbor_size = 15
topTopicNum = 3
train_data, test_data = get_train_test_data(a_dataset.train_data,
a_dataset.test_data)
HINRec_model = HINRec_new(model_name=model_name,
epoch_num=epoch_num,
neighbor_size=neighbor_size,
topTopicNum=topTopicNum)
# 使用LDA处理PasRec的相似度 50 100 150
# HINRec_model = HINRec_new(model_name=model_name, semantic_mode='LDA', LDA_topic_num=50, epoch_num=epoch_num,
# neighbor_size=neighbor_size,
# topTopicNum=topTopicNum)
if os.path.exists(HINRec_model.weight_path):
print('have trained,return!')
else:
# 这里是每隔20epoch测试一下，所以train中输入test_data
HINRec_model.train(test_data)
HINRec_model.save_model()
evalute_by_epoch(
HINRec_model,
HINRec_model,
HINRec_model.model_name,
test_data,
evaluate_by_slt_apiNum=True) # ,if_save_recommend_result=True)
def bl_IsRec(a_dataset):
model_name = 'IsRec' # ''
epoch_nums = [20] # 15,100,1000
neighbor_size = 15
topTopicNums = [3] # [3,4,5,6]
train_data, test_data = get_train_test_data(a_dataset.train_data,
a_dataset.test_data)
for epoch_num in epoch_nums:
for topTopicNum in topTopicNums:
HINRec_model = HINRec_new(model_name=model_name,
epoch_num=epoch_num,
neighbor_size=neighbor_size,
topTopicNum=topTopicNum)
if os.path.exists(HINRec_model.weight_path):
print('have trained,return!')
else:
HINRec_model.train(test_data)
# HINRec_model.test_model(test_data)
HINRec_model.save_model()
evalute_by_epoch(HINRec_model,
HINRec_model,
HINRec_model.model_name,
test_data,
evaluate_by_slt_apiNum=True
) # ,if_save_recommend_result=True)
def bl_DHSR(a_dataset):
dhsr_recommend_model = DHSR_model()
dhsr_model = dhsr_recommend_model.get_model()
# a_dataset.transfer() # 将重复sample删除？ 'newScene'且need_slt_apis=False时
train_data, test_data = get_train_test_data(a_dataset.train_data,
a_dataset.test_data)
dhsr_model = load_preTrained_model(
dhsr_recommend_model, dhsr_model, train_data, test_data,
*new_Para.param.train_paras) # 'monitor loss&acc'
dhsr_recommend_model.save_sth()
evalute_by_epoch(
dhsr_recommend_model, dhsr_model, dhsr_recommend_model.model_name,
test_data
) # ,if_save_recommend_result=True,evaluate_by_slt_apiNum = True)
def bl_DHSR_new(a_dataset):
train_datas, test_datas = a_dataset.transfer_false_test_DHSR(
if_reduct_train=True) # 是否约减训练集
# 选择的服务数目不同，训练对应的模型，并评估效果
for slt_num in range(1, new_Para.param.slt_item_num + 1):
train_data, test_data = train_datas[slt_num - 1], test_datas[slt_num -
1]
# old_new = 'new','new_sigmoid', 'new_reduct'效果最好
dhsr_recommend_model = DHSR_model(old_new='new_reduct',
slt_num=slt_num)
dhsr_model = dhsr_recommend_model.get_model()
dhsr_model = load_preTrained_model(
dhsr_recommend_model, dhsr_model, train_data, test_data,
*new_Para.param.train_paras) # 'monitor loss&acc'
evalute_by_epoch(dhsr_recommend_model,
dhsr_model,
dhsr_recommend_model.model_name,
test_data,
evaluate_by_slt_apiNum=True)
dhsr_recommend_model.save_sth()
print('DHSR, slt_num:{}, train_predict,done!'.format(slt_num))
def DINRec(a_dataset, new_old='new'):
train_data, test_data = a_dataset.train_data, a_dataset.test_data
CI_recommend_model = CI_Model(new_old) # 'old'
CI_recommend_model.prepare()
CI_model_obj = CI_recommend_model.get_model()
CI_model_obj = load_preTrained_model(
CI_recommend_model, CI_model_obj, train_data, test_data,
*new_Para.param.train_paras
) # ,true_candidates_dict=HINRec_model.get_true_candi_apis() 'monitor loss&acc'
DINRec_model = DIN_Rec(CI_recommend_model,
new_Para.param.predict_fc_unit_nums)
DINRec_model.prepare()
DINRec_model_obj = DINRec_model.get_model()
DINRec_model_obj = load_preTrained_model(
