Add test set

gqa.py

@@ -37,7 +37,8 @@ seeking to address key shortcomings of previous visual question answering (VQA)
 
 _URLS = {
     "train": "https://nlp.cs.unc.edu/data/lxmert_data/gqa/train.json",
-    "dev": "https://nlp.cs.unc.edu/data/lxmert_data/gqa/valid.json",
+    "valid": "https://nlp.cs.unc.edu/data/lxmert_data/gqa/valid.json",
+    "testdev": "https://nlp.cs.unc.edu/data/lxmert_data/gqa/testdev.json",
     "img": "https://downloads.cs.stanford.edu/nlp/data/gqa/images.zip",
 }
 
@@ -78,7 +79,11 @@ class Gqa(datasets.GeneratorBasedBuilder):
             ),
             datasets.SplitGenerator(
                 name=datasets.Split.VALIDATION,
-                gen_kwargs={"filepath": dl_dir["dev"], "img_dir": os.path.join(dl_dir["img"], _IMG_DIR)},
+                gen_kwargs={"filepath": dl_dir["valid"], "img_dir": os.path.join(dl_dir["img"], _IMG_DIR)},
+            ),
+            datasets.SplitGenerator(
+                name=datasets.Split.TEST,
+                gen_kwargs={"filepath": dl_dir["testdev"], "img_dir": os.path.join(dl_dir["img"], _IMG_DIR)},
             ),
         ]