savedmodel generation

Author: Zhang-Zhenning

modelzoo/BST/pb_to_pbtxt.py

@@ -0,0 +1,13 @@
+from tensorflow.python.saved_model import loader_impl
+from tensorflow.python.lib.io import file_io
+from tensorflow.python.platform import tf_logging as logging
+
+source_dir="/home/deeprec/DeepRec/modelzoo/BST/savedmodels/1"
+
+logging.info("before _parse_saved_model.")
+saved_model = loader_impl._parse_saved_model(source_dir)
+logging.info("_parse_saved_model done.")
+
+path = source_dir + "/saved_model.pb"
+# write pbtxt graph 
+file_io.write_string_to_file(path+"txt", str(saved_model))

modelzoo/BST/prepare_savedmodel.py

@@ -0,0 +1,143 @@
+#include <iostream>
+#include "serving/processor/serving/processor.h"
+#include "serving/processor/serving/predict.pb.h"
+
+static const char* model_config = "{ \
+    \"omp_num_threads\": 4, \
+    \"kmp_blocktime\": 0, \
+    \"feature_store_type\": \"memory\", \
+    \"serialize_protocol\": \"protobuf\", \
+    \"inter_op_parallelism_threads\": 10, \
+    \"intra_op_parallelism_threads\": 10, \
+    \"init_timeout_minutes\": 1, \
+    \"signature_name\": \"serving_default\", \
+    \"read_thread_num\": 3, \
+    \"update_thread_num\": 2, \
+    \"model_store_type\": \"local\", \
+    \"checkpoint_dir\": \"/root/deeprec/DeepRec/modelzoo/BST/result/\", \
+    \"savedmodel_dir\": \"/root/deeprec/DeepRec/modelzoo/BST/savedmodels/1657183908.2336085/\" \
+  } ";
+
+INPUT_FEATURES = [
+    'pid', 'adgroup_id', 'cate_id', 'campaign_id', 'customer', 'brand',
+    'user_id', 'cms_segid', 'cms_group_id', 'final_gender_code', 'age_level',
+    'pvalue_level', 'shopping_level', 'occupation', 'new_user_class_level',
+    'tag_category_list', 'tag_brand_list', 'price'
+]
+
+struct input_format{
+	string pid;
+  string adgroup_id;
+  string cate_id;
+  string campaign_id;
+  string customer;
+  string brand;
+  string user_id;
+  string cms_segid;
+  string cms_group_id;
+  string final_gender_code;
+  string age_level;
+  string pvalue_level;
+  string shopping_level;
+  string occupation;
+  string new_user_class_level;
+  string tag_category_list;
+  string tag_brand_list;
+  string price;
+	
+};
+
+::tensorflow::eas::ArrayProto get_proto(char* char_input,int dim,::tensorflow::eas::ArrayDataType type){
+  ::tensorflow::eas::ArrayShape array_shape;
+  array_shape.add_dim(1);
+  array_shape.add_dim(dim);
+  // input array
+  ::tensorflow::eas::ArrayProto input;
+  input.add_string_val(char_input);
+  input.set_dtype(type);
+  *(input.mutable_array_shape()) = array_shape;
+
+  return input;
+
+}
+
+int main(int argc, char** argv) {
+  int state;
+  void* model = initialize("", model_config, &state);
+  if (state == -1) {
+    std::cerr << "initialize error\n";
+  }
+   
+  // input format
+  input_format inputs = {"430548_1007","669310","1665","360359","167792","247789","841908","81","10","1","4","2","3","0","3","8153|8153|8153|8154|8154|8154|1673|1673|1673|6115|6115|6115|1665|1665|1665|1665|1665|1665|8188|8188|8188|8188|8188|8188|8188|8188|8188|1665|1665|1665|8188|8188|8188|8188|8188|8188|8188|8188|8188|10747|10747|10747|10747|10747|10747|10747|10747|10747|10747|10747","197848|197848|197848|237004|237004|237004|330898|330898|330898|337445|337445|337445|258262|258262|258262|247789|247789|247789|339517|339517|339517|339517|339517|339517|339517|339517|339517|278878|278878|278878|339517|339517|339517|339517|339517|339517|339517|339517|339517|339517|339517|339517|339517|339517|339517|339517|339517|339517|339517|339517","6"}
+
+  // input type: float
+  ::tensorflow::eas::ArrayDataType dtype =
+      ::tensorflow::eas::ArrayDataType::DT_STRING;
+
+
