代码+实战：TensorFlow Estimator of Deep CTR —— DeepFM/NFM/AFM/FNN/PNN

2018 年 2 月 17 日 AI研习社 lambdaJi

AI 研习社按，本文作者 lambdaJi，本文首发于知乎，AI 研习社获其授权转载。

深度学习在 ctr 预估领域的应用越来越多，新的模型不断冒出。从 ctr 预估问题看看 f(x) 设计—DNN 篇（https://zhuanlan.zhihu.com/p/28202287）整理了各模型之间的联系之后，一直在琢磨这些东西如何在工业界落地。经过几个月的调研，发现目前存在的一些问题：

开源的实现基本都是学术界的人在搞，距离工业应用还有较大的鸿沟

模型实现大量调用底层 API，各版本实现千差万别，代码臃肿难懂，迁移成本较高

单机，放到工业场景下跑不动

针对存在的问题做了一些探索，摸索出一套可行方案，有以下特性：

读数据采用 Dataset API，支持 parallel and prefetch 读取

通过 Estimator model_fn 来实现 f(x)，迁移到其他算法非常方便，只需要改写 model_fn f(x) 部分

支持分布式以及单机多线程训练

支持 export model，然后用 TensorFlow Serving 提供线上预测服务

按工业界的套路，完整的机器学习项目应该包含五个部分：特征框架，训练框架，服务框架，评估框架和监控框架，这里只讨论前三个框架。

特征框架 -- logs in，samples out

实验数据集用 criteo，特征工程参考: https://github.com/PaddlePaddle/models/blob/develop/deep_fm/preprocess.py

#1 连续特征剔除异常值/归一化
#2 离散特征剔掉低频，然后统一编码（特征编码需要保存下来，线上预测的时候要用到）

对大规模离散特征建模是用 DNN 做 ctr 预估的优势，paper 关注点大都放在 ID 类特征如何做 embedding 上，至于连续特征如何处理很少讨论，大概有以下 3 种方式：

--不做 embedding

   |1--concat[continuous, emb_vec] 做 fc

--做 embedding

   |2--离散化之后 embedding

   |3--类似 FM 二阶部分, 统一做 embedding，离散特征 val=1.0

为了模型设计上的简单统一，采用第 3 种方式，感兴趣的读者可是试试前两种的效果。

训练框架 -- samples in，model out

目前实现了 DeepFM/wide_n_deep/NFM/AFM/FNN/PNN 几个算法。以 DeepFM 为例来看看如何使用 TensorFlow Estimator and Datasets API 来实现 input_fn and model_fn：

#1 1:0.5 2:0.03519 3:1 4:0.02567 7:0.03708 8:0.01705 9:0.06296 10:0.18185 11:0.02497 12:1 14:0.02565 15:0.03267 17:0.0247 18:0.03158 20:1 22:1 23:0.13169 24:0.02933 27:0.18159 31:0.0177 34:0.02888 38:1 51:1 63:1 132:1 164:1 236:1def input_fn(filenames, batch_size=32, num_epochs=1, perform_shuffle=False):
   print('Parsing', filenames)
   def decode_libsvm(line):
       columns = tf.string_split([line], ' ')
       labels = tf.string_to_number(columns.values[0], out_type=tf.float32)
       splits = tf.string_split(columns.values[1:], ':')
       id_vals = tf.reshape(splits.values,splits.dense_shape)
       feat_ids, feat_vals = tf.split(id_vals,num_or_size_splits=2,axis=1)
       feat_ids = tf.string_to_number(feat_ids, out_type=tf.int32)
       feat_vals = tf.string_to_number(feat_vals, out_type=tf.float32)
       return {"feat_ids": feat_ids, "feat_vals": feat_vals}, labels

   # Extract lines from input files using the Dataset API, can pass one filename or filename list
   dataset = tf.data.TextLineDataset(filenames).map(decode_libsvm, num_parallel_calls=10).prefetch(500000)    # multi-thread pre-process then prefetch

   # Randomizes input using a window of 256 elements (read into memory)
   if perform_shuffle:
       dataset = dataset.shuffle(buffer_size=256)

   # epochs from blending together.
   dataset = dataset.repeat(num_epochs)
   dataset = dataset.batch(batch_size) # Batch size to use

   iterator = dataset.make_one_shot_iterator()
   batch_features, batch_labels = iterator.get_next()
   return batch_features, batch_labels

def model_fn(features, labels, mode, params):
   """Bulid Model function f(x) for Estimator."""
   #------hyperparameters----
   field_size = params["field_size"]
   feature_size = params["feature_size"]
   embedding_size = params["embedding_size"]
   l2_reg = params["l2_reg"]
   learning_rate = params["learning_rate"]
   layers = map(int, params["deep_layers"].split(','))
   dropout = map(float, params["dropout"].split(','))

