From 868ed52da2eded9a3b20973eb1611af80bf4ac8d Mon Sep 17 00:00:00 2001
From: zhengya01 <zhengya01@baidu.com>
Date: Tue, 26 Feb 2019 13:46:29 +0000
Subject: [PATCH] add ctr ce

---
 fluid/PaddleRec/ctr/.run_ce.sh |  9 +++++++--
 fluid/PaddleRec/ctr/_ce.py     | 10 +++++++++-
 fluid/PaddleRec/ctr/train.py   | 27 +++++++++++++++------------
 3 files changed, 31 insertions(+), 15 deletions(-)

diff --git a/fluid/PaddleRec/ctr/.run_ce.sh b/fluid/PaddleRec/ctr/.run_ce.sh
index b5b1c9dc..8bd8986d 100755
--- a/fluid/PaddleRec/ctr/.run_ce.sh
+++ b/fluid/PaddleRec/ctr/.run_ce.sh
@@ -4,8 +4,13 @@ export MKL_NUM_THREADS=1
 export OMP_NUM_THREADS=1
 
 
-cudaid=${face_detection:=0} # use 0-th card as default
-export CUDA_VISIBLE_DEVICES=$cudaid
+#cudaid=${face_detection:=0} # use 0-th card as default
+#export CUDA_VISIBLE_DEVICES=$cudaid
+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
 
 FLAGS_benchmark=true  python train.py --is_local 1 --cloud_train 0 --train_data_path data/raw/train.txt --enable_ce | python _ce.py
 
diff --git a/fluid/PaddleRec/ctr/_ce.py b/fluid/PaddleRec/ctr/_ce.py
index 67e18566..7597dd52 100644
--- a/fluid/PaddleRec/ctr/_ce.py
+++ b/fluid/PaddleRec/ctr/_ce.py
@@ -13,12 +13,20 @@ each_pass_duration_card1_kpi = DurationKpi('each_pass_duration_card1', 0.08, 0,
 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)
 
 tracking_kpis = [
         each_pass_duration_card1_kpi,
         train_loss_card1_kpi,
         train_auc_val_card1_kpi,
-        train_batch_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
         ]
 
 
diff --git a/fluid/PaddleRec/ctr/train.py b/fluid/PaddleRec/ctr/train.py
index 97c26dc0..bbeebafe 100644
--- a/fluid/PaddleRec/ctr/train.py
+++ b/fluid/PaddleRec/ctr/train.py
@@ -112,10 +112,10 @@ def parse_args():
         action='store_true',
         help='If set, run the task with continuous evaluation logs.')
     parser.add_argument(
-        '--num_devices',
+        '--num_threads',
         type=int,
-        default=0,
-        help='The num of devices, (default: 1)')
+        default=1,
+        help='The num of threads, (default: 1)')
 
     return parser.parse_args()
 
@@ -193,16 +193,17 @@ def train_loop(args, train_program, py_reader, loss, auc_var, batch_auc_var,
 
     # only for ce
     if args.enable_ce:
-        gpu_num = get_cards(args)
+        cpu_num = get_cards(args)
+        print("cpu_num", cpu_num)
         epoch_idx = args.num_passes 
         print("kpis\teach_pass_duration_card%s\t%s" %
-                (gpu_num, total_time / epoch_idx))
+                (cpu_num, total_time / epoch_idx))
         print("kpis\ttrain_loss_card%s\t%s" %
-                (gpu_num, loss_val/args.batch_size))
+                (cpu_num, loss_val/args.batch_size))
         print("kpis\ttrain_auc_val_card%s\t%s" %
-                (gpu_num, auc_val))
+                (cpu_num, auc_val))
         print("kpis\ttrain_batch_auc_val_card%s\t%s" %
-                (gpu_num, batch_auc_val))
+                (cpu_num, batch_auc_val))
         
 
 def train():
@@ -257,11 +258,13 @@ def train():
 
 def get_cards(args):
     if args.enable_ce:
-        cards = os.environ.get('CUDA_VISIBLE_DEVICES')
-        num = len(cards.split(","))
-        return num
+        cards = os.environ.get('NUM_THREADS', 1)
+        print("cards", cards)
+ 
+        return int(cards)
     else:
-        return args.num_devices
+        print("return args.num_threads")
+        return args.num_threads
 
 
 if __name__ == '__main__':
-- 
GitLab