Create new file

create_wikiart.py

@@ -0,0 +1,49 @@
+from datasets import load_dataset
+from torchvision.transforms import InterpolationMode
+from torchvision.transforms import functional as F
+
+
+def art2prompt(example):
+    image = example["image"]
+    image = F.resize(image, 512, InterpolationMode.LANCZOS)
+
+    artist = example["artist_str"].replace("-", " ").title()
+    if example["genre_str"] == "Unknown Genre":
+        genre = "painting"
+    else:
+        genre = example["genre_str"].replace("_", " ")
+
+    style = example["style_str"].replace("_", " ").lower()
+
+    captions = [
+        # a landscape in the style of Vincent Van Gogh
+        f"a {genre} in the style of {artist}",
+        # a landscape in the style of realism
+        f"a {genre} in the style of {style}",
+        # a realism painting by Vincent Van Gogh
+        f"a {style} painting by {artist}",
+        # a landscape by Vincent Van Gogh
+        f"a {genre} by {artist}",
+    ]
+
+    return {"text": captions, "image": image}
+
+
+dataset = load_dataset("huggan/wikiart", split="train")
+
+# map the integer labels to their strings
+dataset = dataset.map(
+    lambda ex: {
+        "artist_str": dataset.features["artist"].int2str(ex["artist"]),
+        "genre_str": dataset.features["genre"].int2str(ex["genre"]),
+        "style_str": dataset.features["style"].int2str(ex["style"]),
+    },
+    remove_columns=["artist", "genre", "style"],
+)
+
+# generate prompts from attributes
+dataset = dataset.map(
+    art2prompt, remove_columns=["artist_str", "genre_str", "style_str"], num_proc=8, writer_batch_size=100
+)
+
+dataset.push_to_hub("fusing/wikiart_captions", split="train")