+// ------------------------------------------------------------------------input setting------------------------------------------------------------------------------
+  
+  ::tensorflow::eas::ArrayProto input0 = get_proto(inputs.pid,strlen(inputs.pid),dtype);
+  ::tensorflow::eas::ArrayProto input1 = get_proto(inputs.adgroup_id,strlen(inputs.adgroup_id),dtype);
+  ::tensorflow::eas::ArrayProto input2 = get_proto(inputs.cate_id,strlen(inputs.cate_id),dtype);
+  ::tensorflow::eas::ArrayProto input3 = get_proto(inputs.campaign_id,strlen(inputs.campaign_id),dtype);
+  ::tensorflow::eas::ArrayProto input4 = get_proto(inputs.customer,strlen(inputs.customer),dtype);
+  ::tensorflow::eas::ArrayProto input5 = get_proto(inputs.brand,strlen(inputs.brand),dtype);
+  ::tensorflow::eas::ArrayProto input6 = get_proto(inputs.user_id,strlen(inputs.user_id),dtype);
+  ::tensorflow::eas::ArrayProto input7 = get_proto(inputs.cms_segid,strlen(inputs.cms_segid),dtype);
+  ::tensorflow::eas::ArrayProto input8 = get_proto(inputs.cms_group_id,strlen(inputs.cms_group_id),dtype);
+  ::tensorflow::eas::ArrayProto input9 = get_proto(inputs.final_gender_code,strlen(inputs.final_gender_code),dtype);
+  ::tensorflow::eas::ArrayProto input10 = get_proto(inputs.age_level,strlen(inputs.age_level),dtype);
+  ::tensorflow::eas::ArrayProto input11 = get_proto(inputs.pvalue_level,strlen(inputs.pvalue_level),dtype);
+  ::tensorflow::eas::ArrayProto input12 = get_proto(inputs.shopping_level,strlen(inputs.shopping_level),dtype);
+  ::tensorflow::eas::ArrayProto input13 = get_proto(inputs.occupation,strlen(inputs.occupation),dtype);
+  ::tensorflow::eas::ArrayProto input14 = get_proto(inputs.new_user_class_level,strlen(inputs.new_user_class_level),dtype);
+  ::tensorflow::eas::ArrayProto input15 = get_proto(inputs.tag_category_list,strlen(inputs.tag_category_list),dtype);
+  ::tensorflow::eas::ArrayProto input16 = get_proto(inputs.tag_brand_list,strlen(inputs.tag_brand_list),dtype);
+  ::tensorflow::eas::ArrayProto input17 = get_proto(inputs.price,strlen(inputs.price),dtype);
+  
+
+ 
+  // PredictRequest
+  ::tensorflow::eas::PredictRequest req;
+  req.set_signature_name("serving_default");
+  req.add_output_filter("output:0");
+ 
+  (*req.mutable_inputs())["pid:0"] = input0;
+  (*req.mutable_inputs())["adgroup_id:0"] = input1;
+  (*req.mutable_inputs())["cate_id:0"] = input2;
+  (*req.mutable_inputs())["campaign_id:0"] = input3;
+  (*req.mutable_inputs())["customer:0"] = input4;
+  (*req.mutable_inputs())["brand:0"] = input5;
+  (*req.mutable_inputs())["user_id:0"] = input6;
+  (*req.mutable_inputs())["cms_segid:0"] = input7;
+  (*req.mutable_inputs())["cms_group_id:0"] = input8;
+  (*req.mutable_inputs())["final_gender_code:0"] = input9;
+  (*req.mutable_inputs())["age_level:0"] = input10;
+  (*req.mutable_inputs())["pvalue_level:0"] = input11;
+  (*req.mutable_inputs())["shopping_level:0"] = input12;
+  (*req.mutable_inputs())["occupation:0"] = input13;
+  (*req.mutable_inputs())["new_user_class_level:0"] = input14;
+  (*req.mutable_inputs())["tag_category_list:0"] = input15;
+  (*req.mutable_inputs())["tag_brand_list:0"] = input16;
+  (*req.mutable_inputs())["price:0"] = input17;
+
+  size_t size = req.ByteSizeLong(); 
+  void *buffer = malloc(size);
+  req.SerializeToArray(buffer, size);
+
+  // do process
+  void* output = nullptr;
+  int output_size = 0;
+  state = process(model, buffer, size, &output, &output_size);
+
+  // parse response
+  std::string output_string((char*)output, output_size);
+  ::tensorflow::eas::PredictResponse resp;
+  resp.ParseFromString(output_string);
+  std::cout << "process returned state: " << state << ", response: " << resp.DebugString();
+
+  return 0;
+}
+

