Upload configs/make_tomls.py with huggingface_hub

configs/make_tomls.py

@@ -0,0 +1,76 @@
+#!/usr/bin/env python3
+"""Gera os TOMLs de cache e de treino a partir das views ja criadas.
+
+Schema H3 atual: `target_video_directory` + `target_modalities`. As chaves
+antigas (`video_directory`, `h3_target_mode`) sao rejeitadas com ValueError
+pelo trainer, nao aceitas como alias.
+
+Cache e treino compartilham fontes/areas/duracoes/caches; a UNICA diferenca e o
+batch_size (alto no cache, obrigatoriamente 1 no treino).
+"""
+from __future__ import annotations
+import argparse
+from pathlib import Path
+
+# Todas as views em 124 frames: minimo do H3 de 5 s respeitado (5,17 s e o
+# primeiro valor do grid 17k+5 acima de 5 s) e igual ao regime de geracao alvo.
+VIEWS = [
+    ("v288", 512, 288, 124),
+    ("v384", 640, 384, 124),
+    ("v480", 832, 480, 124),
+]
+
+HEADER = """# Gerado por make_tomls.py — {kind}
+[general]
+caption_extension = ".txt"
+batch_size = {batch}
+enable_bucket = true
+bucket_no_upscale = true
+"""
+
+ENTRY = """
+[[datasets]]
+target_video_directory = "{videos}"
+target_modalities = ["video"]
+cache_directory = "{cache}"
+resolution = [{w}, {h}]
+source_fps = 24.0
+frame_extraction = "full"
+max_frames = {frames}
+num_repeats = 1
+"""
+
+
+def build(views_root: Path, cache_root: Path, batch: int, kind: str, only: set[str] | None) -> str:
+    out = HEADER.format(kind=kind, batch=batch)
+    for name, w, h, frames in VIEWS:
+        if only and name not in only:
+            continue
+        videos = views_root / name
+        if not videos.is_dir() or not any(videos.glob("*.mp4")):
+            continue
+        out += ENTRY.format(videos=videos, cache=cache_root / name,
+                            w=w, h=h, frames=frames)
+    return out
+
+
+def main() -> None:
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--views", required=True, type=Path)
+    ap.add_argument("--cache", required=True, type=Path)
+    ap.add_argument("--out-dir", required=True, type=Path)
+    ap.add_argument("--prefix", default="dataset_smoothloopv3")
+    ap.add_argument("--cache-batch", type=int, default=24)
+    ap.add_argument("--only", default="", help="views separadas por virgula")
+    a = ap.parse_args()
+
+    only = {x for x in a.only.split(",") if x} or None
+    a.out_dir.mkdir(parents=True, exist_ok=True)
+    for kind, batch in (("cache", a.cache_batch), ("train", 1)):
+        path = a.out_dir / f"{a.prefix}_{kind}.toml"
+        path.write_text(build(a.views, a.cache, batch, kind, only), encoding="utf-8")
+        print(path)
+
+
+if __name__ == "__main__":
+    main()