代碼+實戰：TensorFlow Estimator of Deep CTR——DeepFM/NFM/AFM/FNN/PNN

本文作者：汪思穎

2018-02-26 14:42

導語：本文討論了特征框架，訓練框架，服務框架這三個階段的具體實現。

雷鋒網 AI 研習社按，本文作者 lambdaJi，本文首發于知乎，雷鋒網 AI 研習社獲其授權轉載。

深度學習在 ctr 預估領域的應用越來越多，新的模型不斷冒出。從 ctr 預估問題看看 f(x) 設計—DNN 篇整理了各模型之間的聯系之后，一直在琢磨這些東西如何在工業界落地。經過幾個月的調研，發現目前存在的一些問題：

開源的實現基本都是學術界的人在搞，距離工業應用還有較大的鴻溝
模型實現大量調用底層 API，各版本實現千差萬別，代碼臃腫難懂，遷移成本較高
單機，放到工業場景下跑不動

針對存在的問題做了一些探索，摸索出一套可行方案，有以下特性：

讀數據采用 Dataset API，支持 parallel and prefetch 讀取
通過 Estimator model_fn 來實現 f(x)，遷移到其他算法非常方便，只需要改寫 model_fn f(x) 部分
支持分布式以及單機多線程訓練
支持 export model，然后用 TensorFlow Serving 提供線上預測服務

按工業界的套路，完整的機器學習項目應該包含五個部分：特征框架，訓練框架，服務框架，評估框架和監控框架，這里只討論前三個框架。

特征框架 -- logs in，samples out

實驗數據集用 criteo，特征工程參考: https://github.com/PaddlePaddle/models/blob/develop/deep_fm/preprocess.py

#1 連續特征剔除異常值/歸一化
#2 離散特征剔掉低頻，然后統一編碼（特征編碼需要保存下來，線上預測的時候要用到）

對大規模離散特征建模是用 DNN 做 ctr 預估的優勢，paper 關注點大都放在 ID 類特征如何做 embedding 上，至于連續特征如何處理很少討論，大概有以下 3 種方式：

--不做 embedding
   |1--concat[continuous, emb_vec] 做 fc
--做 embedding
   |2--離散化之后 embedding
   |3--類似 FM 二階部分, 統一做 embedding，離散特征 val=1.0

為了模型設計上的簡單統一，采用第 3 種方式，感興趣的讀者可是試試前兩種的效果。

訓練框架 -- samples in，model out

目前實現了 DeepFM/wide_n_deep/NFM/AFM/FNN/PNN 幾個算法。以 DeepFM 為例來看看如何使用 TensorFlow Estimator and Datasets API 來實現 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;這里不能用label，否則調用predict/export函數會出錯，train/evaluate正常；初步判斷estimator做了優化，用不到label時不傳
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)

封裝成 estimator 之后，調用非常簡單

#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/

完整代碼: lambdaji/tf_repos

服務框架 -- request in，pctr out

TensorFlow Serving 是一個用于機器學習模型 serving 的高性能開源庫。它可以將訓練好的機器學習模型部署到線上，使用 gRPC 作為接口接受外部調用。更加讓人眼前一亮的是，它支持模型熱更新與自動模型版本管理。這意味著一旦部署 TensorFlow Serving 后，你再也不需要為線上服務操心，只需要關心你的線下模型訓練。

首先要導出 TF-Serving 能識別的模型文件

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/

默認以時間戳來管理版本，生成文件如下：

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

然后寫一個client發送請求，這里用C++來寫

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);