Upload tag_char_info.py with huggingface_hub

tag_char_info.py

@@ -0,0 +1,95 @@
+import pandas as pd
+import json
+from tqdm import tqdm
+from concurrent.futures import ProcessPoolExecutor
+from collections import Counter
+
+if __name__ == "__main__":
+    # 文件路径
+    danbooru_parquets_path = "/mnt/data/Booru-parquets/danbooru.parquet"
+    danbooru_parquets_path_add = "/mnt/data/danbooru_newest-all/table.parquet"
+
+    # 读取 Parquet 文件
+    df1 = pd.read_parquet(danbooru_parquets_path)
+    df2 = pd.read_parquet(danbooru_parquets_path_add)
+    df = pd.concat([df1, df2], ignore_index=True)
+    print(df.columns)
+    print(df.head())
+
+    # 获取所有 tag_string_character的list
+    tag_string_character = df['tag_string_character'].unique()
+    
+    CHARACTER_TAG_LIST = []
+    for tag_list in tqdm(tag_string_character):
+        if tag_list is None:
+            continue  # 跳过 None 值
+        for tag in tag_list.split(" "):
+            CHARACTER_TAG_LIST.append(tag)
+    CHARACTER_TAG_LIST = list(set(CHARACTER_TAG_LIST))
+    CHARACTER_TAG_LIST.remove("")
+    print(len(CHARACTER_TAG_LIST))
+    
+    def map_function(tag):
+        """映射函数：获取每个角色的 ID 和相关文本的词频"""
+        # 使用布尔���引来选择包含特定 tag 的行
+        tag_data = df[df['tag_string_character'].str.contains(tag, na=False, regex=False)]
+        
+        tag_id = tag_data['id'].unique()
+        word_counts = Counter()
+
+        # 假设有一列包含文本，如 'description'，替换为实际列名
+        text_column = 'tag_string_general'  # 替换为实际列名
+        for text in tag_data[text_column]:
+            if isinstance(text, str):  # 确保文本不是 None
+                words = text.split(" ")  # 根据需要进行分词
+                word_counts.update(words)
+
+        return tag, tag_id, word_counts
+
+    def reduce_function(results):
+        """归约函数：合并所有结果"""
+        tag_string_character_id = {}
+
+
+        for tag, ids, word_counts in results:
+            tag_string_character_id[tag] = {}
+            tag_string_character_id[tag]["ids"] = ids.tolist()  # 将 NumPy 数组转换为列表
+            tag_string_character_id[tag]["word_counts"] = dict(word_counts)  # 将词频转换为字典
+
+        return tag_string_character_id
+
+    with open("character_tag_list.txt", "w") as f:
+        f.write("\n".join(CHARACTER_TAG_LIST))
+    # CHARACTER_TAG_LIST = CHARACTER_TAG_LIST[:60]
+    # print(CHARACTER_TAG_LIST)
+    # 使用多进程并行处理
+    with ProcessPoolExecutor(max_workers=32) as executor:
+        # 映射阶段
+        mapped_results = list(tqdm(executor.map(map_function, CHARACTER_TAG_LIST), total=len(CHARACTER_TAG_LIST)))
+
+    # 归约阶段
+    tag_string_character_id = reduce_function(mapped_results)
+
+    # 展示前五个艺术家的 ID 和词频
+    # print(list(tag_string_character_id.keys())[:5])
+    # print(list(tag_string_character_id.values())[:5])
+
+    
+    # 保存 ID,词频 为 JSON
+    with open("character_id_with_word_counts.json", "w") as f:
+        json.dump(tag_string_character_id, f)
+
+
+
+    # 保存艺术家 ID 和词频为 Parquet
+    character_data = []
+    for tag, data in tag_string_character_id.items():
+        character_data.append({
+            'tag': tag,
+            'count':len(data["ids"]),
+            'id': data["ids"],
+            'word_counts': data["word_counts"]
+        })
+
+    artist_df = pd.DataFrame(character_data)
+    artist_df.to_parquet("char_id_with_word_counts.parquet")