   #------bulid weights------
   FM_B = tf.get_variable(name='fm_bias', shape=[1], initializer=tf.constant_initializer(0.0))
   FM_W = tf.get_variable(name='fm_w', shape=[feature_size], initializer=tf.glorot_normal_initializer())
   FM_V = tf.get_variable(name='fm_v', shape=[feature_size, embedding_size], initializer=tf.glorot_normal_initializer())

   #------build feaure-------
   feat_ids  = features['feat_ids']
   feat_ids = tf.reshape(feat_ids,shape=[-1,field_size])
   feat_vals = features['feat_vals']
   feat_vals = tf.reshape(feat_vals,shape=[-1,field_size])

   #------build f(x)------
   with tf.variable_scope("First-order"):
       feat_wgts = tf.nn.embedding_lookup(FM_W, feat_ids) # None * F * 1
       y_w = tf.reduce_sum(tf.multiply(feat_wgts, feat_vals),1)

   with tf.variable_scope("Second-order"):
       embeddings = tf.nn.embedding_lookup(FM_V, feat_ids) # None * F * K
       feat_vals = tf.reshape(feat_vals, shape=[-1, field_size, 1])
       embeddings = tf.multiply(embeddings, feat_vals) #vij*xi
       sum_square = tf.square(tf.reduce_sum(embeddings,1))
       square_sum = tf.reduce_sum(tf.square(embeddings),1)
       y_v = 0.5*tf.reduce_sum(tf.subtract(sum_square, square_sum),1)  # None * 1

   with tf.variable_scope("Deep-part"):
       if FLAGS.batch_norm:
           if mode == tf.estimator.ModeKeys.TRAIN:
               train_phase = True
           else:
               train_phase = False

       deep_inputs = tf.reshape(embeddings,shape=[-1,field_size*embedding_size]) # None * (F*K)
       for i in range(len(layers)):
           #if FLAGS.batch_norm:
           #    deep_inputs = batch_norm_layer(deep_inputs, train_phase=train_phase, scope_bn='bn_%d' %i)
               #normalizer_params.update({'scope': 'bn_%d' %i})
           deep_inputs = tf.contrib.layers.fully_connected(inputs=deep_inputs, num_outputs=layers[i], \                #normalizer_fn=normalizer_fn, normalizer_params=normalizer_params, \
               weights_regularizer=tf.contrib.layers.l2_regularizer(l2_reg), scope='mlp%d' % i)
           if FLAGS.batch_norm:
               deep_inputs = batch_norm_layer(deep_inputs, train_phase=train_phase, scope_bn='bn_%d' %i)   #放在RELU之后 https://github.com/ducha-aiki/caffenet-benchmark/blob/master/batchnorm.md#bn----before-or-after-relu
           if mode == tf.estimator.ModeKeys.TRAIN:
               deep_inputs = tf.nn.dropout(deep_inputs, keep_prob=dropout[i])                              #Apply Dropout after all BN layers and set dropout=0.8(drop_ratio=0.2)
               #deep_inputs = tf.layers.dropout(inputs=deep_inputs, rate=dropout[i], training=mode == tf.estimator.ModeKeys.TRAIN)

       y_deep = tf.contrib.layers.fully_connected(inputs=deep_inputs, num_outputs=1, activation_fn=tf.identity, \                weights_regularizer=tf.contrib.layers.l2_regularizer(l2_reg), scope='deep_out')
       y_d = tf.reshape(y_deep,shape=[-1])

   with tf.variable_scope("DeepFM-out"):
       #y_bias = FM_B * tf.ones_like(labels, dtype=tf.float32)  # None * 1  warning;这里不能用label，否则调用predict/export函数会出错，train/evaluate正常；初步判断estimator做了优化，用不到label时不传
       y_bias = FM_B * tf.ones_like(y_d, dtype=tf.float32)     # None * 1
       y = y_bias + y_w + y_v + y_d
       pred = tf.sigmoid(y)

   predictions={"prob": pred}
   export_outputs = {tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY: tf.estimator.export.PredictOutput(predictions)}
   # Provide an estimator spec for `ModeKeys.PREDICT`
   if mode == tf.estimator.ModeKeys.PREDICT:
       return tf.estimator.EstimatorSpec(mode=mode,predictions=predictions,export_outputs=export_outputs)

   #------bulid loss------
   loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=y, labels=labels)) + \        l2_reg * tf.nn.l2_loss(FM_W) + l2_reg * tf.nn.l2_loss(FM_V)

   # Provide an estimator spec for `ModeKeys.EVAL`
   eval_metric_ops = {
       "auc": tf.metrics.auc(labels, pred)
   }
   if mode == tf.estimator.ModeKeys.EVAL:
       return tf.estimator.EstimatorSpec(mode=mode,predictions=predictions,loss=loss,eval_metric_ops=eval_metric_ops)

