Revise dataset processing order

* shuffle before applying the preprocessing function
* call repeat after batching the dataset otherwise batches can contain
  examples from future epochs

opennmt/models/model.py

@@ -375,6 +375,9 @@ def _input_fn_impl(self,
       padded_shapes_fn = lambda: (
           feat_padded_shapes_fn(), labels_padded_shapes_fn())
 
+    if mode == tf.estimator.ModeKeys.TRAIN:
+      dataset = dataset.shuffle(buffer_size, seed=int(time.time()))
+
     dataset = dataset.map(
         process_fn,
         num_parallel_calls=num_parallel_process_calls).prefetch(buffer_size)
@@ -386,8 +389,6 @@ def _input_fn_impl(self,
           labels,
           maximum_features_length=maximum_features_length,
           maximum_labels_length=maximum_labels_length))
-      dataset = dataset.shuffle(buffer_size, seed=int(time.time()))
-      dataset = dataset.repeat()
 
     num_buckets = num_buckets or 1
 
@@ -432,6 +433,9 @@ def _reduce_func(unused_key, dataset):
           batch_size,
           padded_shapes=padded_shapes)
 
+    if mode == tf.estimator.ModeKeys.TRAIN:
+      dataset = dataset.repeat()
+
     iterator = dataset.make_initializable_iterator()
 
     # Add the initializer to a standard collection for it to be initialized.