Spaces:
Running on Zero
Running on Zero
File size: 5,037 Bytes
0f5513d | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 | """Build the bundled gallery for the EchoLVFM Space.
For each chosen patient, this script:
1. Copies the latent `.pt` from `sample_data/CAMUS_Latents_4f4/<patient>.pt`
into `space/samples/<patient>.pt`.
2. Copies the matching real frames from
`<source_frames_root>/<patient>/frame_*.png` into
`space/samples/<patient>/`. If the patient has more frames than the
model's `max_frames` (32), the frames are subsampled with the same
`linspace` indices that `EchoDataset.resample_sequence` would use, so
the real frames stay aligned 1:1 with the latent.
3. Writes/updates `space/samples/manifest.json` with one entry per patient
containing: `id`, `latent_path`, `real_frames_dir`, `ef_true`,
`t_real` (= min(NbFrame, 32)), `fps_orig`, `view`.
Usage:
python space/scripts/build_samples.py \
--frames-root C:\\path\\to\\flow_matching\\data\\CAMUS_Processed_Frames
Defaults pick the 5 long-video patients added to `sample_data/` (NbFrame
30–37).
"""
from __future__ import annotations
import argparse
import json
import shutil
from pathlib import Path
import pandas as pd
import torch
from PIL import Image
REPO_ROOT = Path(__file__).resolve().parents[2]
SAMPLE_DATA_DIR = REPO_ROOT / "sample_data" / "CAMUS_Latents_4f4"
SAMPLES_DIR = REPO_ROOT / "space" / "samples"
DEFAULT_PATIENTS = [
"patient0082_4CH",
"patient0106_4CH",
"patient0310_4CH",
"patient0326_4CH",
"patient0422_4CH",
]
MAX_FRAMES = 32 # Matches `cfg.dataset.max_frames` for all three checkpoints.
def _resample_indices(t_orig: int, target: int) -> list[int]:
"""Mirror `EchoDataset.resample_sequence` index selection for T > target."""
return torch.linspace(0, t_orig - 1, target).round().long().tolist()
def _copy_real_frames(src_dir: Path, dst_dir: Path, t_real: int) -> None:
"""Copy frame_*.png files. If the source has > MAX_FRAMES, subsample with
linspace indices so frames align with the (resampled) latent."""
src_frames = sorted(src_dir.glob("frame_*.png"),
key=lambda p: int(p.stem.split("_")[-1]))
if not src_frames:
raise FileNotFoundError(f"No frame_*.png in {src_dir}")
n_src = len(src_frames)
if n_src > MAX_FRAMES:
idxs = _resample_indices(n_src, MAX_FRAMES)
chosen = [src_frames[i] for i in idxs]
else:
chosen = src_frames # T_real == n_src
if len(chosen) != t_real:
raise RuntimeError(
f"Expected {t_real} frames after resampling for {src_dir.name}, "
f"got {len(chosen)} (n_src={n_src})"
)
dst_dir.mkdir(parents=True, exist_ok=True)
for new_idx, p in enumerate(chosen):
# Write 0-indexed to match `frame_to_mp4` ordering.
Image.open(p).convert("RGB").save(dst_dir / f"frame_{new_idx}.png")
def build(patients: list[str], frames_root: Path) -> None:
if not SAMPLE_DATA_DIR.exists():
raise FileNotFoundError(f"Sample data dir not found: {SAMPLE_DATA_DIR}")
meta_csv = SAMPLE_DATA_DIR / "metadata.csv"
df = pd.read_csv(meta_csv).set_index("video_name")
SAMPLES_DIR.mkdir(parents=True, exist_ok=True)
manifest = {"samples": []}
for pid in patients:
if pid not in df.index:
raise KeyError(f"{pid} not in {meta_csv}")
row = df.loc[pid]
n_orig = int(row["NbFrame"])
t_real = min(n_orig, MAX_FRAMES)
latent_src = SAMPLE_DATA_DIR / f"{pid}.pt"
if not latent_src.exists():
raise FileNotFoundError(latent_src)
shutil.copyfile(latent_src, SAMPLES_DIR / f"{pid}.pt")
frames_src = frames_root / pid
if not frames_src.is_dir():
raise FileNotFoundError(frames_src)
_copy_real_frames(frames_src, SAMPLES_DIR / pid, t_real)
manifest["samples"].append({
"id": pid,
"latent_path": f"{pid}.pt",
"real_frames_dir": pid,
"ef_true": float(row["EF_AL"]),
"t_real": t_real,
"n_orig": n_orig,
"fps_orig": float(row["FrameRate"]),
"view": str(row["view"]),
})
print(f" + {pid}: t_real={t_real} (n_orig={n_orig}), "
f"ef_true={row['EF_AL']}, fps={row['FrameRate']}")
manifest_path = SAMPLES_DIR / "manifest.json"
manifest_path.write_text(json.dumps(manifest, indent=2))
print(f"\nWrote {manifest_path} with {len(manifest['samples'])} samples")
def _parse_args() -> argparse.Namespace:
p = argparse.ArgumentParser(description=__doc__,
formatter_class=argparse.RawDescriptionHelpFormatter)
p.add_argument("--frames-root", required=True, type=Path,
help="Path to CAMUS_Processed_Frames root.")
p.add_argument("--patients", nargs="+", default=DEFAULT_PATIENTS,
help="Patient ids to bundle.")
return p.parse_args()
def main() -> None:
args = _parse_args()
build(args.patients, args.frames_root)
if __name__ == "__main__":
main()
|