DINRec_model, DINRec_model_obj, train_data, test_data,
*new_Para.param.train_paras
) # ,true_candidates_dict=HINRec_model.get_true_candi_apis() 'monitor loss&acc'
evalute_by_epoch(DINRec_model,
DINRec_model_obj,
DINRec_model.simple_name,
test_data,
if_save_recommend_result=True,
evaluate_by_slt_apiNum=True)
def train_save_by_early_stop(recommend_model, model, train_data, test_data):
"""
训练时按照验证集的loss，early stopping得到最优的模型；最后基于该模型测试
:return:
"""
if_Train = True if new_Para.param.pairwise else False
train_labels = train_data[-1]
train_instances_tuple = recommend_model.get_instances(
*train_data[:-1], test_phase_flag=if_Train)
train_model = recommend_model.get_pairwise_model(
) if new_Para.param.pairwise else model
if new_Para.param.pairwise:
train_model.compile(optimizer=recommend_model.optimizer,
loss=lambda y_true, y_pred: y_pred,
metrics=['accuracy'])
else:
train_model.compile(optimizer=recommend_model.optimizer,
loss='binary_crossentropy',
metrics=['accuracy'])
early_stopping = EarlyStopping(monitor='val_loss',
patience=10,
verbose=2,
mode='min')
hist = train_model.fit([*train_instances_tuple],
train_labels,
epochs=new_Para.param.num_epochs,
batch_size=new_Para.param.small_batch_size,
callbacks=[early_stopping],
validation_split=new_Para.param.validation_split,
shuffle=True) #
model.save_weights(
dataset.crt_ds.new_model_para_path.format(recommend_model.model_dir,
'min_loss')) # !!! 改正
model_name = recommend_model.get_simple_name() + recommend_model.get_name(
) + '_min_loss'
save_loss_acc(hist, model_name, if_multi_epoch=True)
epoch_evaluate_result = evalute_by_epoch(recommend_model, model,
model_name, test_data)
return model
def test_model(self, test_data):
evalute_by_epoch(self, self, self.model_name, test_data)
def train_save_best_NDCG_model(recommend_model,
model,
train_data,
test_data,
start_epoch=0,
true_candidates_dict=None,
CI_start_test_epoch=2,
earlyStop_epochs=5):
"""
训练多个epoch，每个之后均测试，选择并返回NDCG等最终指标最优的模型
:param recommend_model: 整体的推荐模型
:param model: model_core
:param train_data:
:param test_data:
:param start_epoch: 之前该模型已经训练过多个epoch，在这个基础上接着训练
:param true_candidates_dict: ？
:return:
"""
print('training_save_best_NDCG_model...')
epoch_evaluate_results = []
train_labels = train_data[-1]
if new_Para.param.final_activation == 'softmax': # 针对softmax变换labels
train_labels = utils.to_categorical(train_labels, num_classes=2)
# ??? 跟获取的实例有关
test_phase_flag = False if (new_Para.param.pairwise
or new_Para.param.NI_OL_mode == 'OL_GE'
or new_Para.param.train_mashup_best
) else True # 在线NI时，也要区分训练和测试过程得到隐式表示的方法
train_instances_tuple = recommend_model.get_instances(
*train_data[:-1],
test_phase_flag=test_phase_flag) # 有必要才获取负例:pairwise的训练
# 读取之前训练过的最优指标
if start_epoch > 0:
with open(
dataset.crt_ds.new_best_epoch_path.format(
recommend_model.get_simple_name()), 'r') as f:
best_epoch = int(f.read().strip())
with open(
dataset.crt_ds.new_best_NDCG_path.format(
recommend_model.get_simple_name()), 'r') as f:
best_NDCG_5 = float(f.read().strip())
else:
best_epoch, best_NDCG_5 = 0, 0
# 有必要转化为pairwise模型
train_model = recommend_model.get_pairwise_model(
) if new_Para.param.pairwise else model
for epoch in range(new_Para.param.num_epochs - start_epoch):
if start_epoch == 0 and epoch == 0: # 首次训练要编译
if new_Para.param.pairwise:
train_model.compile(optimizer=recommend_model.optimizer,
loss=lambda y_true, y_pred: y_pred,
metrics=['accuracy'])
else:
train_model.compile(optimizer=recommend_model.optimizer,
loss='binary_crossentropy',
metrics=['accuracy'])
print('model compile,done!')
