1.Introduction
This's the tool for CTR, including FM, FFM, NFFM, XdeepFM and so on.
Note: only implement FM, FFM and NFFM. More detail and another models will be implemented
2.Requirements
- python3
- sklearn
- TensorFlow>=1.6
3.Kernel for NFFM
You can find kaggle kernel for NFFM in the following link: https://www.kaggle.com/guoday/nffm-baseline-0-690-on-lb
4.Kernel for Xdeepfm
You can find kaggle kernel of Xdeepfm in the following link: https://www.kaggle.com/guoday/xdeepfm-baseline
5.Quick Start
Loading dataset
import pandas as pd
import numpy as np
import tensorflow as tf
import ctrNet
from sklearn.model_selection import train_test_split
from src import misc_utils as utils
import os
train_df=pd.read_csv('data/train_small.txt',header=None,sep='\t')
train_df.columns=['label']+['f'+str(i) for i in range(39)]
train_df, dev_df,_,_ = train_test_split(train_df,train_df,test_size=0.1, random_state=2019)
dev_df, test_df,_,_ = train_test_split(dev_df,dev_df,test_size=0.5, random_state=2019)
features=['f'+str(i) for i in range(39)]Creating hparams
hparam=tf.contrib.training.HParams(
model='ffm', #['fm','ffm','nffm']
k=16,
hash_ids=int(1e5),
batch_size=64,
optimizer="adam", #['adadelta','adagrad','sgd','adam','ftrl','gd','padagrad','pgd','rmsprop']
learning_rate=0.0002,
num_display_steps=100,
num_eval_steps=1000,
epoch=3,
metric='auc', #['auc','logloss']
init_method='uniform', #['tnormal','uniform','normal','xavier_normal','xavier_uniform','he_normal','he_uniform']
init_value=0.1,
feature_nums=len(features))
utils.print_hparams(hparam)Building model
os.environ["CUDA_DEVICE_ORDER"]='PCI_BUS_ID'
os.environ["CUDA_VISIBLE_DEVICES"]='0'
model=ctrNet.build_model(hparam)Training model
#You can use control-c to stop training if the model doesn't improve.
model.train(train_data=(train_df[features],train_df['label']),\
dev_data=(dev_df[features],dev_df['label']))Testing model
from sklearn import metrics
preds=model.infer(dev_data=(test_df[features],test_df['label']))
fpr, tpr, thresholds = metrics.roc_curve(test_df['label']+1, preds, pos_label=2)
auc=metrics.auc(fpr, tpr)
print(auc)