Fixed the nightly testing failure of transformer_lt_official model (#54)

* Fixed the nightly testing failure of transformer_lt_official model

* commented out unnecessary excessive logging info

* Updated the in_graph parameters and remove KMP setting, which could be different on different hardware

Author: cuixiaom

benchmarks/common/tensorflow/start.sh

@@ -1280,7 +1280,6 @@ function transformer_lt_official() {
     fi
 
     CMD="${CMD}
-    --input_graph=${input_graph} \
     --vocab_file=${DATASET_LOCATION}/${vocab_file} \
     --file=${DATASET_LOCATION}/${file} \
     --file_out=${OUTPUT_DIR}/${file_out} \

benchmarks/language_translation/tensorflow/transformer_lt_official/inference/fp32/config.json

@@ -1,7 +1,2 @@
 {
-    "optimization_parameters": {
-        "KMP_AFFINITY": "granularity=fine,compact,1,0",
-        "KMP_BLOCKTIME": 0,
-        "KMP_SETTINGS": 1
-    }
 }

benchmarks/language_translation/tensorflow/transformer_lt_official/inference/fp32/model_init.py

@@ -18,6 +18,7 @@
 #
 
 import os
+import sys
 from argparse import ArgumentParser
 
 from common.base_model_init import BaseModelInitializer
@@ -26,10 +27,22 @@
 
 class ModelInitializer(BaseModelInitializer):
     """Model initializer for Transformer LT FP32 inference"""
+    def run_inference_sanity_checks(self, args, custom_args):
+        if not args.input_graph:
+            sys.exit("Please provide a path to the frozen graph directory"
+                     " via the '--in-graph' flag.")
+        if not args.data_location:
+            sys.exit("Please provide a path to the data directory via the "
+                     "'--data-location' flag.")
+        if args.socket_id == -1 and args.num_cores == -1:
+            print("***Warning***: Running inference on all cores could degrade"
+                  " performance. Pass a '--socket-id' to specify running on a"
+                  " single socket instead.\n")
 
     def __init__(self, args, custom_args, platform_util=None):
         super(ModelInitializer, self).__init__(args, custom_args, platform_util)
 
+        self.run_inference_sanity_checks(self.args, self.custom_args)
         self.cmd = self.get_command_prefix(self.args.socket_id)
         self.bleu_params = ""
 
@@ -61,10 +74,6 @@ def __init__(self, args, custom_args, platform_util=None):
                                 help='input vocable file for translation',
                                 dest="vocab_file",
                                 default="vocab.txt")
-        arg_parser.add_argument('--in_graph',
-                                help='input fp32 frozen graph file for inference',
-                                dest="fp32_graph",
-                                default="fp32_graphdef.pb")
         arg_parser.add_argument('--file',
                                 help='decode input file with path',
                                 dest="decode_from_file",
@@ -85,7 +94,7 @@ def __init__(self, args, custom_args, platform_util=None):
         translate_file = os.path.join(self.args.output_dir,
                                       self.args.decode_to_file)
         cmd_args = " --param_set=" + self.args.param_set + \
-                   " --in_graph=" + self.args.fp32_graph + \
+                   " --in_graph=" + self.args.input_graph + \
                    " --batch_size=" + \
                    (str(self.args.batch_size)
                     if self.args.batch_size != -1 else "1") + \

models/language_translation/tensorflow/transformer_lt_official/inference/fp32/utils/tokenizer.py

@@ -63,8 +63,8 @@ class Subtokenizer(object):
 
   def __init__(self, vocab_file, reserved_tokens=None):
     """Initializes class, creating a vocab file if data_files is provided."""
-    tf.compat.v1.logging.info("Initializing Subtokenizer from file %s." %
-                              vocab_file)
+#    tf.compat.v1.logging.info("Initializing Subtokenizer from file %s." %
+#                              vocab_file)
 
     if reserved_tokens is None:
       reserved_tokens = RESERVED_TOKENS