render_full.py

#!/usr/bin/env python
"""Render 24 views (az 0..315 step 45 x elevations low/-20, eye/10, high/40) of ALL
Google Scanned Objects from suvadityamuk/google-scanned-objects webdataset shards.

768x768 RGBA, transparent background, soft three-point lighting, no ground plane.
Z-up meshes converted to Y-up. One camera distance per object, from the bounding
sphere (eye-level projected span ~60% of frame, fov 40 deg). Per-object rejection:
any view with alpha coverage < 8% or projected span > 95% of frame.

Outputs (uploaded to ysharma/gso-orbit-rgba):
  renders/{object_id}/az{000..315}_el{low|eye|high}.png
  metadata.jsonl          one line per view
  split.json              heldout_40 / rejected_objects / train_objects
  contact_sheets/*.jpg    one sheet per sampled category (uploaded after core data)
"""
import os
os.environ['PYOPENGL_PLATFORM'] = 'osmesa'  # must precede pyrender import
import io, json, math, random, tarfile, time, traceback
from concurrent.futures import ProcessPoolExecutor

import numpy as np
import pyrender
import trimesh
from scipy.ndimage import distance_transform_edt
from huggingface_hub import HfApi

DATA = '/data/data'
OUT = '/out'
REPO_ID = 'ysharma/gso-orbit-rgba'
N_SHARDS = 43
SZ = 1536
OUTSZ = 768
FOV_Y = math.radians(40.0)
FOCAL_F = 0.60
DIST = 0.5 / (FOCAL_F * math.tan(FOV_Y / 2.0))
ZNEAR, ZFAR = 0.05, 10.0
ELEVS = [('low', -20.0), ('eye', 10.0), ('high', 40.0)]
AZS = list(range(0, 360, 45))
HALO_RING = 8
BG = [0.0, 0.0, 0.0]
COV_MIN = 0.08
SPAN_MAX = 0.95
WORKERS = 8

T_UP = np.eye(4); T_UP[1, 1] = 0.0; T_UP[1, 2] = -1.0; T_UP[2, 1] = 1.0; T_UP[2, 2] = 0.0


def camera_pose(az_deg, el_deg, dist):
    az, el = math.radians(az_deg), math.radians(el_deg)
    eye = dist * np.array([math.cos(el) * math.cos(az), math.sin(el),
                           math.cos(el) * math.sin(az)])
    zax = eye / np.linalg.norm(eye)
    up = np.array([0.0, 1.0, 0.0])
    xax = np.cross(up, zax); xax /= np.linalg.norm(xax)
    yax = np.cross(zax, xax)
    pose = np.eye(4)
    pose[:3, 0], pose[:3, 1], pose[:3, 2], pose[:3, 3] = xax, yax, zax, eye
    return pose


def rot_y(deg):
    a = math.radians(deg)
    R = np.eye(4)
    R[:3, :3] = [[math.cos(a), 0, math.sin(a)], [0, 1, 0], [-math.sin(a), 0, math.cos(a)]]
    return R


def build_scene(glb_bytes):
    st = trimesh.load(io.BytesIO(glb_bytes), file_type='glb', force='scene')
    geoms = list(st.geometry.values())
    for g in geoms:
        g.apply_transform(T_UP)
    mesh = pyrender.Mesh.from_trimesh(geoms)
    scene = pyrender.Scene(bg_color=BG, ambient_light=[0.30, 0.30, 0.30])
    scene.add(mesh)
    return scene


def add_lights(scene, cam_pose):
    base = cam_pose.copy(); base[:3, 3] = 0.0
    scene.add(pyrender.DirectionalLight(intensity=3.0), pose=base @ rot_y(0.0))    # key
    scene.add(pyrender.DirectionalLight(intensity=1.2), pose=base @ rot_y(35.0))   # fill
    scene.add(pyrender.DirectionalLight(intensity=2.0), pose=base @ rot_y(160.0))  # rim


def tar_read(tf, name):
    m = None
    for cand in (name, name.lower(), name.replace(" ", "_")):
        try:
            m = tf.getmember(cand)
            break
        except KeyError:
            continue
    if m is None:
        for mem in tf.getmembers():
            if mem.name.lower() == name.lower():
                m = mem
                break
    if m is None:
        raise KeyError(name)
    f = tf.extractfile(m)
    return f.read() if f else None


