Upload vndb_img.py

vndb_img.py

@@ -0,0 +1,212 @@
+import os
+import datasets
+import pandas as pd
+import numpy as np
+from tqdm import tqdm
+import os
+import glob
+
+
+def read_header(path: str) -> list[str]:
+    with open(path, 'r', encoding='utf-8') as f:
+        return f.read().strip('\n\r').split('\t')
+
+
+def read_table(root: str, name: str) -> pd.DataFrame:
+    return pd.read_table(
+        os.path.join(root, 'db', name),
+        sep='\t',
+        header=None,
+        names=read_header(os.path.join(root, 'db', name + '.header')),
+        encoding='utf-8',
+    )
+
+
+def get_image_path(id: str):
+    cat = id[:2]
+    i = int(id[2:])
+    sub = i % 100
+    path = os.path.join(cat, f'{sub:02}', f'{i}.jpg')
+    return path
+
+
+def get_class(avg: np.ndarray) -> np.ndarray:
+    # https://code.blicky.net/yorhel/vndb/src/commit/cac41c8361194bd1d010357a946f056454231509/lib/VNWeb/Images/Lib.pm#L155
+    cls = np.zeros_like(avg)
+    cls[40 < avg] = 1
+    cls[130 < avg] = 2
+    return cls
+
+
+def load_image_df(db_root: str, img_root: str):
+    df = read_table(db_root, 'images')
+
+    df_st = df[df['id'].map(lambda x: x.startswith('sf'))].copy()
+    df_st.loc[:, 'id'] = df_st['id'].map(lambda x: 'st' + x[2:])
+    df = pd.concat([df, df_st], ignore_index=True)
+
+    types = {
+        'ch': 0,
+        'cv': 1,
+        'sf': 2,
+        'st': 3,
+    }
+    df['type'] = df['id'].map(lambda x: types[x[:2]])
+
+    df['sexual_class'] = get_class(df['c_sexual_avg'].values)
+    df['violence_class'] = get_class(df['c_violence_avg'].values)
+
+    df['file_name'] = df['id'].apply(get_image_path)
+    df['full_path'] = df['file_name'].apply(lambda x: os.path.join(img_root, x))
+
+    return df
+
+
+def load_metadata(db_root: str, img_root: str) -> pd.DataFrame:
+    db_root = os.path.abspath(os.path.expanduser(db_root))
+    img_root = os.path.abspath(os.path.expanduser(img_root))
+
+    print('Loading metadata')
+    df = load_image_df(db_root, img_root)
+
+    print('Scanning images')
+    paths_from_files = set(glob.glob(os.path.join(img_root, '*', '*', '*.jpg')))
+    paths_from_db = set(df['full_path'].values)
+
+    print('Removing images without metadata')
+    count = 0
+    for im in tqdm(paths_from_files):
+        if im not in paths_from_db:
+            os.remove(im)
+            count += 1
+    print(f'Removed {count} images')
+
+    print('Removing images without files')
+    count = len(df)
+    df.drop(df[~df['full_path'].isin(paths_from_files)].index, inplace=True)
+    count = count - len(df)
+    print(f'Removed {count} images')
+
+    print(f'{len(df)} left')
+
+    return df
+
+
+def load_vndb_img_df(db_root: str, img_root: str):
+    metadata = load_metadata(db_root, img_root)
+    metadata.reset_index(inplace=True)
+    metadata['image'] = metadata['full_path']
+    return metadata
+
+
+TYPE_NAMES = [
+    'character',
+    'cover',
+    'screenshot_full',
+    'screenshot_thumb',
+]
+
+
+SEXUAL_NAMES = [
+    'safe',
+    'suggestive',
+    'explicit',
+]
+
+
+VIOLENCE_NAMES = [
+    'tame',
+    'violent',
+    'brutal',
+]
+
+
+class VNDBIMG(datasets.GeneratorBasedBuilder):
+    @property
+    def manual_download_instructions(self):
+        return '''\
+    Please download the vndb.org database dump manually from https://vndb.org/d14.
+    Download the 'Near-complete database' vndb-db-latest.tar.zst file.
+    Use `rsync` to download the 'Images' collection.
+
+    Create the following directory structure:
+
+    ```
+    my/dataset/path
+    ├── db
+    │   └── vndb-db-latest.tar.zst
+    └── vndb-img # this is the directory you downloaded with rsync
+        ├── ch
+        ├── cv
+        ├── sf
+        ├── st
+        └── ...
+    ```
+
+    Inside `my/dataset/path/db` run `zstd -d vndb-db-latest.tar.zst` and `tar -xf vndb-db-latest.tar`.
+
+    The final directory structure should look like this:
+
+    ```
+    my/dataset/path
+    ├── db
+    │   ├── vndb-db-latest.tar
+    │   ├── vndb-db-latest.tar.zst
+    │   ├── db
+    │   └── ...
+    └── vndb-img
+        ├── ch
+        ├── cv
+        ├── sf
+        ├── st
+        └── ...
+    ```
+
+    Finally, use `datasets.load_dataset('carbon225/vndb_img', data_dir='my/dataset/path')`.
+    '''
+
+    def _info(self):
+        features = datasets.Features(
+            {
+                'index': datasets.Value('int32'),
+                'id': datasets.Value('string'),
+                'width': datasets.Value('int32'),
+                'height': datasets.Value('int32'),
+                'c_votecount': datasets.Value('int32'),
+                'c_sexual_avg': datasets.Value('int32'),
+                'c_sexual_stddev': datasets.Value('int32'),
+                'c_violence_avg': datasets.Value('int32'),
+                'c_violence_stddev': datasets.Value('int32'),
+                'c_weight': datasets.Value('int32'),
+                'type': datasets.ClassLabel(names=TYPE_NAMES),
+                'sexual_class': datasets.ClassLabel(names=SEXUAL_NAMES),
+                'violence_class': datasets.ClassLabel(names=VIOLENCE_NAMES),
+                'file_name': datasets.Value('string'),
+                'full_path': datasets.Value('string'),
+                'image': datasets.Image(),
+            }
+        )
+        return datasets.DatasetInfo(
+            features=features,
+        )
+
+    def _split_generators(self, dl_manager):
+        root_path = os.path.abspath(os.path.expanduser(dl_manager.manual_dir))
+        db_path = os.path.join(root_path, 'db')
+        img_path = os.path.join(root_path, 'vndb-img')
+        if not os.path.exists(db_path) or not os.path.exists(img_path):
+            raise FileNotFoundError(f'Dataset not found at {root_path}. Please follow the manual download instructions.')
+        return [
+            datasets.SplitGenerator(
+                name=datasets.Split.TRAIN,
+                gen_kwargs={
+                    'db_path': db_path,
+                    'img_path': img_path,
+                },
+            ),
+        ]
+
+    def _generate_examples(self, db_path, img_path):
+        metadata = load_vndb_img_df(db_path, img_path)
+        for i, row in metadata.iterrows():
+            yield i, row.to_dict()