fix lr layer decay and shuffle

Author: kimiyoung

model_utils.py

@@ -144,7 +144,7 @@ def get_train_op(FLAGS, total_loss, grads_and_vars=None):
   gradients, variables = zip(*grads_and_vars)
   clipped, gnorm = tf.clip_by_global_norm(gradients, FLAGS.clip)
 
-  if FLAGS.lr_layer_decay_rate != 1.0:
+  if getattr(FLAGS, "lr_layer_decay_rate", 1.0) != 1.0:
     n_layer = 0
     for i in range(len(clipped)):
       m = re.search(r"model/transformer/layer_(\d+?)/", variables[i].name)

run_classifier.py

@@ -406,7 +406,6 @@ def file_based_convert_examples_to_features(
 
   writer = tf.python_io.TFRecordWriter(output_file)
 
-  np.random.shuffle(examples)
   if num_passes > 1:
     examples *= num_passes
 
@@ -704,6 +703,7 @@ def tokenize_fn(text):
     tf.logging.info("Use tfrecord file {}".format(train_file))
 
     train_examples = processor.get_train_examples(FLAGS.data_dir)
+    np.random.shuffle(train_examples)
     tf.logging.info("Num of train samples: {}".format(len(train_examples)))
 
     file_based_convert_examples_to_features(