def halo_fill(rgb, mask):
    inv = (~mask)
    if not inv.any() or not mask.any():
        return rgb
    dist, idx = distance_transform_edt(inv, return_indices=True)
    fill = inv & (dist <= HALO_RING)
    rgb = rgb.copy()
    ys, xs = np.nonzero(fill)
    rgb[ys, xs] = rgb[idx[0][ys, xs], idx[1][ys, xs]]
    return rgb


def downsample_rgba(rgb, mask):
    premul = rgb.astype(np.float32) * mask[..., None]
    a = mask.astype(np.float32)
    h2 = SZ // 2
    pm = premul.reshape(h2, 2, h2, 2, 3).mean(axis=(1, 3))
    am = a.reshape(h2, 2, h2, 2).mean(axis=(1, 3))
    out = np.zeros((h2, h2, 4), np.uint8)
    good = am > (1.0 / 255.0)
    denom = np.maximum(am, 1e-6)
    for c in range(3):
        ch = np.clip(pm[..., c] / denom, 0, 255)
        out[..., c] = np.where(good, ch, 0).astype(np.uint8)
    out[..., 3] = np.clip(am * 255.0 + 0.5, 0, 255).astype(np.uint8)
    return out


def render_object(args):
    shard_path, object_id, category = args
    t_start = time.time()
    try:
        with tarfile.open(shard_path, 'r') as tf:
            glb_bytes = tar_read(tf, f'{object_id}.glb')
        os.makedirs(os.path.join(OUT, 'renders', object_id), exist_ok=True)
        scene = build_scene(glb_bytes)
        cam = pyrender.PerspectiveCamera(yfov=FOV_Y, znear=ZNEAR, zfar=ZFAR)
        scene.add(cam, pose=camera_pose(0.0, 10.0, DIST))
        add_lights(scene, camera_pose(0.0, 10.0, DIST))
        renderer = pyrender.OffscreenRenderer(SZ, SZ)
        rows, t_render = [], 0.0
        eye_covs = []
        for el_name, el in ELEVS:
            for az in AZS:
                t0 = time.time()
                scene.set_pose(scene.main_camera_node, camera_pose(az, el, DIST))
                color, depth = renderer.render(scene)
                t_render += time.time() - t0
                mask = (depth > 0.0)
                cov = float(mask.mean())
                ys_m, xs_m = np.nonzero(mask)
                span = max((xs_m.max() - xs_m.min()) / SZ, (ys_m.max() - ys_m.min()) / SZ) if len(xs_m) else 0.0
                if el_name == 'eye':
                    eye_covs.append(cov)
                rgb = halo_fill(color, mask)
                rgba = downsample_rgba(rgb, mask)
                fn = f'az{az:03d}_el{el_name}.png'
                from PIL import Image
                Image.fromarray(rgba, 'RGBA').save(os.path.join(OUT, 'renders', object_id, fn))
                rows.append({
                    'object_id': object_id, 'category': category,
                    'azimuth_deg': az, 'elevation_deg': el, 'elevation_name': el_name,
                    'camera_azimuth_deg': az, 'camera_elevation_deg': el,
                    'distance': round(DIST, 6), 'coverage': round(cov, 6), 'bbox_span': round(span, 6),
                    'file': f'renders/{object_id}/{fn}',
                })
        renderer.delete()
        rejected = any(r['coverage'] < COV_MIN or r['bbox_span'] > SPAN_MAX for r in rows)
        return {'ok': True, 'rows': rows, 'object_id': object_id, 'rejected': rejected,
                's_per_view': round(t_render / 24.0, 3),
                'render_s': round(t_render, 3)}
    except Exception:
        return {'ok': False, 'object_id': object_id, 'error': traceback.format_exc()}


def list_all_objects():
    jobs = []
    for i in range(N_SHARDS):
        shard = os.path.join(DATA, f'gso-train-{i:05d}.tar')
        with tarfile.open(shard, 'r') as tf:
            stems, seen = [], set()
            for m in tf.getnames():
                ext = m.rsplit('.', 1)[1].lower()
                if ext not in ('glb', 'json'):
                    continue
                stem = m.rsplit('.', 1)[0]
                if stem not in seen:
                    seen.add(stem)
                    stems.append(stem)
        for stem in stems:
            with tarfile.open(shard, 'r') as tf:
                j = json.loads(tar_read(tf, f'{stem}.json'))
            jobs.append((shard, stem, j.get('category', '') or 'unlabeled'))
    return jobs


