SKING_DDJ_Dataset / dataset_layout.py
EntropyDrop
update
6ab2f07
Raw History Blame Contribute Delete
1.3 kB
"""Shared file naming and directory layout for dataset readers and writers."""
from pathlib import Path
import re
SHARD_PREFIX_LENGTH = 3
DATASET_FILE_PATTERN = re.compile(
r"(?P<id>.+?)_(?:(?:source|intermediate|edited|result|perfect)"
r"(?:_[A-Za-z0-9][A-Za-z0-9_-]*)?\.(?:png|jpe?g|webp)"
r"|(?:prompt|best|perfect|all_rejected)\.txt"
r"|(?:author|decode)\.json"
r"|(?:meat|semantic)\.(?:png|json)"
r"|result(?:_[A-Za-z0-9][A-Za-z0-9_-]*)?\.uv_review\.json)"
)
def generation_directory(root: Path, model_version: str, generation_id: str) -> Path:
"""Use the first three ID characters for every model; short IDs stay intact."""
for value in (model_version, generation_id):
if not value or value in {".", ".."} or any(c in value for c in "/\\\0"):
raise ValueError("Model version and generation ID must be single path components")
return Path(root) / model_version / generation_id[:SHARD_PREFIX_LENGTH]
def model_directory(sample_directory: Path, generation_id: str) -> Path:
"""Recognize current shards and the previous one-character layout for readers."""
directory = Path(sample_directory)
if directory.name in {generation_id[:SHARD_PREFIX_LENGTH], generation_id[:1]}:
return directory.parent
return directory