modelzoo/BST/result/README.md

@@ -7,6 +7,7 @@
 import collections
 from tensorflow.python.client import timeline
 import json
+from glob import glob
 
 from tensorflow.python.ops import partitioned_variables
 
@@ -55,6 +56,10 @@
     'price': 50
 }
 
+# next_element = iter.next()
+
+# dict("pid":"1 2 3 4 5 ... 1000",)
+
 
 class BST():
     def __init__(self,
@@ -78,9 +83,10 @@ def __init__(self,
         if not inputs:
             raise ValueError("Dataset is not defined.")
         self._feature = inputs[0]
-        self._label = inputs[1]
-
+        self._label = inputs[1]  
+        
         self._unseq_column = user_column + item_column
+
         self._tag_column = tag_column
         self._key_column = key_column
         self._batch_size = batch_size
@@ -101,7 +107,7 @@ def __init__(self,
         self._input_layer_partitioner = input_layer_partitioner
         self._dense_layer_partitioner = dense_layer_partitioner
 
-        self._create_model()
+        self.r = self._create_model()
         with tf.name_scope('head'):
             self._create_loss()
             self._create_optimizer()
@@ -262,6 +268,7 @@ def _create_model(self):
                         self._feature,
                         self._unseq_column,
                         cols_to_output_tensors=key_dict)
+                
 
             # bst input
             with tf.variable_scope('bst_input_layer', reuse=tf.AUTO_REUSE):
@@ -298,6 +305,7 @@ def _create_model(self):
                                          seq_size=self._max_seqence_length,
                                          head_count=self._multi_head_size,
                                          name='bst')
+        
 
         net = tf.concat([unseq_emb, bst_output], axis=1)
 
@@ -323,8 +331,14 @@ def _create_model(self):
             net = tf.cast(net, dtype=tf.float32)
         self._logits = tf.layers.dense(inputs=net, units=1)
 
-        self.probability = tf.math.sigmoid(self._logits)
-        self.output = tf.round(self.probability)
+        self.probability = tf.math.sigmoid(self._logits,name="probability")
+        self.output = tf.round(self.probability,name="output")
+
+        return self.output
+
+     
+
+
 
     # compute loss
     def _create_loss(self):
@@ -380,7 +394,7 @@ def parse_csv(value):
         all_columns.pop(BUY_COLUMN[0])
         features = all_columns
         return features, labels
-
+    
     '''Work Queue Feature'''
     if args.workqueue and not args.tf:
         from tensorflow.python.ops.work_queue import WorkQueue
@@ -493,6 +507,43 @@ def build_feature_columns():
 
