def esmm_model_train(train_data_path):
# 确定嵌入层维度
cat2id = read_data_pickle("cat2id.pkl")
cat12id = read_data_pickle("cat12id.pkl")
brand2id = read_data_pickle("brand2id.pkl")
pagecat2id = read_data_pickle("pagecat2id.pkl")
cat_feat_size = len(cat2id.keys())
cat1_feat_size = len(cat12id.keys())
brand_feat_size = len(brand2id.keys())
pagecat_feat_size = len(pagecat2id.keys())
# 初始化模型
model = EsmmModel(cat_names_feature_size=cat_feat_size,
cat_names_embedding_size=16,
cat1_names_feature_size=cat1_feat_size,
cat1_names_embedding_size=8,
brand_names_feature_size=brand_feat_size,
brand_names_embedding_size=32,
pagecat_names_feature_size=pagecat_feat_size,
pagecat_names_embedding_size=64,
output_dir=get_data_path() +
"../saved_sessions/result/model/")
# 训练模型，并保存
df_train = pd.read_csv(get_data_path() + train_data_path,
dtype={
"pagecat": int,
"brand": int,
"cat": int,
"cat1": int
})
model.train(df_train, df_train)
model.clear_all()
def get_trained_model():
cat2id = read_data_pickle("cat2id.pkl")
cat12id = read_data_pickle("cat12id.pkl")
brand2id = read_data_pickle("brand2id.pkl")
pagecat2id = read_data_pickle("pagecat2id.pkl")
cat_feat_size = len(cat2id.keys())
cat1_feat_size = len(cat12id.keys())
brand_feat_size = len(brand2id.keys())
pagecat_feat_size = len(pagecat2id.keys())
# 初始化模型
model = EsmmModel(cat_names_feature_size=cat_feat_size,
cat_names_embedding_size=16,
cat1_names_feature_size=cat1_feat_size,
cat1_names_embedding_size=8,
brand_names_feature_size=brand_feat_size,
brand_names_embedding_size=32,
pagecat_names_feature_size=pagecat_feat_size,
pagecat_names_embedding_size=64,
output_dir=get_data_path() +
"../saved_sessions/result/model/")
model.build_place_holders()
model.build_dense_weight()
model.build_deep_weight()
model.build_dense_layer()
model.build_input_layer()
model.build_logit_layer()
model.restore_session()
return model
def evaluate(self, test_data):
total_sample_num = len(test_data["purchased"].values.tolist())
fd = self.feed_batch(test_data, 0, total_sample_num)
loss, purchase_pred, click_pred = self.sess.run(
[self.loss, self.purchase_predictions, self.ctr_predictions],
feed_dict=fd)
cvr_label = np.array(test_data["purchased"].values.tolist())
ctr_label = np.array(test_data["clicked"].values.tolist())
purchase_pred = [1 if label >= 0.5 else 0 for label in purchase_pred]
click_pred = [1 if label >= 0.5 else 0 for label in click_pred]
cvr_label = [int(label) for label in cvr_label]
ctr_label = [int(label) for label in ctr_label]
log_handler.log.info("actually purchased: %s" % sum(cvr_label))
log_handler.log.info("actually clicked: %s" % sum(ctr_label))
log_handler.log.info("predict purchased: %s" % sum(purchase_pred))
log_handler.log.info("predict clicked: %s" % sum(click_pred))
test_data["purchased_pred"] = np.array(purchase_pred)
test_data["clicked_pred"] = np.array(click_pred)
test_data[[
"goodssn", "keyword", "expose_total", "click_total",
"purchase_total", "click_rate", "purchase_rate", "clicked",
"purchased", "clicked_pred", "purchased_pred"
]].to_csv(get_data_path() + "final_result.csv")
purchase_pred_score = self.eval_metric(cvr_label, purchase_pred)
click_pred_score = self.eval_metric(ctr_label, click_pred)
return loss, purchase_pred_score, click_pred_score
def model_train():
try:
log_handler.log.info(
"----------------Training Esmm Model----------------")
rmdir_if_exist(get_data_path() + "../saved_sessions/result/model")
esmm_model_train(FEATURE_TRAIN_SOURCE)
log_handler.log.info(
"----------------Finish training Esmm Model----------------")
time.sleep(5)
except Exception:
log_handler.log.info(
"----------------Error training Esmm Model----------------")
log_handler.log.info(str(excp_trace()))
raise Exception
def predict(self, test_data, result_file):
total_sample_num = len(test_data["pagecat"].values.tolist())
test_data["purchased"] = np.zeros(total_sample_num, dtype="float")
test_data["clicked"] = np.zeros(total_sample_num, dtype="float")
fd = self.feed_batch(test_data, 0, total_sample_num)
purchase_prob, click_prob = self.sess.run(
[self.purchase_predictions, self.ctr_predictions], feed_dict=fd)
test_data["click_prob"] = click_prob
test_data["purchase_prob"] = purchase_prob
test_data["overall_prob"] = test_data["purchase_prob"].multiply(
test_data["click_prob"])
test_data = test_data[[
"goodssn", "keyword", "purchase_prob", "click_prob", "overall_prob"
]]
pd.to_pickle(test_data, get_data_path() + result_file)