Fix benchmark (PaddlePaddle#451)

* fix benchmark

* add fomm and stylegan benchmark

benchmark/analysis_log.py

@@ -0,0 +1,51 @@
+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+# encoding=utf-8 vi:ts=4:sw=4:expandtab:ft=python
+
+import re
+import sys
+import json
+
+def analyze(model_name, log_file, res_log_file):
+    time_pat = re.compile(r"ips: (.*) images/s")
+
+    logs = open(log_file).readlines()
+    logs = ";".join(logs)
+    time_res = time_pat.findall(logs)
+
+    fail_flag = 0
+    run_mode = ""
+    gpu_num = 0
+    ips = 0
+
+    if time_res == []:
+        fail_flag = 1
+    else:
+        gpu_num = log_file.split('_')[-1]
+        run_mode = "sp" if gpu_num == 1 else "mp"
+
+        skip_num = 4
+        total_time = 0
+        for i in range(skip_num, len(time_res)):
+            total_time += float(time_res[i])
+        ips = total_time / (len(time_res) - skip_num)
+
+    info = {"log_file": log_file, "model_name": model_name, "mission_name": "图像生成",
+            "direction_id": 0, "run_mode": run_mode, "index": 1, "gpu_num": gpu_num,
+            "FINAL_RESULT": ips, "JOB_FAIL_FLAG": fail_flag, "UNIT": "images/s"}
+    json_info = json.dumps(info)
+    with open(res_log_file, "w") as of:
+        of.write(json_info)
+
+if __name__ == "__main__":
+    if len(sys.argv) != 4:
+        print("Usage:" + sys.argv[0] + " model_name path/to/log/file path/to/res/log/file")
+        sys.exit()
+
+    model_name = sys.argv[1]
+    log_file = sys.argv[2]
+    res_log_file = sys.argv[3]
+
+    analyze(model_name, log_file, res_log_file) 
+
+

benchmark/run_benchmark.sh

@@ -18,7 +18,13 @@ function _set_params(){
     arr=(${device})
     num_gpu_devices=${#arr[*]}
     log_file=${run_log_path}/${model_name}_${run_mode}_bs${batch_size}_${fp_item}_${num_gpu_devices}
+    res_log_file=${run_log_path}/${model_name}_${run_mode}_bs${batch_size}_${fp_item}_${num_gpu_devices}_speed
 }
+
+function _analysis_log(){
+    python benchmark/analysis_log.py ${model_name} ${log_file} ${res_log_file}
+}
+
 function _train(){
     echo "Train on ${num_gpu_devices} GPUs"
     echo "current CUDA_VISIBLE_DEVICES=$CUDA_VISIBLE_DEVICES, gpus=$num_gpu_devices, batch_size=$batch_size"
@@ -49,6 +55,9 @@ function _train(){
         rm ${log_file}
         cp mylog/workerlog.0 ${log_file}
     fi
+
+    _analysis_log
+
 }
 
 _set_params $@