add batch_size

python/examples/resnet_v2_50/benchmark.py

@@ -56,9 +56,13 @@ def single_func(idx, resource):
                 ])
                 image_file = "daisy.jpg"
                 img = seq(image_file)
+                feed_data = np.array(img)
+                feed_data = np.expand_dims(feed_data, 0).repeat(
+                    args.batch_size, axis=0)
                 result = client.predict(
-                    feed={"image": img},
-                    fetch=["save_infer_model/scale_0.tmp_0"])
+                    feed={"image": feed_data},
+                    fetch=["save_infer_model/scale_0.tmp_0"],
+                    batch=True)
                 l_end = time.time()
                 if latency_flags:
                     latency_list.append(l_end * 1000 - l_start * 1000)

python/examples/resnet_v2_50/benchmark.sh

@@ -20,7 +20,7 @@ python3.6 benchmark.py --thread 1 --batch_size 1 --model $2/serving_client_conf.
 echo -e "import psutil\nimport time\nwhile True:\n\tcpu_res = psutil.cpu_percent()\n\twith open('cpu.txt', 'a+') as f:\n\t\tf.write(f'{cpu_res}\\\n')\n\ttime.sleep(0.1)" > cpu.py
 for thread_num in 1 2 4 8
 do
-for batch_size in 1
+for batch_size in 1 4 8 16 32 64
 do
     job_bt=`date '+%Y%m%d%H%M%S'`
     nvidia-smi --id=$gpu_id --query-compute-apps=used_memory --format=csv -lms 100 > gpu_memory_use.log 2>&1 &