add infer net define

fleetrec/examples/user_define_trainer.py

@@ -28,7 +28,6 @@ class UserDefineTrainer(TranspileTrainer):
         self.regist_context_processor('train_pass', self.train)
 
     def init(self, context):
-        self.model.input()
         self.model.net()
         self.model.metrics()
         self.model.avg_loss()

fleetrec/models/base.py

@@ -63,11 +63,11 @@ def create(config):
     model = None
     if config['mode'] == 'fluid':
         model = YamlModel(config)
-        model.net()
+        model.train_net()
     return model
 
 
-class Model(object):
+class ModelBase(object):
     """R
     """
     __metaclass__ = abc.ABCMeta
@@ -80,6 +80,9 @@ class Model(object):
         self._data_var = []
         self._fetch_interval = 20
 
+    def get_input(self):
+        return self._data_var
+
     def get_cost_op(self):
         """R
         """
@@ -94,20 +97,23 @@ class Model(object):
         return self._fetch_interval
 
     @abc.abstractmethod
-    def net(self):
+    def train_net(self):
         """R
         """
         pass
 
+    def infer_net(self):
+        pass
+
 
-class YamlModel(Model):
+class YamlModel(ModelBase):
     """R
     """
 
     def __init__(self, config):
         """R
         """
-        Model.__init__(self, config)
+        ModelBase.__init__(self, config)
         self._config = config
         self._name = config['name']
         f = open(config['layer_file'], 'r')
@@ -116,7 +122,7 @@ class YamlModel(Model):
         self._build_param = {'layer': {}, 'inner_layer': {}, 'layer_extend': {}, 'model': {}}
         self._inference_meta = {'dependency': {}, 'params': {}}
 
-    def net(self):
+    def train_net(self):
         """R
         build a fluid model with config
         Return:

fleetrec/models/ctr_dnn/model.py

@@ -16,12 +16,12 @@ import math
 import paddle.fluid as fluid
 
 from fleetrec.utils import envs
-from fleetrec.models.base import Model
+from fleetrec.models.base import ModelBase
 
 
-class TrainModel(Model):
+class Model(ModelBase):
     def __init__(self, config):
-        Model.__init__(self, config)
+        ModelBase.__init__(self, config)
         self.namespace = "train.model"
 
     def input(self):
@@ -52,8 +52,10 @@ class TrainModel(Model):
         self.dense_input = dense_input()
         self.label_input = label_input()
 
-    def inputs(self):
-        return [self.dense_input] + self.sparse_inputs + [self.label_input]
+        for input in self.sparse_inputs:
+            self._data_var.append(input)
+        self._data_var.append(self.dense_input)
+        self._data_var.append(self.label_input)
 
     def net(self):
         def embedding_layer(input):
@@ -112,15 +114,17 @@ class TrainModel(Model):
         self._metrics["AUC"] = auc
         self._metrics["BATCH_AUC"] = batch_auc
 
+    def train_net(self):
+        self.input()
+        self.net()
+        self.avg_loss()
+        self.metrics()
+
     def optimizer(self):
         learning_rate = envs.get_global_env("hyper_parameters.learning_rate", None, self.namespace)
         optimizer = fluid.optimizer.Adam(learning_rate, lazy_mode=True)
         return optimizer
 
-
-class EvaluateModel(object):
-    def input(self):
-        pass
-
-    def net(self):
-        pass
+    def infer_net(self):
+        self.input()
+        self.net()

fleetrec/trainer/cluster_trainer.py

@@ -66,15 +66,11 @@ class ClusterTrainer(TranspileTrainer):
         return strategy
 
     def init(self, context):
-        self.model.input()
-        self.model.net()
-        self.model.metrics()
-        self.model.avg_loss()
+        self.model.train_net()
         optimizer = self.model.optimizer()
-
         strategy = self.build_strategy()
         optimizer = fleet.distributed_optimizer(optimizer, strategy)
-        optimizer.minimize(self.model._cost)
+        optimizer.minimize(self.model.get_cost_op())
 
         if fleet.is_server():
             context['status'] = 'server_pass'

fleetrec/trainer/single_trainer.py

@@ -37,12 +37,9 @@ class SingleTrainer(TranspileTrainer):
         self.regist_context_processor('terminal_pass', self.terminal)
 
     def init(self, context):
-        self.model.input()
-        self.model.net()
-        self.model.metrics()
-        self.model.avg_loss()
+        self.model.train_net()
         optimizer = self.model.optimizer()
-        optimizer.minimize(self.model._cost)
+        optimizer.minimize((self.model.get_cost_op()))
 
         self.fetch_vars = []
         self.fetch_alias = []