actaug/code/compose_2x3.py

#!/usr/bin/env python3
"""compose_2x3.py -- 2x3 comparison video: row 1 = the ORIGINAL episode's three
dataset camera views, row 2 = the augmented (perturbed + re-earned) rollout's
same three views. The shorter row freezes on its last frame.

    compose_2x3.py <episode_idx> <rollout_dump_dir> <out.mp4> [dataset_root]
"""
import sys
from pathlib import Path

import imageio.v2 as iio
import numpy as np

from actaug_core import DEFAULT_DATASET, write_mp4

VIEWS = ["robot0_agentview_left", "robot0_agentview_right", "robot0_eye_in_hand"]


def read_video(path):
    return [f for f in iio.get_reader(str(path))]


def main(ep_idx, dump_dir, out_path, dataset_root=str(DEFAULT_DATASET)):
    ep_idx = int(ep_idx)
    root = Path(dataset_root)
    chunk = ep_idx // 1000
    orig = [read_video(root / "videos" / f"chunk-{chunk:03d}" /
                       f"observation.images.{v}" / f"episode_{ep_idx:06d}.mp4")
            for v in VIEWS]
    dump = Path(dump_dir)
    aug = [read_video(dump / f"{n}.mp4") for n in ("left", "right", "wrist")]
    H, W = aug[0][0].shape[:2]
    orig = [[np.asarray(f)[..., :3] for f in v] for v in orig]
    aug = [[np.asarray(f)[..., :3] for f in v] for v in aug]
    # resize originals to the rollout render size if they differ
    if orig[0][0].shape[:2] != (H, W):
        import cv2
        orig = [[cv2.resize(f, (W, H)) for f in v] for v in orig]
    T = max(len(orig[0]), len(aug[0]))
    pad = lambda v: v + [v[-1]] * (T - len(v))
    orig, aug = [pad(v) for v in orig], [pad(v) for v in aug]
    frames = [np.concatenate([np.concatenate([o[t] for o in orig], axis=1),
                              np.concatenate([a[t] for a in aug], axis=1)],
                             axis=0) for t in range(T)]
    write_mp4(out_path, frames, fps=20)
    print(f"wrote {out_path} ({T} frames, {frames[0].shape[1]}x{frames[0].shape[0]})")


if __name__ == "__main__":
    main(*sys.argv[1:])