Nadeem-Shoukath115113/opensarm / tools /make_reward_collage.py
Nadeem-Shoukath115113's picture
download
raw
5.4 kB
"""Frame collage of one episode with each reward model's progress curve beside it.
One row per reward model: the same 5 overhead frames, and that model's progress for the whole
episode on the right. Layout follows the VLAC-style figure (frames left, curve right).
Values are read from what the video scripts already wrote under final_videos/:
<model>_raw/episode_XXXXXX/pred.npy per-frame prediction (make_sarm2_video.py, model mode)
<model>_spiral/episode_XXXXXX/reward.npy the progress column SPIRAL trains on
Usage (from Thesis_Code/):
Thesis/.venv/bin/python opensarm/tools/make_reward_collage.py --episode 18 \
--frames 0 141 231 348 545
"""
import argparse
from pathlib import Path
import cv2
import matplotlib.pyplot as plt
import numpy as np
from matplotlib.patches import FancyBboxPatch
FPS = 15
ROOT = Path(__file__).resolve().parents[1] / "final_videos"
VIDEO = (Path.home() / ".cache/huggingface/lerobot/local/teleop57_daggerrlt1_v21/videos/chunk-000/"
"observation.images.left_top/episode_{:06d}.mp4")
# teleop57_daggerrlt1_v21 holds the 57 teleop demos first, so DAgger episode e is episode 57 + e.
DAGGER_OFFSET = 57
BLUE, ORANGE = "#2a78d6", "#eb6834"
ROWS = [ # folder, file, title, colour
("rm1_raw", "pred.npy", "RM1: raw per-frame prediction", BLUE),
("rm2_raw", "pred.npy", "RM2: raw per-frame prediction", ORANGE),
("rm1_spiral", "reward.npy", "RM1: progress label used by SPIRAL", BLUE),
("rm2_spiral", "reward.npy", "RM2: progress label used by SPIRAL", ORANGE),
]
def read_frames(path, indices):
cap = cv2.VideoCapture(str(path))
out = []
for i in indices:
cap.set(cv2.CAP_PROP_POS_FRAMES, i)
ok, img = cap.read()
if not ok:
raise RuntimeError(f"cannot read frame {i} of {path}")
out.append(cv2.cvtColor(img, cv2.COLOR_BGR2RGB))
return out
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--episode", type=int, default=18)
ap.add_argument("--frames", type=int, nargs=5, default=[0, 141, 231, 348, 545])
ap.add_argument("--task", default="Fold the long sleeved T-shirt on the table")
ap.add_argument("--labels", nargs=5, default=None,
help="stage name under each frame, e.g. 'flat' 'first sleeve in' ...")
ap.add_argument("--out", default=None)
args = ap.parse_args()
ep = f"episode_{args.episode:06d}"
frames = read_frames(Path(str(VIDEO).format(DAGGER_OFFSET + args.episode)), args.frames)
curves = [(np.load(ROOT / d / ep / f), t, c) for d, f, t, c in ROWS]
n = len(ROWS)
fig = plt.figure(figsize=(13, 2.15 * n + 0.9), dpi=200)
fig.patch.set_facecolor("white")
# left block: 5 frames per row; right block: one curve per row
left, right = 0.015, 0.78
top, bot = 0.92, 0.065
row_h = (top - bot) / n
fw = (right - left - 0.03) / 5
fig.text((left + right) / 2, 0.965, f"Task: {args.task}", ha="center", va="center",
fontsize=13, style="italic")
for r, (y, title, color) in enumerate(curves):
y0 = top - (r + 1) * row_h
# tinted box behind the frames
# the last row carries the frame captions underneath, so its box reaches lower
ext = (0.035 if args.labels else 0.012) if r == n - 1 else 0.0
box = FancyBboxPatch((left, y0 + 0.012 - ext), right - left, row_h - 0.02 + ext,
boxstyle="round,pad=0.002,rounding_size=0.008",
transform=fig.transFigure, facecolor=color, alpha=0.08,
edgecolor=color, linewidth=1.2)
fig.patches.append(box)
fig.text(left + 0.008, y0 + row_h - 0.022, title, fontsize=10.5, weight="bold",
color="#222222", va="center")
fig.text(right - 0.008, y0 + row_h - 0.022, f"final progress {y[-1]:.2f}", fontsize=9.5,
color="#444444", va="center", ha="right")
for k, img in enumerate(frames):
ax = fig.add_axes([left + 0.015 + k * fw, y0 + 0.03, fw - 0.006, row_h - 0.075])
ax.imshow(img)
ax.set_xticks([]); ax.set_yticks([])
for s in ax.spines.values():
s.set_edgecolor("#bbbbbb"); s.set_linewidth(0.6)
if r == n - 1:
cap = f"t = {args.frames[k] / FPS:.1f} s"
if args.labels:
cap = f"{args.labels[k]}\n{cap}"
ax.set_xlabel(cap, fontsize=9, color="#444444")
ax = fig.add_axes([right + 0.04, y0 + 0.04, 0.17, row_h - 0.085])
t = np.arange(len(y)) / FPS
ax.plot(t, y, color=color, linewidth=2)
ax.set_ylim(-0.05, 1.05); ax.set_xlim(0, t[-1])
ax.set_yticks([0, 0.5, 1.0])
ax.tick_params(labelsize=8, colors="#555555")
ax.grid(color="#e6e6e6", linewidth=0.6)
for s in ("top", "right"):
ax.spines[s].set_visible(False)
for s in ("left", "bottom"):
ax.spines[s].set_color("#999999")
ax.set_ylabel("Progress", fontsize=8.5, color="#444444")
if r == n - 1:
ax.set_xlabel("time (s)", fontsize=8.5, color="#444444")
out = Path(args.out) if args.out else ROOT / f"collage_{ep}.png"
fig.savefig(out, facecolor="white")
fig.savefig(out.with_suffix(".pdf"), facecolor="white")
print(out, out.with_suffix(".pdf"))
if __name__ == "__main__":
main()

Xet Storage Details

Size:
5.4 kB
·
Xet hash:
7634da849bb83209793ab7dc8957004f09bec218b4e3aae4a9f30bbdc817d70b

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.