     return user_column, item_column, tag_column, key_column
 
+def deldir(dir):
+    if not os.path.exists(dir):
+        return False
+    if os.path.isfile(dir):
+        os.remove(dir)
+        return
+    for i in os.listdir(dir):
+        t = os.path.join(dir, i)
+        if os.path.isdir(t):
+            deldir(t)
+        else:
+            os.unlink(t)
+    os.removedirs(dir)
+
+class MyHook(tf.train.SessionRunHook):
+        def __init__(self,cur_model,export_dir):
+            self.model = cur_model
+            self.dir = export_dir
+        
+        def before_run(self, run_context):
+            """返回SessionRunArgs和session run一起跑"""
+            v1 = tf.get_collection('logis')
+            prob = tf.get_collection('prob')
+            return tf.train.SessionRunArgs(fetches=[v1, prob])
+
+        
+        def end(self,session):
+            if os.path.exists(self.dir):
+                deldir(self.dir)
+                os.mkdir(self.dir)
+            
+            tf.saved_model.simple_save(
+                session,
+                self.dir,
+                inputs = self.model._feature,
+                outputs = {"predict":self.model.output}
+            )
 
 def train(sess_config,
           input_hooks,
@@ -509,7 +560,7 @@ def train(sess_config,
     scaffold = tf.train.Scaffold(
         local_init_op=tf.group(tf.local_variables_initializer(), data_init_op),
         saver=tf.train.Saver(max_to_keep=args.keep_checkpoint_max))
-
+    # save_hook = MyHook(model,"/root/deeprec/DeepRec/modelzoo/BST/result/savedmodels")
     stop_hook = tf.train.StopAtStepHook(last_step=steps)
     log_hook = tf.train.LoggingTensorHook(
         {
@@ -518,6 +569,8 @@ def train(sess_config,
         }, every_n_iter=100)
     hooks.append(stop_hook)
     hooks.append(log_hook)
+    dir = "/home/deeprec/DeepRec/modelzoo/BST/result/savedmodels"
+    # hooks.append(save_hook)
     if args.timeline > 0:
         hooks.append(
             tf.train.ProfilerHook(save_steps=args.timeline,
@@ -545,8 +598,12 @@ def train(sess_config,
             summary_dir=checkpoint_dir,
             save_summaries_steps=args.save_steps,
             config=sess_config) as sess:
+        
         while not sess.should_stop():
             sess.run([model.loss, model.train_op])
+        
+            
+        
     print("Training completed.")
 
 
@@ -575,6 +632,9 @@ def eval(sess_config, input_hooks, model, data_init_op, steps, checkpoint_dir):
                 writer.add_summary(events, _in)
                 print("Evaluation complate:[{}/{}]".format(_in, steps))
                 print("ACC = {}\nAUC = {}".format(eval_acc, eval_auc))
+          
+
+
 
 
 def main(tf_config=None, server=None):
@@ -629,6 +689,8 @@ def main(tf_config=None, server=None):
 
     # create feature column
     user_column, item_column, tag_column, key_column = build_feature_columns()
+    
+    
 
     # create variable partitioner for distributed training
     num_ps_replicas = len(tf_config['ps_hosts']) if tf_config else 0
@@ -682,6 +744,9 @@ def main(tf_config=None, server=None):
     if not (args.no_eval or tf_config):
         eval(sess_config, hooks, model, test_init_op, test_steps,
              checkpoint_dir)
+    
+
+    
 
 
 def boolean_string(string):

modelzoo/BST/start_serving.cc

@@ -0,0 +1,13 @@
+from tensorflow.python.saved_model import loader_impl
+from tensorflow.python.lib.io import file_io
+from tensorflow.python.platform import tf_logging as logging
+
+source_dir="/home/deeprec/DeepRec/modelzoo/DBMTL/savedmodels/1657784766"
+
+logging.info("before _parse_saved_model.")
+saved_model = loader_impl._parse_saved_model(source_dir)
+logging.info("_parse_saved_model done.")
+
+path = source_dir + "/saved_model.pb"
+# write pbtxt graph 
+file_io.write_string_to_file(path+"txt", str(saved_model))