add ctr ce

fluid/PaddleRec/ctr/.run_ce.sh

@@ -6,11 +6,17 @@ export OMP_NUM_THREADS=1
 
 #cudaid=${face_detection:=0} # use 0-th card as default
 #export CUDA_VISIBLE_DEVICES=$cudaid
+export CPU_NUM=1
 export NUM_THREADS=1
 
 FLAGS_benchmark=true  python train.py --is_local 1 --cloud_train 0 --train_data_path data/raw/train.txt --enable_ce | python _ce.py
 
-export NUM_THREADS=4
+export CPU_NUM=1
+export NUM_THREADS=8
 
 FLAGS_benchmark=true  python train.py --is_local 1 --cloud_train 0 --train_data_path data/raw/train.txt --enable_ce | python _ce.py
 
+export CPU_NUM=8
+export NUM_THREADS=8
+
+FLAGS_benchmark=true  python train.py --is_local 1 --cloud_train 0 --train_data_path data/raw/train.txt --enable_ce | python _ce.py

fluid/PaddleRec/ctr/_ce.py

@@ -8,24 +8,32 @@ from kpi import DurationKpi
 from kpi import AccKpi
 
 
-each_pass_duration_card1_kpi = DurationKpi('each_pass_duration_card1', 0.08, 0, actived=True)
-train_loss_card1_kpi = CostKpi('train_loss_card1', 0.08, 0)
-train_auc_val_card1_kpi = AccKpi('train_auc_val_card1', 0.08, 0)
-train_batch_auc_val_card1_kpi = AccKpi('train_batch_auc_val_card1', 0.08, 0)
-each_pass_duration_card4_kpi = DurationKpi('each_pass_duration_card4', 0.08, 0, actived=True)
-train_loss_card4_kpi = CostKpi('train_loss_card4', 0.08, 0)
-train_auc_val_card4_kpi = AccKpi('train_auc_val_card4', 0.08, 0)
-train_batch_auc_val_card4_kpi = AccKpi('train_batch_auc_val_card4', 0.08, 0)
+each_pass_duration_cpu1_thread1_kpi = DurationKpi('each_pass_duration_cpu1_thread1', 0.08, 0, actived=True)
+train_loss_cpu1_thread1_kpi = CostKpi('train_loss_cpu1_thread1', 0.08, 0)
+train_auc_val_cpu1_thread1_kpi = AccKpi('train_auc_val_cpu1_thread1', 0.08, 0)
+train_batch_auc_val_cpu1_thread1_kpi = AccKpi('train_batch_auc_val_cpu1_thread1', 0.08, 0)
+each_pass_duration_cpu1_thread8_kpi = DurationKpi('each_pass_duration_cpu1_thread8', 0.08, 0, actived=True)
+train_loss_cpu1_thread8_kpi = CostKpi('train_loss_cpu1_thread8', 0.08, 0)
+train_auc_val_cpu1_thread8_kpi = AccKpi('train_auc_val_cpu1_thread8', 0.08, 0)
+train_batch_auc_val_cpu1_thread8_kpi = AccKpi('train_batch_auc_val_cpu1_thread8', 0.08, 0)
+each_pass_duration_cpu8_thread8_kpi = DurationKpi('each_pass_duration_cpu8_thread8', 0.08, 0, actived=True)
+train_loss_cpu8_thread8_kpi = CostKpi('train_loss_cpu8_thread8', 0.08, 0)
+train_auc_val_cpu8_thread8_kpi = AccKpi('train_auc_val_cpu8_thread8', 0.08, 0)
+train_batch_auc_val_cpu8_thread8_kpi = AccKpi('train_batch_auc_val_cpu8_thread8', 0.08, 0)
 
 tracking_kpis = [
-        each_pass_duration_card1_kpi,
-        train_loss_card1_kpi,
-        train_auc_val_card1_kpi,
-        train_batch_auc_val_card1_kpi,
-        each_pass_duration_card4_kpi,
-        train_loss_card4_kpi,
-        train_auc_val_card4_kpi,
-        train_batch_auc_val_card4_kpi
+        each_pass_duration_cpu1_thread1_kpi,
+        train_loss_cpu1_thread1_kpi,
+        train_auc_val_cpu1_thread1_kpi,
+        train_batch_auc_val_cpu1_thread1_kpi,
+        each_pass_duration_cpu1_thread8_kpi,
+        train_loss_cpu1_thread8_kpi,
+        train_auc_val_cpu1_thread8_kpi,
+        train_batch_auc_val_cpu1_thread8_kpi,
+        each_pass_duration_cpu8_thread8_kpi,
+        train_loss_cpu8_thread8_kpi,
+        train_auc_val_cpu8_thread8_kpi,
+        train_batch_auc_val_cpu8_thread8_kpi,
         ]
 
 

fluid/PaddleRec/ctr/train.py

@@ -116,6 +116,11 @@ def parse_args():
         type=int,
         default=1,
         help='The num of threads, (default: 1)')
+    parser.add_argument(
+        '--num_cpu',
+        type=int,
+        default=1,
+        help='The num of cpu, (default: 1)')
 
     return parser.parse_args()
 
@@ -193,16 +198,16 @@ def train_loop(args, train_program, py_reader, loss, auc_var, batch_auc_var,
 
     # only for ce
     if args.enable_ce:
-        cpu_num = get_cards(args)
+        threads_num, cpu_num = get_cards(args)
         epoch_idx = args.num_passes 
-        print("kpis\teach_pass_duration_card%s\t%s" %
-                (cpu_num, total_time / epoch_idx))
-        print("kpis\ttrain_loss_card%s\t%s" %
-                (cpu_num, loss_val/args.batch_size))
-        print("kpis\ttrain_auc_val_card%s\t%s" %
-                (cpu_num, auc_val))
-        print("kpis\ttrain_batch_auc_val_card%s\t%s" %
-                (cpu_num, batch_auc_val))
+        print("kpis\teach_pass_duration_cpu%s_thread%s\t%s" %
+                (cpu_num, threads_num, total_time / epoch_idx))
+        print("kpis\ttrain_loss_cpu%s_thread%s\t%s" %
+                (cpu_num, threads_num, loss_val/args.batch_size))
+        print("kpis\ttrain_auc_val_cpu%s_thread%s\t%s" %
+                (cpu_num, threads_num, auc_val))
+        print("kpis\ttrain_batch_auc_val_cpu%s_thread%s\t%s" %
+                (cpu_num, threads_num, batch_auc_val))
         
 
 def train():
@@ -257,10 +262,11 @@ def train():
 
 def get_cards(args):
     if args.enable_ce:
-        cards = os.environ.get('NUM_THREADS', 1)
-        return int(cards)
+        threads_num = os.environ.get('NUM_THREADS', 1)
+        cpu_num = os.environ.get('CPU_NUM', 1)
+        return int(threads_num), int(cpu_num)
     else:
-        return args.num_threads
+        return args.num_threads, args.num_cpu
 
 
 if __name__ == '__main__':