visualize_policy_input_keyframes.py

#!/usr/bin/env python3
"""Visualize policy-input subgoal keyframes (same style as compute_subgoal_embedding).

Sampling matches train_policy.py / RoboSuiteDataset.sample_goal_sequence_paths.
Preprocessing matches encode_goals_with_r3m: raw demo PNG -> Resize(224) -> [0,1] tensor.
(No CenterCrop; RoboSuiteDataset transform is NOT used for subgoal encoding.)
"""

import argparse
import glob
import os

import matplotlib.pyplot as plt
import numpy as np
import torchvision.transforms as T
from PIL import Image


def save_keyframe_visualization(sampled_images, save_path):
    """Same layout as multi-task-tcc-robosuite/compute_subgoal_embedding.py."""
    sampled_images = np.squeeze(sampled_images)
    num_trajectories = sampled_images.shape[0]
    num_keyframes = sampled_images.shape[1]

    fig, axes = plt.subplots(
        num_trajectories,
        num_keyframes,
        figsize=(num_keyframes * 2, num_trajectories * 2),
    )

    for i in range(num_trajectories):
        for j in range(num_keyframes):
            ax = axes[i, j] if num_trajectories > 1 else axes[j]
            img = sampled_images[i, j].transpose(1, 2, 0)
            ax.imshow(img)
            ax.axis("off")

    plt.tight_layout()
    plt.savefig(save_path)
    plt.close()
    print(f"Saved keyframe visualization to {save_path}.")


def build_r3m_display_transform(size=224):
    """Same resize path as train_policy.encode_goals_with_r3m (display only)."""
    return T.Compose([
        T.ToPILImage(),
        T.Resize(size),
        T.ToTensor(),
    ])


def load_demo_dirs(demo_root):
    dirs = sorted(
        glob.glob(os.path.join(demo_root, "*/")),
        key=lambda p: int(os.path.basename(os.path.normpath(p))),
    )
    return [d for d in dirs if glob.glob(os.path.join(d, "*.png"))]


def keyframe_paths_for_demo(seq_dir, num_keyframes=8):
    seq = sorted(
        glob.glob(os.path.join(seq_dir, "*.png")),
        key=lambda x: int(os.path.splitext(os.path.basename(x))[0]),
    )
    if not seq:
        return []
    n = len(seq)
    indices = np.linspace(0, n - 1, num=num_keyframes, dtype=int)
    return [seq[i] for i in indices]


def collect_policy_keyframes(demo_root, num_keyframes=8, display_size=224):
    transform = build_r3m_display_transform(display_size)
    traj_keyframes = []
    for seq_dir in load_demo_dirs(demo_root):
        paths = keyframe_paths_for_demo(seq_dir, num_keyframes)
        if len(paths) != num_keyframes:
            continue
        frames = []
        for path in paths:
            raw = np.array(Image.open(path).convert("RGB"))
            tensor = transform(raw)  # [c, h, w] in [0, 1]
            frames.append(tensor.numpy())
        traj_keyframes.append(np.stack(frames, axis=0))
    return np.array(traj_keyframes)


def save_cursor_previews(full_png_path, output_stem):
    """Save small JPG rows + HTML viewer (works when IDE image preview fails)."""
    im = Image.open(full_png_path).convert("RGB")
    w, h = im.size
    out_dir = os.path.dirname(os.path.abspath(full_png_path))
    rows_dir = os.path.join(out_dir, "policy_keyframes_rows")
    os.makedirs(rows_dir, exist_ok=True)

    num_demos = h // max(1, w // 8)  # infer ~square cells; fallback below
    # matplotlib grid: each row one demo, row height = h / num_trajectories
    # count rows by scanning or use fixed 35 for lift
    row_h = h // 35 if h >= 35 * 8 else h // max(1, int(h / (w / 8)))
    num_demos = max(1, h // row_h)

    rows_html = []
    for i in range(num_demos):
        top = i * row_h
        bottom = h if i >= num_demos - 1 else (i + 1) * row_h
        row = im.crop((0, top, w, bottom))
        row = row.resize(
            (640, max(1, int(640 * row.height / row.width))), Image.Resampling.LANCZOS
        )
        fname = f"demo_{i:02d}.jpg"
        row.save(os.path.join(rows_dir, fname), format="JPEG", quality=88, optimize=True)
        rows_html.append(
            f'<div class="row"><span class="label">demo {i}</span>'
            f'<img src="policy_keyframes_rows/{fname}" width="640"></div>'
        )

    html_path = f"{output_stem}_viewer.html"
    with open(html_path, "w", encoding="utf-8") as f:
        f.write(
            "<!DOCTYPE html><html><head><meta charset=\"utf-8\">"
            "<title>Policy Keyframes</title><style>"
            "body{font-family:system-ui;margin:12px;background:#1a1a1a;color:#ddd}"
            "h1{font-size:16px}.row{margin:6px 0;display:flex;align-items:center;gap:8px}"
            ".label{width:56px;font-size:11px;color:#888;flex-shrink:0}img{border:1px solid #444}"
            "</style></head><body>"
            "<h1>Policy input keyframes — rows=demos, cols=k0→k7</h1>"
            "<p style=\"font-size:12px;color:#888\">"
            "在浏览器打开此 HTML 文件查看（Cursor 图片预览可能不支持远程大图）</p>"
            + "\n".join(rows_html)
            + "</body></html>"
        )
    print(f"Saved HTML viewer to {html_path}")
    print(f"Saved {num_demos} row previews under {rows_dir}/")


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument(
        "--demo_root",
        default="/home/lei/Documents/tong/irl4idm/multi-task-tcc-robosuite/experiments/datasets/mimicgen/train/lift",
    )
    parser.add_argument("--num_keyframes", type=int, default=8)
    parser.add_argument(
        "--output",
        default="./logs/policy_input_keyframes.png",
    )
    parser.add_argument(
        "--preview",
        default=None,
        help="Unused; kept for compatibility. Previews are auto-generated.",
    )
    args = parser.parse_args()

    sampled_images = collect_policy_keyframes(
        args.demo_root, num_keyframes=args.num_keyframes
    )
    os.makedirs(os.path.dirname(os.path.abspath(args.output)), exist_ok=True)
    save_keyframe_visualization(sampled_images, args.output)

    output_stem, _ = os.path.splitext(args.output)
    save_cursor_previews(args.output, output_stem)


if __name__ == "__main__":
    main()