Explicity set the shuffle_buffer_size parameter to address meta-datas…

…et issue #54 (google-research/meta-dataset#54).
cambridge-mlg · Dec 18, 2020 · 7db98d3 · 7db98d3
1 parent df67dab
commit 7db98d3
Showing 1 changed file with 4 additions and 2 deletions.
diff --git a/src/meta_dataset_reader.py b/src/meta_dataset_reader.py
@@ -61,7 +61,8 @@ def _init_multi_source_dataset(self, items, split, episode_description):
             use_bilevel_ontology_list=use_bilevel_ontology_list,
             split=split,
             episode_descr_config=episode_description,
-            image_size=84)
+            image_size=84,
+            shuffle_buffer_size=1000)
 
         iterator = multi_source_pipeline.make_one_shot_iterator()
         return iterator.get_next()
@@ -85,7 +86,8 @@ def _init_single_source_dataset(self, dataset_name, split, episode_description):
             use_bilevel_ontology=use_bilevel_ontology,
             split=split,
             episode_descr_config=episode_description,
-            image_size=84)
+            image_size=84,
+            shuffle_buffer_size=1000)
 
         iterator = single_source_pipeline.make_one_shot_iterator()
         return iterator.get_next()