Add hierarchies to dataset

captions_images.zip → captions_hierarchies_images.zip

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:850703b04c0723b698e8c72ee3bb8540d45f34632cd49a81014dac4aa390b4d8
-size 5850078360
+oid sha256:06be80e71cfeedcf22f20bed55a276587fc5831605fae37df9cc829ba18dba5e
+size 6478261393

rico_captions.py

@@ -50,7 +50,7 @@ _LICENSE = ""
 # The HuggingFace dataset library don't host the datasets but only point to the original files
 # This can be an arbitrary nested dict/list of URLs (see below in `_split_generators` method)
 _DATA_URLs = {
-    "screenshots_captions": "https://huggingface.co/datasets/ncoop57/rico_captions/resolve/main/captions_images.zip",
+    "screenshots_captions": "https://huggingface.co/datasets/ncoop57/rico_captions/resolve/main/captions_hierarchies_images.zip",
 }
 
 
@@ -86,8 +86,9 @@ class RicoDataset(datasets.GeneratorBasedBuilder):
             features = datasets.Features(
                 {
                     "screenshot_path": datasets.Value("string"),
-                    # This is a JSON obj, but will be coded as a string
                     "caption": datasets.Value("string"),
+                    # This is a JSON obj, but will be coded as a string
+                    "hierarchy": datasets.Value("string"),
                 }
             )
 
@@ -131,11 +132,14 @@ class RicoDataset(datasets.GeneratorBasedBuilder):
         # The `key` is here for legacy reason (tfds) and is not important in itself.
 
         screen_glob = sorted(glob.glob(os.path.join(root_dir, "**/*.jpg")))
+        hierarchy_glob = sorted(glob.glob(os.path.join(root_dir, "**/*.json")))
         caption_glob = sorted(glob.glob(os.path.join(root_dir, "**/*.txt")))
-        for idx, (screen_filepath, caption_filepath) in enumerate(
-            zip(screen_glob, caption_glob)
+        for idx, (screen_filepath, hierarchy_filepath, caption_filepath) in enumerate(
+            zip(screen_glob, hierarchy_glob, caption_glob)
         ):
+            with open(hierarchy_filepath, "r", encoding="utf-8") as f:
+                hierarchy = f.read()
             with open(caption_filepath, "r", encoding="utf-8") as f:
                 caption = f.read()
 
-            yield idx, {"screenshot_path": screen_filepath, "caption": caption}
+            yield idx, {"screenshot_path": screen_filepath, "hierarchy": hierarchy, "caption": caption}