def make_split(jobs, all_flags):
    cats = {}
    for _, oid, cat in jobs:
        cats.setdefault(cat, []).append(oid)
    rng = random.Random(20260923)
    heldout = []
    total = sum(len(v) for v in cats.values())
    for cat in sorted(cats):
        pool = [o for o in cats[cat] if not all_flags.get(o, False)]
        rng.shuffle(pool)
        n = max(1, round(40.0 * len(pool) / total)) if pool else 0
        heldout.extend(pool[:n])
    rng.shuffle(heldout)
    heldout = heldout[:40]
    hs = set(heldout)
    train = [o for _, o, c in jobs
             if o not in hs and not all_flags.get(o, False)]
    return heldout, train


def contact_sheets(sheets_ids):
    from PIL import Image
    os.makedirs(os.path.join(OUT, 'contact_sheets'), exist_ok=True)
    cell = 256
    for sample_id in sheets_ids:
        sheet = Image.new('RGB', (8 * cell, 3 * cell), (128, 128, 128))
        for r, (el_name, _el_deg) in enumerate(ELEVS):
            for c, az in enumerate(AZS):
                im = Image.open(os.path.join(OUT, 'renders', sample_id,
                                             f'az{az:03d}_el{el_name}.png'))
                im = im.resize((cell, cell), Image.LANCZOS)
                sheet.paste(im, (c * cell, r * cell), im)
        sheet.save(os.path.join(OUT, 'contact_sheets', f'{sample_id}.jpg'), quality=88)


def main():
    t0 = time.time()
    jobs = list_all_objects()
    print(f'{len(jobs)} objects listed across {N_SHARDS} shards', flush=True)
    os.makedirs(OUT, exist_ok=True)
    all_rows, ok_flags, sheets_pool = [], {}, []
    s_per_view = []
    n_ok = n_fail = 0
    with ProcessPoolExecutor(max_workers=WORKERS) as ex:
        for k, res in enumerate(ex.map(render_object, jobs)):
            if not res['ok']:
                print(f"FAILED {res['object_id']}: {res['error'][-300:]}", flush=True)
                n_fail += 1
                continue
            n_ok += 1
            all_rows.extend(res['rows'])
            ok_flags[res['object_id']] = res['rejected']
            s_per_view.append(res['s_per_view'])
            if not res['rejected'] and len(sheets_pool) < 8:
                sheets_pool.append(res['object_id'])
            if (k + 1) % 100 == 0:
                print(f'  {k+1}/{len(jobs)} objects | mean s/view {np.mean(s_per_view):.3f} '
                      f'| elapsed {time.time()-t0:.0f}s', flush=True)
    with open(os.path.join(OUT, 'metadata.jsonl'), 'w') as f:
        for r in all_rows:
            f.write(json.dumps(r) + '\n')
    heldout, train = make_split(jobs, ok_flags)
    split = {'heldout_40': heldout, 'train_objects': train,
             'rejected_objects': sorted([o for o, r in ok_flags.items() if r]),
             'failed_objects': [], 'seed': 20260923,
             'n_renders': len(all_rows), 'instructions_version': 'orbit-relative-v1'}
    with open(os.path.join(OUT, 'split.json'), 'w') as f:
        json.dump(split, f, indent=2)
    bad = sum(1 for r in all_rows if r['coverage'] < COV_MIN)
    n_rej = sum(1 for r in ok_flags.values() if r)
    print(f'\nDONE {n_ok} ok / {n_fail} failed | rejected {n_rej} objects | '
          f'views <8% cov: {bad}/{len(all_rows)} | {time.time()-t0:.0f}s total '
          f'| mean {np.mean(s_per_view):.3f} s/view ({WORKERS} workers)', flush=True)
    print(f'split: {len(heldout)} heldout, {len(train)} train', flush=True)

    api = HfApi(token=os.environ.get('HF_TOKEN'))
    print('uploading core data to Hub (upload_large_folder)...', flush=True)
    t_up = time.time()
    api.upload_large_folder(repo_id=REPO_ID, repo_type='dataset', folder_path=OUT)
    print(f'UPLOAD DONE in {time.time()-t_up:.0f}s', flush=True)

    # contact sheets last, guarded: a bug here must never cost the render sweep
    try:
        contact_sheets(sheets_pool[:6])
        api.upload_folder(repo_id=REPO_ID, repo_type='dataset',
                          folder_path=os.path.join(OUT, 'contact_sheets'),
                          path_in_repo='contact_sheets')
        print('contact sheets uploaded', flush=True)
    except Exception:
        print('contact sheets failed (non-fatal):', flush=True)
        traceback.print_exc()


if __name__ == '__main__':
    main()