update the profiler of seq2seq (#3959)

* update the profiler of seq2seq

* modify profiler args

PaddleNLP/PaddleTextGEN/seq2seq/args.py

@@ -123,15 +123,10 @@ def parse_args():
     parser.add_argument(
         "--profile", action='store_true', help="Whether enable the profile.")
     # NOTE: profiler args, used for benchmark
-    parser.add_argument(
-        "--is_profiler",
-        type=int,
-        default=0,
-        help="The switch of profiler tools. (used for benchmark)")
     parser.add_argument(
         "--profiler_path",
         type=str,
-        default='./',
+        default='./seq2seq.profile',
         help="the profiler output file path. (used for benchmark)")
     args = parser.parse_args()
     return args

PaddleNLP/PaddleTextGEN/seq2seq/train.py

@@ -47,9 +47,9 @@ import pickle
 
 
 @contextlib.contextmanager
-def profile_context(profile=True):
+def profile_context(profile=True, profiler_path='./seq2seq.profile'):
     if profile:
-        with profiler.profiler('All', 'total', 'seq2seq.profile'):
+        with profiler.profiler('All', 'total', profiler_path):
             yield
     else:
         yield
@@ -216,10 +216,9 @@ def main():
                     word_count = 0.0
                 
                 # profiler tools
-                if args.is_profiler and epoch_id == 0 and batch_id == 100:
-                    profiler.start_profiler("All")
-                elif args.is_profiler and epoch_id == 0 and batch_id == 105:
-                    profiler.stop_profiler("total", args.profiler_path)
+                if args.profile and epoch_id == 0 and batch_id == 100:
+                    profiler.reset_profiler()
+                elif args.profile and epoch_id == 0 and batch_id == 105:
                     return
 
             end_time = time.time()
@@ -252,7 +251,7 @@ def main():
             print("kpis\ttrain_duration_card%s\t%s" % (card_num, _time))
             print("kpis\ttrain_ppl_card%s\t%f" % (card_num, _ppl))
 
-    with profile_context(args.profile):
+    with profile_context(args.profile, args.profiler_path):
         train()