if start_epoch > 0: # 载入原模型，直接在原来的基础上训练
train_model = load_trained_model(recommend_model, model)
epoch = epoch + start_epoch
print('Epoch {}'.format(epoch))
# test_model = model if not new_Para.param.pairwise else recommend_model.get_single_model() # pairwise时需要复用相关参数!!!
if type(train_instances_tuple) == tuple:
hist = train_model.fit(
[*train_instances_tuple],
np.array(train_labels),
batch_size=new_Para.param.batch_size,
epochs=1,
verbose=2,
shuffle=True,
validation_split=new_Para.param.validation_split
) #可以观察过拟合欠拟合 ,validation_split=0.1
else:
hist = train_model.fit(
train_instances_tuple,
np.array(train_labels),
batch_size=new_Para.param.batch_size,
epochs=1,
verbose=2,
shuffle=True,
validation_split=new_Para.param.validation_split)
print('model train,done!')
# 记录：数据集情况，模型新旧完整，模型架构，训练设置
# recommend_model.get_simple_name()+ '---'+
model_name = dataset.crt_ds.data_name + recommend_model.get_name(
) + new_Para.param.train_name if epoch == 0 else '' # 记录在测试集的效果，写入evalute.csv
# 每个epoch的测试
save_loss_acc(hist, model_name, epoch=epoch)
if not os.path.exists(recommend_model.model_dir):
os.makedirs(recommend_model.model_dir)
if isinstance(recommend_model, CI_Model) and not isinstance(
recommend_model, NI_Model_online):
first_test_epoch = CI_start_test_epoch # 前3轮效果差，一般不用测
else:
first_test_epoch = 0
if epoch < first_test_epoch: # 暂不测试，提高速度
epoch_evaluate_results.append(None)
continue
if epoch == first_test_epoch: # 记录第一个epoch的测试时间
with open(new_Para.param.time_path, 'a+') as f1:
f1.write(recommend_model.get_simple_name())
f1.write('\n')
# test_model = model if not new_Para.param.pairwise else recommend_model.get_single_model() # pairwise时需要复用相关参数!!!
# 没必要使用get_model再获取，传入的model是对象引用，pairwise更新后model也变化
# 每个epoch的测试
epoch_evaluate_result = evalute_by_epoch(
recommend_model,
model,
model_name,
test_data,
record_time=True if epoch == 1 else False,
true_candidates_dict=true_candidates_dict)
epoch_evaluate_results.append(epoch_evaluate_result)
if epoch_evaluate_result[0][3] >= best_NDCG_5: # 优于目前的best_NDCG_5才存储
best_NDCG_5 = epoch_evaluate_result[0][3]
best_epoch = epoch
model.save_weights(
dataset.crt_ds.new_model_para_path.format(
recommend_model.model_dir, epoch)) # 记录该epoch下的模型参数***
else:
if epoch - best_epoch >= earlyStop_epochs: # 大于若干个epoch，效果没有提升，即时终止
break
#@@@# # 第一个 epoch之后存储HIN_sim对象？？？ 删去only_MLP_model的判断，换成了CI? NI为什么要记录？or isinstance(recommend_model,NI_Model)
# if epoch==0 and (isinstance(recommend_model,gx_text_tag_continue_only_MLP_model) ):
# recommend_model.save_HIN_sim()
# 看word embedding矩阵是否发生改变，尤其是padding的0
# print('some embedding parameters after {} epoch:'.format(epoch))
# print (recommend_model.embedding_layer.get_weights ()[0][:2])
# 记录最优epoch和最优NDCG@5
with open(
dataset.crt_ds.new_best_epoch_path.format(
recommend_model.model_dir), 'w') as f:
f.write(str(best_epoch))
with open(
dataset.crt_ds.new_best_NDCG_path.format(
recommend_model.model_dir), 'w') as f:
f.write(str(best_NDCG_5))
print('best epoch:{},best NDCG@5:{}'.format(best_epoch, best_NDCG_5))
# 记录最优指标
csv_table_name = 'best_indicaters\n' # 命名格式！！！
summary(new_Para.param.evaluate_path, csv_table_name,
epoch_evaluate_results[best_epoch], new_Para.param.topKs)
# 把记录的非最优的epoch模型参数都删除
try:
for i in range(new_Para.param.num_epochs):
temp_path = dataset.crt_ds.new_model_para_path.format(
recommend_model.model_dir, i)
if i != best_epoch and os.path.exists(temp_path):
os.remove(temp_path)
model.load_weights(
dataset.crt_ds.new_model_para_path.format(
recommend_model.model_dir, best_epoch))
finally:
return model