   #------bulid optimizer------
   if FLAGS.optimizer == 'Adam':
       optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate, beta1=0.9, beta2=0.999, epsilon=1e-8)
   elif FLAGS.optimizer == 'Adagrad':
       optimizer = tf.train.AdagradOptimizer(learning_rate=learning_rate, initial_accumulator_value=1e-8)
   elif FLAGS.optimizer == 'Momentum':
       optimizer = tf.train.MomentumOptimizer(learning_rate=learning_rate, momentum=0.95)
   elif FLAGS.optimizer == 'ftrl':
       optimizer = tf.train.FtrlOptimizer(learning_rate)

   train_op = optimizer.minimize(loss, global_step=tf.train.get_global_step())

   # Provide an estimator spec for `ModeKeys.TRAIN` modes
   if mode == tf.estimator.ModeKeys.TRAIN:
       return tf.estimator.EstimatorSpec(mode=mode,predictions=predictions,loss=loss,train_op=train_op)

封装成 estimator 之后，调用非常简单

#train

python DeepFM.py --task_type=train --learning_rate=0.0005 --optimizer=Adam --num_epochs=1 --batch_size=256 --field_size=39 --feature_size=117581 --deep_layers=400,400,400 --dropout=0.5,0.5,0.5 --log_steps=1000 --num_threads=8 --model_dir=./model_ckpt/criteo/DeepFM/ --data_dir=../../data/criteo/



#predict

python DeepFM.py --task_type=infer --learning_rate=0.0005 --optimizer=Adam --num_epochs=1 --batch_size=256 --field_size=39 --feature_size=117581 --deep_layers=400,400,400 --dropout=0.5,0.5,0.5 --log_steps=1000 --num_threads=8 --model_dir=./model_ckpt/criteo/DeepFM/ --data_dir=../../data/criteo/

完整代码: lambdaji/tf_repos

https://github.com/lambdaji/tf_repos/tree/master/deep_ctr/Model_pipeline

服务框架 -- request in，pctr out

TensorFlow Serving 是一个用于机器学习模型 serving 的高性能开源库。它可以将训练好的机器学习模型部署到线上，使用 gRPC 作为接口接受外部调用。更加让人眼前一亮的是，它支持模型热更新与自动模型版本管理。这意味着一旦部署 TensorFlow Serving 后，你再也不需要为线上服务操心，只需要关心你的线下模型训练。

首先要导出 TF-Serving 能识别的模型文件

python DeepFM.py --task_type=export --learning_rate=0.0005 --optimizer=Adam --batch_size=256 --field_size=39 --feature_size=117581 --deep_layers=400,400,400 --dropout=0.5,0.5,0.5 --log_steps=1000 --num_threads=8 --model_dir=./model_ckpt/criteo/DeepFM/ --servable_model_dir=./servable_model/

默认以时间戳来管理版本，生成文件如下：

$ ls -lh servable_model/1517971230
|--saved_model.pb
|--variables
 |--variables.data-00000-of-00001
 |--variables.index

然后写一个client发送请求，这里用C++来写

PredictRequest predictRequest;

PredictResponse response;

ClientContext context;



predictRequest.mutable_model_spec()->set_name(model_name);

predictRequest.mutable_model_spec()->set_signature_name(model_signature_name); //serving_default

google::protobuf::Map<tensorflow::string, tensorflow::TensorProto>& inputs = *predictRequest.mutable_inputs();



//feature to tfrequest

std::vector<long>  ids_vec = {1,2,3,4,5,6,7,8,9,10,11,12,13,15,555,1078,17797,26190,26341,28570,35361,35613,35984,48424,51364,64053,65964,66206,71628,84088,84119,86889,88280,88283,100288,100300,102447,109932,111823};

std::vector<float> vals_vec = {0.05,0.006633,0.05,0,0.021594,0.008,0.15,0.04,0.362,0.1,0.2,0,0.04,

1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1};

tensorflow::TensorProto feat_ids;

for (uint32_t i = 0; i < ids_vec.size(); i++) {
feat_ids.add_int64_val(ids_vec[i]);

}

feat_ids.mutable_tensor_shape()->add_dim()->set_size(1);  //batch_size

feat_ids.mutable_tensor_shape()->add_dim()->set_size(feat_ids.int64_val_size());

feat_ids.set_dtype(tensorflow::DataType::DT_INT64);

inputs["feat_ids"] = feat_ids;



tensorflow::TensorProto feat_vals;

for (uint32_t i = 0; i < vals_vec.size(); i++) {
feat_vals.add_float_val(vals_vec[i]);

}

feat_vals.mutable_tensor_shape()->add_dim()->set_size(1);  //batch_size

feat_vals.mutable_tensor_shape()->add_dim()->set_size(feat_vals.float_val_size());  //sample size

feat_vals.set_dtype(tensorflow::DataType::DT_FLOAT);

inputs["feat_vals"] = feat_vals;



Status status = _stub->Predict(&context, predictRequest, &response);